The AI guide.

The job ad and the description

Most job ads are last year's ad with the date changed. Give an assistant the role, the pay, the hours and what a good first year looks like, and it drafts the ad and the description in your voice for you to edit. It is quick at the dull part, the clear wording, and it keeps every ad consistent. You still decide what the job is and what it pays.

Sorting applications, not rejecting people

An assistant can read a pile of applications against the must-haves you set, pull out the relevant experience and flag the ones worth a call. That saves hours. It must not reject anyone on its own: tools like this can repeat the bias in the examples they learned from, and the rules on automated hiring decisions are spreading, with some places already requiring notice to candidates or an audit of the tool. A person reads every shortlist, and looks at the ones it passed over.

Interviews, booked and written up

Arranging interviews is back-and-forth messaging, which booking by text and chat does well. With the candidate's permission, meeting notes capture the interview, so you compare what people said rather than what you remember. The judgement about fit stays with the people in the room.

The first week

New starters ask the same questions: where things are, how a job is done here, who to ask. An assistant that answers from your own handbook and procedures gives them an answer at any hour and frees whoever used to field the questions. It only knows what you have written down, which is a good reason to write it down.

What a person still does

Decides who is interviewed, who is hired and on what terms. Tells people no, kindly. Checks that the tools treat every candidate the same way. Hiring is where a business is most exposed if an automated decision goes wrong, so it is where people stay closest to the work.

Drafts in your voice

Posts, emails, website pages and replies to comments. Give an assistant examples of how you actually write and the facts it must not change, and it drafts in your voice for you to edit. A draft is a starting point, not a publish button. Check every claim, number and name before it goes out, because a confident mistake in public costs more than a slow post.

Ideas worth posting

The hard part is often what to say, not how to say it. An assistant is a good partner for turning a week of jobs, questions and lessons into ideas worth posting, and for testing which angle reads clearer. Brainstorming and content are where we use it most in our own marketing.

Reviews, asked for and answered

Happy customers rarely leave a review unless someone asks at the right moment. Review requests can go out after a finished job, and replies can be drafted for a person to approve. Never post a reply to a complaint without reading it yourself.

Your listings, kept right

Hours, services, service area and photos drift out of date across the places people find you. Keeping them consistent is dull, repetitive work, and worth doing well, because it is often the first thing a new customer sees.

No enquiry left waiting

Marketing that brings in an enquiry nobody answers is money spent twice. Missed-call text back, a chat assistant on the website and follow-up that keeps going make sure the work you paid to attract reaches a person. Sales picks it up from there.

What a person still does

Decides what the business says and what it will not say. Approves anything that goes out under your name. Handles complaints and anything personal. The voice is yours; AI only helps you use it more often.

The first answer, fast

The business that answers first often wins the job. Missed calls get a text back in a minute, the website answers the common questions at any hour, and every enquiry lands in one place with what the customer asked. Speed is the cheapest sales advantage there is.

Quotes from your own history

For repeat and near-repeat work, AI reads the request and drafts a quote from what similar jobs actually took, for a person to check and sign. The quoting page goes into where that works and where it goes wrong.

Follow-up that keeps going

A quote with no follow-up is a coin toss. Follow-up on a sensible schedule, in your voice, until the customer says yes, no or not now, wins work that would have gone quiet. It stops the moment a person takes over the conversation.

Notes from every call

Call summaries mean the customer's needs, the promises made and the next step are written down without anyone typing them up, so the next person to call knows the history. Ask before recording; in some places the law requires everyone on the call to agree.

What a person still does

Sets the price and signs the quote. Handles the negotiation, the awkward call and anything that needs judgement. Decides when to walk away. AI keeps the pipeline moving; people close it.

Orders and paperwork without the re-key

Orders arrive by email, PDF and phone, and someone types them into the system. AI reads them into your format, fills in what it can and sends the rest to a person. The same goes for supplier invoices matched to their orders. The win is time, and fewer typing errors.

A schedule that survives the week

A schedule is right on Monday and wrong by Wednesday. AI can draft the week from the real jobs, people and constraints, and redraft it when something changes, for a person to approve. It is only as good as the information it can see, which is usually the first thing to fix.

Stock and purchasing

Counts you can trust, reorder points that move with demand, and supplier quotes gathered in one place. The value is fewer stock-outs and less cash sitting on the shelf.

Getting paid

Invoices go out when the job closes, and reminders follow on a schedule, politely, until they are paid. A person handles the customer who is struggling or disputing a charge.

Answers from your own documents

Procedures, price lists, manuals and past jobs, searchable with a plain question. The team stops interrupting the one person who knows, and that person gets their day back.

What a person still does

Approves anything that commits money or changes a promise to a customer. Owns the exceptions the tools hand them. Decides what is worth automating at all: some processes should be fixed or dropped before anything is automated.

Every call answered

An AI receptionist answers when you cannot: after hours, on the tools, on another call. It books the job, takes the message and passes anything urgent or unusual to a person straight away. Clients get an answer instead of voicemail.

The same questions, answered

Opening hours, prices, what to expect, how to prepare. A chat assistant on your website and in your inbox answers from your own information, at any hour, and hands over when it does not know.

Bookings and reminders

Booking by text and chat, with confirmations and reminders, cuts the back-and-forth and the no-shows. The client picks a time that suits them, and your calendar stays right.

Kept in the loop

Short updates when a job is booked, on the way, delayed or done. Most chasing calls are a client asking where things are, and an update that arrives before they ask is worth more than a good answer after.

Reviews and the relationship

Asking for a review after a good job, thanking the people who leave one, and noticing the clients who go quiet. The tools do the remembering; the relationship is still yours.

What a person still does

Handles complaints, bad news and anything emotional. Makes the judgement calls on exceptions and goodwill. Picks up the phone to the client who is unhappy. AI should make it easier to reach a person, never harder.

The test

A use case is worth money when all five of these are true. Most fail on the second or the fourth.

• The outcome is valuable: hours, errors or delays you can point to.
• The information it needs already exists somewhere you can get at.
• A person checks the output before it matters, at least at first.
• Being wrong occasionally is tolerable, because it will be.
• Somebody wants it to work and has the time to make it work.

The usual suspects

The same handful of workflows come up in almost every small business, because they are all the same shape: information arrives in a messy form and someone spends hours turning it into a tidy one.

• Quoting: turning emails, drawings and spreadsheets into a priced quote, especially for repeat work.
• Order entry: pulling orders out of emails and PDFs into the system that runs the business.
• Scheduling and dispatch: who goes where, what runs next, what to do when something slips.
• Customer service: the first reply, the status update, the question that has been answered a hundred times.
• Sales follow-up: the note after the call, the reminder that never gets sent.
• Finding things: the procedure, the drawing, the last email about this customer, without asking the one person who knows.
• Reporting: the weekly numbers that someone assembles by hand every Monday.

Where it fails

It fails when nobody can describe the process, because there is nothing to automate. It fails when the data lives in people's heads. It fails when the stakes are high and there is no human check: a wrong quote sent, a wrong part ordered, a wrong promise to a customer. And it fails when a business buys a tool because of what it can do rather than because of a job it has to do. A licence is not an outcome.

Start small and reversible

Pick one workflow. Measure how long it takes now, for a fortnight, honestly. Put the tool in with a person still checking every output. Run it for sixty days and measure again. Then decide, with numbers, whether to keep it, widen it or switch it off.

Everything about that plan is boring, and boring is the point. The businesses that get value from this do one unglamorous thing well and then do the next one.

You can describe the process

Write down how the job is done today, step by step, including the parts that live in one person's head and the exceptions everyone knows about. If that takes an afternoon and produces something two people agree on, you are ready to talk about tools. If it cannot be written down, the tool has nothing to do yet, and the writing down is the first project.

The data exists and someone can get at it

Spreadsheets count. Emails count. A folder of PDFs counts. What matters is that the information the process depends on exists somewhere, that you know where, and that it can be reached without a favour from whoever set up the system years ago. Data you cannot reach is data you do not have.

Someone owns it

A name, with hours in the week. Not a committee and not the owner in their spare time. Tools that nobody owns drift, stop being fed, and are quietly abandoned within a year. The owner does not need to be technical. They need to care whether it works and have the standing to change the process around it.

You know what wrong looks like

Decide before you start what a bad output is, who catches it, and what happens then. Where does a person check the work. What is the one thing it must never do on its own. How do you switch it off, and what happens to the process when you do. Answering those turns a risk into a procedure.

Your people are in on it

The people whose work changes will decide whether it works, and they will decide early. Tell them what it is for, what changes for them, and what does not. Train them before it goes live, and let the person who does the job today help shape how it is done tomorrow. Anything introduced as a surprise is resisted, and usually rightly.

Risk, privacy and the questions to ask

Customer information, employee information, drawings and pricing all end up inside these tools, and each carries obligations you already have. Ask where the data goes and who can see it. Ask what your customer contracts say about sharing their information with a third party. Ask what your staff are allowed to paste into a tool, and write that down as a policy before someone guesses.

Those are questions for your lawyer, your insurer and your customers' contracts, and we do not answer them for you. We make sure they are asked before the tool is switched on rather than after.

A thirty-day way to find out

Take one workflow that passes the first five checks. Write it down. Find its data. Name its owner. Define wrong. Tell the team. Then run the smallest possible pilot with a person checking everything, and measure. At the end of thirty days you will know more about your readiness than any assessment could tell you, and you will have spent almost nothing finding out.

The head problem

Ask how a job gets done and you will often be told it depends, and the depends lives with two people who have been there fifteen years. No tool can retrieve that, because it was never recorded. Either it gets written down, which is valuable regardless of AI, or this particular task is not a candidate yet. Say that early rather than discovering it in month four.

Contradiction beats quality

Four versions of the same policy, all findable, none marked current, is worse than having only the old one, because the tool will quote whichever it found and sound equally sure each time. Before anything else, decide which copy is the live one and remove or clearly mark the rest. Unglamorous, and it moves accuracy more than a model upgrade.

Locked in the wrong shape

Scanned paper, information that only exists inside a drawing, numbers that live in a system with no way to get them out, spreadsheets where the meaning is in the colour of a cell. All of this is technically present and practically unreachable. Find out which of your critical material is in this state before you scope anything.

The afternoon test

Write down ten questions you would want it to answer. Then try to answer them yourself using only the documents and systems a tool would have. Note how long each takes and what you had to know that was not written anywhere. If you struggle, the tool will fail, and you have learned it for the price of an afternoon.

How we help

We run that test with you against your real material, not a sample, and come back with what is ready, what needs a week of tidying, and what is not a candidate this year. Sometimes the outcome is that the data work is the whole project and AI is the easy part afterwards.

The two pricing shapes

Per seat is predictable and you pay for people who may not use it. Per use is cheap while you are experimenting and moves with volume, which is fine until a new workflow sends ten times the traffic and the bill follows. Find out which you are buying, then model your busiest realistic month rather than your average one, because the average is not the month that causes the phone call.

The costs that never appear on the quote

Connecting it to the systems that hold your information. Cleaning up the documents so the answers are worth having. Training, and the slower weeks while people learn. The internal owner's time, which is the largest line and the one nobody writes down. Add these as their own rows before you look at a total.

Cost per useful answer

Divide what you will really pay in a year by the number of times it will genuinely be used. A tool at thirty a seat across two hundred people where eleven people use it weekly is an expensive tool. This number embarrasses a lot of proposals, and it is the one to compute before the meeting rather than after the renewal.

The cost of not knowing

Usage that nobody meters produces the surprise invoice. Before signing, ask where you see spend, how often it updates, whether you can cap it, and whether you are alerted before a threshold rather than after. If the answer is a monthly statement, you have a month of exposure and no brake.

How we help

We build the five-year sheet with you, filled from your volumes and each vendor's real numbers, with the internal time named and costed. Then we compare proposals on the total rather than the front page. What fits comes first, partner or not, and we say so when a vendor is a partner.

Subscribe when the job is ordinary

Meeting notes, drafting replies, summarising, transcription, general assistance. Hundreds of companies have the same need, so products are mature, cheap and improving without you. Building your own version of a solved problem is the most common expensive mistake, and it is usually made by whoever on the team most enjoys building things.

Build when the advantage is yours

If the task runs on knowledge only your business has, your quoting logic, your inspection standards, the way your schedulers really decide, no product encodes it because no product can. That is where a build earns its place. Notice that the valuable part is your knowledge, not the software, which means the project is mostly about capturing how you work.

The middle, where most people land

Buy the product and connect it to your systems. Most of the value, a fraction of the build. The catch is the twenty per cent it does not do, which you will be tempted to close with customisation. Price that customisation as a permanent commitment, because it has to survive every upgrade the vendor ships.

The road nobody proposes

Not yet. If the process is about to change, if the data it would rely on is a mess, if nobody has time to own it, then the right move is to fix that first and buy in six months into a better market. This is the correct answer more often than it is given, because no vendor and no internal champion is incentivised to say it.

How we help

We work out which of your tasks are ordinary and which are genuinely yours, then price all three roads honestly, including doing nothing for now. We resell nothing, and what fits comes first, partner or not.

Buy, configure, integrate, or build

There is a ladder, and every rung up it costs more and ties you closer to whoever built it. Buy a finished product and use it as is. Configure one to your process. Integrate it with the systems you already run. Build something of your own. Most small businesses should stop at configure, and should be suspicious of anyone proposing to build before the first two rungs have been tried.

The questions to ask any vendor

Ask them in writing, and keep the answers with the contract.

• Where does our data go, who can see it, and is it used to train anything?
• Can we export everything we put in and everything it produced, in a form we can use elsewhere?
• What happens when it is wrong, and how do we find out?
• What does the price look like at three times our current usage?
• What does support cost, and what is the response time when it breaks on a Monday morning?
• Can we switch it off, and what do we lose when we do?

The demo trap

A demo runs on their data, on their best day. Your business runs on your data on your worst week: the scanned fax, the drawing with the coffee ring, the customer who writes orders in the subject line. Insist on a pilot on your ugliest real examples before you decide, and measure how much a person still has to fix.

Total cost

The licence is the visible part. Seats for every person who touches it. Usage charges that climb with success. The integration work to connect it to what you already run. The person inside the business who owns it, keeps it fed and answers for it. Add those up over three years and compare that number, not the monthly figure on the pricing page.

Lock-in and the way out

Every tool wants to become the place your work lives. That is fine as long as you can leave. Before signing, know how you would get your data and your process out, what it would cost, and how long it would take. A tool you cannot leave is a tool that sets its own price at renewal.

Nothing here is legal advice; the contract itself is a question for your lawyer. What we do is make sure you walk in knowing what to ask.

Where it goes and who holds it

Which countries is it processed and stored in. Which other companies does my data pass through, naming the model provider and anyone else in the chain. Is it encrypted in transit and at rest. These have crisp answers at any vendor worth buying from, and the chain question catches the resellers who have not read their own suppliers' terms.

Training and retention, asked separately

Is my content ever used to train or improve models, for me or for anyone else, and is that in the contract rather than the marketing page. How long is it kept after a request is answered. Can I require deletion, and how long does that take. Vendors frequently answer the training question warmly and the retention question vaguely, which is why you ask them apart.

Access and proof

Who at the vendor can see my content, under what process, and is it logged. Do you have an independent audit report, and may I read it rather than see the badge. Has this product been tested by an outside party against prompt injection and data leakage, and what did they find. A serious vendor has a document ready. A wrapper has a logo.

Your own obligations

Check what you promised your customers. Confidentiality clauses and data protection terms you signed years ago often bind what you may put into a third-party tool, and nobody re-reads them before adopting software. If regulated information is involved, the vendor's compliance posture must cover your obligation, not just sound impressive.

The exit

How do I get my content out, in what format, and is there a charge. What is deleted on termination, and when. If you cannot leave cheaply you have not bought a tool, you have joined something, and the price of that shows up at renewal when both sides know it.

What the job really is

Deciding which task is next and which is not worth it. Keeping the source documents current. Collecting what went wrong and getting it fixed. Answering the team's questions so they do not quietly give up. Watching the spend. Saying no to the enthusiastic suggestion that would put it somewhere it does not belong. Mostly unglamorous, mostly continuous.

Who should hold it

Someone who knows how the work actually gets done and is trusted by the people doing it. Operations, not information technology, in most businesses. They need to be senior enough to say no and close enough to the floor to hear that it is being worked around. The most technical person is rarely the right choice, because the hard part is judgment about the work.

Be honest about the hours

Half a day a week for a single tool in one department, rising sharply during a rollout. If you cannot free that, you have not decided to do this, you have decided to buy it. Name what they will stop doing. An owner with a full plate and a new title produces a tool that fades by month three.

What they need from you

Authority to turn something off. A budget line they can see. Direct access to whoever signed the vendor contract. And air cover when they tell an executive that a favoured idea is a bad candidate, because that conversation happens in every company and the owner loses it without backing.

How we help

We help you pick the person and write down what the role actually covers, then we stay on call through the first months so they have somewhere to take a question that is not the vendor who sold it. The aim is that they need us less each month.

Nobody agreed what success was

Without a number written down beforehand, the result is a discussion about impressions, and impressions do not survive a budget meeting. Decide in advance: it must handle seventy per cent of these without correction, or save this team four hours a week, measured this way. Vague pilots produce enthusiasm, and enthusiasm does not sign a purchase order.

It was tested on the easy pile

Pilots run on clean examples chosen by someone who wants it to work. Real weeks contain the scanned fax, the customer who writes in fragments, the job with three revisions. A tool that shines on the tidy set and stumbles on the real one has told you nothing, and you find out in production with an audience.

The pilot was an extra job

Given to an enthusiastic person on top of their work, it gets attention for two weeks and then whenever there is time. The result is a half-run pilot that proves little. Either free real hours or do not start, because an inconclusive pilot costs the same as a conclusive one and buys nothing.

Nobody owned the decision

Pilots that end without a named decision-maker become a report that circulates and dies. Before it begins, name who says yes or no and when. That single line does more for momentum than any feature the vendor could add.

The gap nobody scoped

A pilot uses exported files and a person pasting things in. Production needs it connected, permissioned, monitored and supported. That work is often larger than the pilot and appears only once everyone has decided it works. Scope it before you start, so the yes at the end is a yes to the real number.

Take the before, or lose the argument

For two weeks before anything changes, record how long the task takes, how many are done, and how often one comes back wrong. This is the least popular fortnight of any project and it is the only reason the after number will mean anything. Without it you are comparing against a memory, and memories bend toward whoever is paying.

Time saved is the weakest number

Four hours a week freed across eight people is thirty-two hours of nothing in particular unless you can say what filled them. Finance departments know this, which is why time-saved cases die quietly. Stronger claims: quotes going out the same day instead of in three, invoices paid a week sooner, a backlog cleared without a hire, a customer answered at nine at night.

The two numbers to keep

The correction rate, meaning how often a person has to fix the output, tells you whether it is fit for the task and whether that is drifting. The usage rate, meaning what share of the people who have it actually use it weekly, tells you whether it solved a real problem. A tool with low corrections and low usage solved a problem nobody had.

Watch for the harm that is off the sheet

Faster drafts and slower approvals because reviewers do not trust it. A team quietly maintaining a parallel spreadsheet. More output of something nobody needed more of. These show up in mood before they show up in numbers, so ask the people doing the work, plainly, a month in.

Decide the review date now

Put a date in the calendar, ninety days out, with the numbers to be presented and the named person to present them. Renewals arrive on their own; reviews do not. Most tools that outlived their usefulness did so because nobody had a scheduled excuse to ask whether they still earned their place.

Exact answers with nobody checking

Prices, quantities, tolerances, dates, dosages, anything going out unread. The tool produces something plausible when it does not know, and plausible is indistinguishable from correct at a glance. If the workflow has no human gate and no mechanical check, the right answer is a lookup in the system that actually holds the number.

Decisions that need a reason you can defend

Who gets hired, who gets credit, who gets let go, who gets the claim paid. Beyond the legal exposure, which is real and growing, you cannot reconstruct why it landed where it did. If you would have to explain the decision to the person affected or to a regulator, a model's output is not a reason.

Processes that are about to change

Automating a workflow three months before you replace the system it runs on is paying twice for something that will be thrown away. Fix the order of operations first. This one catches enthusiastic teams, because the new tool is more interesting than the sequencing conversation.

When something simpler would do

A shared template. A rule in the system you already own. A ten-line script. One conversation with the department that keeps sending the bad file. Plenty of problems presented to us as AI opportunities are a form that asks the wrong question, and no model repairs a form.

When nobody has the time

Without an owner with real hours, the answer is not yet, whatever the business case says. This is the most common no we give, and it is not a judgment about the tool. It is arithmetic about the year you are having, and it saves people a subscription they would have cancelled in month four.

Vision inspection

Mature, and the most common place AI already earns its keep on a floor. Rule-based vision has checked dimensions and presence for decades. Learning-based inspection adds the ability to catch defects nobody can write a rule for: a scratch, a smear, a fold, a colour that is slightly off. The price of that is examples. It has to be shown hundreds of good and bad parts from your line, under your lighting, before it can judge one. Budget for collecting them.

Predictive maintenance

Real, but it needs two things most plants do not have yet: sensors on the machines that matter and a history of what those sensors said before each failure. Without the history there is nothing to predict from. The honest first step is usually condition monitoring, which tells you a bearing is getting hot today, and the prediction comes later once the data exists.

Scheduling and planning

Software that sequences jobs against machines, people and due dates has existed for a long time, and the newer tools are better at handling the mess of a real week. What decides whether it works is not the algorithm. It is whether your order data, routings and machine states are accurate enough to plan from, and whether the system that runs your business can feed them across. Fix the data and the scheduling problem gets much smaller before any tool is bought.

Quoting from drawings and documents

Reading a drawing or a request for quote and producing a first-pass estimate works well for repeat and near-repeat work, where the tool has your history to learn from. It is a draft, not a quote, and a person with judgement checks it before it leaves the building. Used that way it can take hours out of every week in an estimating department.

Help for the operator

Procedures, troubleshooting steps and the answer to what do I do when this light comes on, available at the machine instead of in a binder or in one person's memory. Unglamorous, low risk, and often the best first project on a floor because it changes nothing about how the line runs.

The data question

Every item above depends on data from your own floor, and most plants have less of it than they think. Machine states not logged, quality results on paper, downtime reasons written as other. Before buying any of this, ask what data it needs, whether you have it, and how long it would take to start collecting it. The tool can wait. The collecting cannot.

The systems you already run

ERP is the system of record for the business: quotes, orders, inventory, purchasing, invoices. MES tracks jobs as they move across the floor. Beside them sit the quality system, the maintenance system with its work orders, the warehouse system, the scheduler, and the engineering data that holds drawings and revisions. Together they are the manufacturing software stack. Most small shops run two or three of them and a great many spreadsheets in between, and the spreadsheets are where the hours go.

What changed

Two things. The companies that make these systems are shipping assistants inside them, so an AI that reads your orders or drafts an invoice entry is now a feature, not a project. And an open standard, the Model Context Protocol, gives an AI a governed two-way bridge into an ERP, an MES, a quality system or a database, the same way from any of them. The software that supervises equipment on the floor is getting the same bridge, which means the path from an AI to a cell exists. The intelligence layer that connects it all is within reach of a small or mid-sized shop. The hardware was never the problem.

Read first, write later

The way in is an AI that reads and drafts, with a person deciding. It reads the order book and proposes a schedule. It reads inspection results and writes the report. It reads a request for quote and pulls your history on that part. None of that changes a record, so nothing can go wrong that a person does not see first. Write access comes later, one system at a time, once the reading has earned it, and always with someone checking what it changed.

Dispatch, not control

On the floor, an AI can decide what to run next, send the job to a cell, and react when inspection data drifts. What it should not do is run the machine. The PLC, the controller that runs the equipment, keeps real-time control, safety and motion, because that is what it is built and trusted for. AI writing controller logic directly is still weak: what it produces can fail to compile or behave oddly on a controller that scans its program thousands of times a second, and every maker has its own dialect. The defensible claim is that AI dispatches, monitors and orchestrates. Nobody should tell you it programs your robot.

An integration project, not a plug-in

The connectors that join an AI to a manufacturing system are early. Many are read-only and thin, and the good ones still need someone to decide which data they may see, which system owns each fact, who governs access, and what happens when the AI is wrong. Plan it as an integration project with a scope, an owner and a budget, not as a switch to flip. Priced that way it is still far cheaper than it was, and far more likely to work.

The question for your vendor

Before you plan around any system you have or are about to buy, ask its maker one question and listen for a specific answer: do you have, or have you promised, a way for an AI to read and act on our data? Then the follow-ups. What does it expose. What can it change. Who governs it. What does it cost. When. A vendor with a real answer will have dates and a document. A vendor without one will have a demo.

What it does well

Answers at once, at any hour. Says it is automated. Books into the real calendar against your rules. Answers the questions whose answers are written down: hours, location, services, what to bring. Takes a message with the details and sends it to the right person. Texts a confirmation.

What it must not do

Quote a price you have not written down. Promise a time the calendar does not hold. Give advice that needs a qualified person. Keep an upset caller talking. Each of those is a hand-off, written into it before it takes a call, and tested before it goes live.

The hand-off

Warm transfer to a person when one is free, a callback promise with a real time when not, and an urgent route for emergencies in trades where they happen. The hand-off is where these tools are won or lost; listen to it in the demo.

What it connects to

The phone system, the calendar or booking system, the customer record, and wherever messages should land. The page on the call at seven at night under AI for small business shows it working for a trade.

Business plan, not personal accounts

A team plan with training use off, retention stated, administrators you chose, and a contract that says so. Everyone on it, personal accounts retired for work. The lesson on settings to change on day one is the ten-minute checklist.

Connected to your own work

An assistant that can read the shared drive, the email and the job system answers questions about your business. One that cannot is a good writer with no context. Connect the sources you trust, one at a time, and check what each connection lets it see.

Standing instructions for the team

Who the business is, how it writes, what never to say, what to do when unsure, set once for everyone. The lesson on the prompt you write once covers it. It is what makes the whole team's drafts sound like one business.

What to expect

Drafting, summarising, rewriting, searching documents, first passes at spreadsheets. Time saved across the whole team in small amounts every day. Not decisions, not numbers without checking, not anything sent without a read.

Where it saves time

Sales calls, where the notes should land in the CRM. Site visits and client meetings, where the actions belong on the job. Internal meetings, where nobody wants to be the note-taker. The summary arrives before people are back at their desks.

Consent first

Recording a call has rules that depend on where the people on it are, and some require everyone to agree. Say so at the start of every recorded call, and have the tool announce itself. For sensitive conversations, clients' finances, patients, staff issues, decide whether it should be in the room at all.

Where the notes go

Into the customer or job record, not a separate app nobody opens. Check where the recordings are stored, for how long, and who can read them. A library of every call the business has had is valuable and sensitive in equal measure.

Check the actions

The summary is a draft. Whoever ran the meeting reads the action list before it is sent, because a misheard name or number becomes a task for the wrong person. Ten seconds, every time.

What it does well

Answers the questions whose answers you have written down: services, areas covered, hours, what to expect, how booking works. Asks the questions your office would ask to qualify the enquiry. Books into the real calendar where the rules allow, or collects the details and passes them on. Says it is automated, and hands to a person when asked.

What it must not do

Invent a price, a date or a policy. Give advice that needs a qualified person. Argue with an upset visitor. It answers only from what you gave it, says when it does not know, and hands over. The lessons on made-up facts and on telling it what to do when unsure are the setup checklist.

What it connects to

Your written information: the services page, the FAQ, the price list if prices are public. The booking system. The CRM or inbox where enquiries land. Without the last one it is a chat window whose leads get lost like any other.

How to know it worked

Count enquiries and bookings that came through it, and read twenty conversations a month. The ones where it should have handed over and did not are the settings to change. If nobody reads the conversations, nobody will know what it said.

How it works

The phone system sees the missed call and sends a text from the business number. If the caller replies, an AI carries on the conversation: what they need, where, when, and books it or passes it to a person with the details gathered. The caller never hears voicemail again.

The first text

Short, from the business name, honest that it is automatic, and asking one question. Not a sales message and not a link to a form. The reply rate depends almost entirely on this line, so try two versions and keep the one that gets answered.

Where it stops

Emergencies go to a person by the route you set, at once. Prices, diagnoses and promises go to a person. Texting has its own rules about consent and opting out; the tool should handle opt-out words automatically and you should know the rules where you work.

What it connects to

The phone system, which must support texting from the business number, the calendar, and the customer record.

What it needs to know

The calendar, live. The appointment types and how long each takes. Who can do what. The gaps, the travel, the deposits and the cancellation rules. It books only what those rules allow, and asks the customer one clear question when something is missing.

Changes and no-shows

Rebooking and cancelling by a reply to the reminder. A waitlist that offers the freed slot to the next customer. The no-show rate is the number that moves first; write it down before you start.

What goes to a person

New customers for complex work, anything needing a quote, the booking that breaks a rule for a good reason, and the unhappy customer. It hands over with the conversation attached.

What it connects to

The booking or job system that owns the calendar, the phone and messaging, and reminders.

Asking

A text or email after the job is done, with a direct link, sent to every customer rather than only the ones you think are happy. Asking only happy customers, or offering something in return, breaks the review sites' rules and in many places the law. Ask everyone, honestly, once, with one reminder.

Replying

A drafted reply to every review in your voice, specific to what the customer said, for a person to approve with one tap. Replies are read by the next customer more than the one who wrote the review.

The bad review

Flagged to a person at once, with the job history attached. The reply is written or at least read by the owner, and the customer is contacted. A tool that auto-replies to a complaint makes it worse.

What you learn

A monthly summary of what customers praise and complain about, across every review. Useful for the business, not just the listing.

The sequence

An answer to every new enquiry within minutes. For every open quote, a follow-up after a couple of days, another after a week, and a last one that makes it easy to say no. Each drafted from the quote and the customer's own words, in your voice, and stopped the moment they reply.

What it says

A question, not a push: did the quote answer what you needed, would a different option help, is the timing wrong. No invented urgency, no discounts it was not given. The lesson on one good example is the fastest way to set the tone.

When a person takes over

The moment the customer replies with anything but a simple yes or no, and for every large quote. The tool keeps the thread going; the salesperson closes.

What it connects to

The CRM or job system that holds the quotes, the email and texting, and the calendar for the booking at the end.

The reminders

Before the due date, on it, and after it on a schedule you set, each with the invoice and a payment link. Firmer with time, never rude. Stopped the moment payment lands.

The replies

Can you resend it, what is this charge, we need a purchase order number: drafted answers from the invoice and the job, sent after a person approves at first, and on their own for the simple ones once you trust them.

Where it stops

Disputes, hardship, large amounts, and your best customers go to a person. Payment plans, write-offs and anything that changes what is owed are decisions, not reminders.

What it connects to

The accounting system that holds the invoices and records the payments, and the email or texting it sends from.

How it works

It searches your documents for the passages closest to the question, reads them, and answers from them with the source shown. It is only as good as what it can find, which is why the lesson on retrieval step by step is worth ten minutes before buying one.

The documents first

One home for files, one current version of each, old versions archived where the tool does not look. A tidy drive gets right answers; a drive with three price lists gets a confident wrong one.

Who can ask what

The tool can see whatever the account it uses can see. Connect it with permissions that match your people, or the junior member of staff will be told what is in the salary folder, helpfully.

Checking answers

Every answer carries the document and the passage. Staff read the passage, not the paraphrase. Anything not in the documents gets an I don't know, and you test that before it goes live.

Asking well

The questions your best person would ask, in plain language, one at a time, skipping what does not apply. Photos where a photo says more: the fault, the damage, the nameplate. A summary at the end for the customer to confirm.

Into the system

The answers written into the right fields of the CRM, the job system or the practice system, with a one-paragraph brief for the person who picks it up. No retyping, and nobody asking the customer the same thing twice.

What it should not collect

Card numbers, passwords, and anything your rules say needs a secure form or a person. Health and financial details only through a tool with the agreement your rules require, as the healthcare and finance pages under AI for small business explain.

From real work

The job photos, the reviews, the new service, the seasonal reminder: raw material the business already has. Drafted into posts, a short newsletter, a page update. Real specifics beat generic content every time, for customers and for search.

In your voice

Standing instructions on how the business writes, and three examples of your best posts or emails. The lessons on the prompt you write once and on one good example are the whole method.

The checks

Every claim true, every photo yours with permission, no testimonial invented, no customer named without asking. A person reads everything before it goes out under the business name. Marketing is the one place a fluent invented detail is published to everyone at once.

Reading the request

Requests arrive as emails, PDFs, drawings, spreadsheets and customer portals, and the first hour of every quote is spent finding the quantity, the material, the tolerances and the date inside them. An AI reads all of that and puts it in one place, in your format, with the gaps flagged. The win is intake, not judgement. It gets the estimator to the estimating faster.

Quoting from your own history

For repeat and near-repeat work, the tool finds the last similar jobs and what they actually took, and drafts a number from them. That is where it is strongest, because the evidence is yours. Ask it to show the comparables it used. A draft with its reasoning beside it is something an estimator can correct in a minute; a bare number is something they have to redo.

Where it goes wrong

New geometry, a material you have not run, a tolerance tighter than anything in your history, a customer whose drawings lie. On those the tool produces a confident number from thin evidence, and confidence is the problem. Make it say when it has fewer than a handful of comparables, and route those quotes to a person first, not last.

The signature stays human

A person checks every quote before it leaves. That is not a limitation to engineer away; it is the control that makes the speed safe. Measure after the fact: hit rate, margin against estimate, the jobs that ran long. The tool only learns from actuals if you record them, so the discipline of closing out a job with real hours is what makes next year's quotes better.

Where it plugs in

The quote module in your ERP, the customer list, the estimating spreadsheet, the job history with actual hours. Read-only first: the AI reads those and drafts into a queue. Only when the drafts are consistently right does it earn the right to create the quote record itself, and even then a person releases it. The connection to your data matters more than which model sits behind it.

On arrival

The email, the attachments and the drawing read together. The parts listed with quantities and dates. The drawing numbers and revisions pulled. The specs, finishes, certifications and packaging requirements found wherever they were hidden. The terms noted. All of it in one summary on one page, with each item pointing at where it came from.

What it flags

The revision you have never seen. The spec you do not hold. The delivery date that is before your lead time. The quantity that is ten times the last order. The customer whose last three quotes went nowhere. Each flag is a decision the estimator makes in a minute instead of discovering on day three.

Why it changes the quote

The estimator starts from knowing, which means the quote starts on the day the request arrives instead of the day someone got to it. The requests not worth quoting are declined the same day, politely, and the hours go to the ones that are. Turnaround is the number to watch, and it moves first.

The three closest jobs

By material, size, operations, tolerance and quantity, from the job history in the ERP: the part number, the quoted hours, the actual hours, the material used, the margin. Put beside the new request so the estimator sees at a glance that this family runs over on the second op every time. The draft quote uses the actuals, and the estimator adjusts.

Why history beats the card

The rate card is what you meant to charge. The history is what happened. Shops that quote from history find their quotes drift toward true, which means fewer jobs lost on price and fewer won at a loss. The drift is small per quote and large per year.

What it must never guess

A material price it has not looked up today. A lead time nobody confirmed. A tolerance it inferred. The tool reads prices from the current list and hours from real jobs; where it has neither, it leaves a marked blank. The estimator's initials go on the quote, and the estimator has read the blanks.

The five

The material price from a list that changed. The drawing revision from the last order rather than this request. A quantity break that looks like the pattern of your other quotes and is not on your list. A lead time assumed from the usual, when the shop is full. A customer term dropped because the tool was drafting from a different customer's template.

The checks

Price against today's list, not memory. Revision against the request. Breaks against the price policy. Lead time against the schedule this week. Terms against the customer's file. Five glances. The tool helps by marking anything it did not read from a source, and a good one shows the source beside each number.

The review that holds up

Narrow it to the five. Time it. When the reviewer has not changed a draft in a fortnight, ask whether the tool is perfect or the reviewer has stopped, and the answer is never the first one. The lesson on when the reviewer stops reading is short and it applies here more than anywhere in the office.

Order entry

Orders arrive as emails, attachments, portal downloads and the occasional fax. An AI reads them, finds the part numbers, quantities, dates, ship-to and terms, matches them to your customer and item records, and drafts the sales order. The person who used to type it now checks it, which takes a tenth of the time, and catches the things the AI flags: a part number it could not match, a date that is before today, a quantity that is ten times the usual.

Supplier invoices

Matching a supplier's invoice to the purchase order and the receipt is the same job in reverse. The AI reads the invoice, finds the order, checks the quantities and prices against what was received, and drafts the entry. Matches go through on a person's approval; mismatches go to a person first. Most shops discover that the mismatches are where the money was leaking, and that finding them was the whole point.

Exceptions to a person

The design rule for all of this is that the AI never guesses in silence. A field it cannot read with confidence, a customer it cannot match, a price that differs from the contract: each is a question to a named person, not a value filled in. Set that rule up front and the system stays trustworthy as the volume grows. Skip it and the first quiet error costs more than the year of typing it saved.

What still needs a signature

Credit decisions. Anything that changes a price. A customer's terms. The release of an invoice to a customer. Keep those with a person, on purpose, and write it down so nobody later automates them because it looked easy. The point is not to remove people from the paperwork; it is to give them the paperwork worth their attention.

Where it plugs in

The order entry and accounts payable modules of your ERP, the customer and supplier records, the mailbox the orders arrive in, and the scanner. Read and draft first. Creating records comes once the drafts are consistently right, and even then a person releases them. If your ERP has a built-in assistant that does some of this, start there; if it does not, this is one of the simplest integrations to add on.

From the document to the draft

The PDF or email read: customer, ship-to, lines, part numbers matched to yours, quantities, prices checked against the customer's agreed list, dates, the PO number. Drafted into the ERP as an order awaiting release, with anything that did not match flagged in red. The desk opens a draft with three things to look at instead of a document with forty things to type.

Where the errors were

The transposed digit in a quantity. The customer's part number typed as yours. The wrong ship-to from the dropdown. Each one used to be found at the dock or by the customer. Now the tool matches against your records and flags what does not fit, and the errors that remain are the ones in the customer's document, which is a different conversation.

What gets released

Everything, by a person, in a glance: the flags looked at, the draft released. New customers, new ship-tos, prices that differ from the agreement and quantities far from the usual always get a proper look. The clean orders take seconds each, which is the whole return.

The match

The invoice read: supplier, number, lines, quantities, prices, total, terms. The PO found. The receipt found. Line by line: quantity invoiced against quantity received, price invoiced against price ordered. A match within tolerance goes through for approval. A mismatch goes to a person with the three documents side by side and the difference highlighted.

What it finds

The price that rose without notice. The freight line that was not on the quote. The quantity invoiced for the full order when half arrived. The invoice sent twice. In most offices the first month of matching everything pays for the year, because the discrepancies were always there and nobody had time to look.

What a person does

Decides the mismatches. Approves the batch. Talks to the supplier about the pattern. Payment is a stop point: the tool prepares the run, a person releases it, and the early-payment discount the office used to miss for want of time gets taken.

What goes to the pile

The document the tool could not read. The order that did not match a price. The invoice with no PO. The customer the system has never seen. Anything above a value you set. Anything where the tool's confidence was low, which a good tool can say. The rule is simple: when in doubt, a person, and the tool is told to doubt generously at first.

How it arrives

With the document, the reason it was flagged, the tool's best guess, and the one-tap fix if there is one. In the place the person already works: the ERP, the inbox, the queue. A bare notification that says exception is worse than no tool at all, because now the person has to reconstruct what a tool already knew.

How it shrinks

Each exception is a lesson: a new customer added, a supplier's odd format learned, a tolerance widened. The pile should get smaller each month and someone should be watching the count. A pile that grows is the tool telling you something changed upstream, and it is the earliest warning you will get.

What lives where

The model, the drawings, the bill of materials and the revision history live in engineering, usually in a product data system or a well-kept folder. The floor works from a printed copy, a PDF on a shared drive, or a photo on someone's phone. The gap between the two is where the wrong revision gets built, and it is the first thing to close before any tool is worth buying.

Drafting instructions from the model

Given the model and the bill of materials, an AI can draft an assembly sequence, a tool list, torque values, the checks at each step, and the photos to go with them. The draft is a starting point. The person who knows the job edits it, and nothing reaches an operator until someone with authority approves it. Written that way, instructions that took a week take an afternoon, and they get written at all.

Revision control is the whole game

An instruction generated from the wrong revision is worse than no instruction, because it is confident and wrong. Before any of this, every part needs one place where the current revision lives and a rule for how a change gets there. If two people would give different answers to which revision is current, that is the project, and it is not an AI project.

At the machine

The instruction on a screen at the station, the current revision by construction. The operator asks what to do when a step fails and gets an answer drawn from your approved documents, not from the internet. Unglamorous, low risk, and one of the best first projects on a floor, because it changes what happens on the second shift when the person who knew is not there.

Where it plugs in

The engineering data system, the item master in the ERP, the approved-document store in the quality system. Read-only first. The AI drafts from them and never writes back into engineering; a person releases a revision. That boundary is what keeps the drawing the drawing.

A typical shop

The model in the CAD system. The drawing in the PLM, or a folder, or both. A PDF of it in the quote email. A print on the bench, marked up in pencil. The spec it references in a binder. The work instruction in a document nobody has opened since the person who wrote it left. Five places, one truth, and the tool has to be pointed at the truth.

Naming the truth

One afternoon, with the engineer who knows: for each kind of document, where the current version lives, and the rule that says so. The PLM holds the drawing and the revision. The folder holds the specs, by number. The instruction lives with the routing. Written down as a page, and the tool is connected to those places and no others. The pencil-marked print is a document to reconcile, not a source.

Why this comes first

Every page under this application assumes the tool reads the right document. An instruction drafted from the wrong model, an answer given from the wrong revision: each is worse than no tool, because it is fluent. The lesson on retrieval says the answer is only as good as the fetch, and the fetch is only as good as what it is pointed at.

The draft

From the drawing and the model: the features, the datums, the critical dimensions. From the routing: the operations and the machines. From the three most similar instructions you have: the structure, the wording, the checks at each step. Drafted in your template with the views from the model placed at the steps that need them. What took a day per part takes an hour, and the format is the same for all of them.

What gets checked

The engineer reads it against the drawing: every dimension, every tolerance, every step. The operator runs the first part from it and marks what was unclear. Both before it is released. A drafted instruction that goes straight to the bench is the fluent mistake this whole site warns about, on the one document the floor trusts.

What it must never contain

Safety steps the tool wrote. Isolation, guarding, lifting, anything that protects a person: written by the person responsible for it, in the words the plant uses, and the tool told to leave a marked block for them. The tool drafts the how of the part. The how of staying safe is not its to draft.

How they get out of step

The revision changes in the PLM and the print on the bench does not. The quote was made to the old one and the order references it. The instruction was written for rev B and nobody reissued it. The PDF in someone's inbox is whatever it was when it was sent. Each is a copy that drifted, and the tool will read a copy as happily as the truth.

What the tool can do

Read only from the system that holds the revision, never from a copy. Show the revision it read on every answer. Flag when an order, a quote, an instruction or a print references a revision that is not current. Chase the reissue of the instruction when the drawing changes. It cannot know which revision the customer meant; it can make sure everyone sees which one it used.

The discipline

One system holds the revision. The tool reads that one. Prints on the floor carry the revision and are pulled when it changes. A revision release triggers the instruction review. None of this is new to a shop that runs a quality system; what is new is that the tool makes the drift visible instead of quietly reading through it.

Counts you can trust

No AI fixes a bad count. If the system says forty and the shelf says twelve, every decision downstream is wrong, and a smarter tool makes it wrong faster. Scanning at receipt, issue and shipment comes first, and a cycle count that people actually do. Once the number in the system is the number on the shelf, the system can start to reason about it.

Reorder points that move

A reorder point set once and never touched is wrong most of the year. A tool that watches demand, lead time and the season, and suggests when the point should move, is a real improvement, and the suggestion is the right shape for it. The buyer decides. The buyer knows the supplier is closing for two weeks in August; the tool does not until someone tells it.

Supplier quotes in one place

Requests for quote go out and the replies come back in every format a supplier can invent. An AI reads them, pulls out price, lead time, minimums and terms, and lines them up in one comparison. That is hours a week for a small purchasing desk, and the judgement stays where it was: a person picks the supplier, and knows why.

The dock

Pick lists, packing slips, labels, the advance notice the customer wants, the carrier booking. Most of it is the same order typed again. Done once from the order, with the exceptions routed to a person, the dock stops being the place where the day runs late. A robot on the dock is a separate decision, and a robotics one.

Where it plugs in

The purchasing and inventory modules of the ERP, the warehouse system if you run one, the carrier accounts. Read first: the AI drafts purchase orders and shipping documents into a queue. Writing to stock levels comes last, if ever, because that is the number everything else depends on.

Finding where it is wrong

The tool watches receipts, issues, sales, returns and the counts that do happen, and points at the products whose movements do not add up: sold more than it had, issued to a job that never used it, received and never booked. Those go on today's count list. The count is entered on a phone at the shelf, and the variance is traced back to the transaction that caused it.

The three habits

Receipts booked late or against the wrong line. Issues to jobs taken from stock without a transaction. Returns put on the shelf and never put in the system. Every plant has a version of these, and the variance analysis names which. Fixing the habit is a conversation, not a feature, and it does more for the count than any tool.

When to trust it

When the count list gets short and the variances get small. Then the reorder points, the availability and the purchasing proposals built on the count are worth reading. Before that, they are arithmetic on fiction. Most shops get there in a quarter if someone owns it.

What it proposes

For each item: the rate of use over the last year and the last quarter, the supplier's measured lead time, the variability of both, and the point and quantity that follow from them at the service level you set. Shown beside the current point, with the difference and the reason. The buyer sees the twenty items where the change matters, not the two thousand.

What the buyer knows

That the big customer's order in the spring will not repeat. That the supplier has just changed hands. That the item is being phased out. The tool proposes from the numbers; the buyer overrides from the knowledge, and the override is recorded so the tool does not propose it again next week.

The two numbers

Stockouts, and stock sitting more than a year. Both should fall. Write them down before the tool starts, look at them quarterly, and the page called how do I know it worked has its answer. If only one falls, the service level is set wrong, and that is a dial, not a failure.

The comparison

Each quote read: unit price, quantity breaks, lead time, minimum order, freight, currency, validity, payment terms. Put side by side on one page, landed cost calculated at your usual quantity, the differences that matter flagged: the minimum you cannot meet, the freight that doubles the small order, the lead time that misses your date. The buyer decides from a page, not from three inboxes.

What it must not assume

That freight is included when the quote does not say. That the currency is yours. That the price holds past the validity date. The tool reads what is there and marks what is missing; a marked blank is a question for the supplier, and a guess is a surprise on the invoice.

Afterwards

The quotes filed against the item and the supplier, searchable, so that the next request for that part starts from what was quoted last time. Over a year, that history is what tells the buyer which supplier drifts up between quotes, and that is worth more than any single comparison.

What the scheduler knows

Which machine really runs that part well. Which operator can set up the tricky job. Which customer will accept a day late and which will not. Which supplier's material always arrives short. None of that is in the ERP, and a scheduling tool that does not know it will produce a beautiful plan that the floor ignores. The first project is writing the unwritten rules down, and it is worth doing whether or not a tool ever arrives.

Sequencing against the real week

Given clean orders, routings, machine capacities and due dates, a modern scheduler produces a sequence that honours the constraints, and re-runs it in seconds when something changes. That is genuinely useful, and the AI layer on top makes it easier to ask why: why this job first, what happens if we pull that order in, what breaks if the mill is down until Thursday. The plan is a proposal. The person with the unwritten rules approves it.

Reacting to the floor

The value multiplies when the schedule can see the floor: a machine state from the MES, a downtime reason, an inspection result. An AI that notices a cell has stopped and proposes the re-plan before anyone walks over is the difference between a schedule and a wall chart. That needs the floor data to be flowing, which is the same data every other plant-floor AI depends on, so the investment pays more than once.

Dispatch, not control

The step after scheduling is sending the job: the program to the machine, the recipe to the cell, the work order to the operator's screen. An AI can do that dispatching, and can react when inspection data drifts. What it must not do is run the equipment. The PLC, the controller that runs the machine, keeps real-time control, safety and motion, because that is what it is built and trusted for. AI dispatches, monitors and orchestrates. It does not program your robot, whatever the demo implies.

Where it plugs in

The order book and routings in the ERP, machine states and job progress from the MES, and the scheduling tool if you already have one. Read first: the AI proposes and explains, a person releases the plan. Dispatching to a cell comes later, one cell at a time, with the same rule: the person decides which jobs the system may send on its own.

In the system

The orders, the due dates, the routings, the standard times, the machines, the shifts. Usually present, often stale: the routing from the first time the part ran, the standard time nobody updated when the fixture changed. The tool reads these and the schedule it proposes is only as good as they are. Check them against a week of actuals before believing the proposal.

In the planner

That one operator can run the big press and is on days. That the setup from part A to part B is an hour and from B to A is ten minutes. That the customer who calls the owner comes first. That the old lathe cannot hold the tolerance on that job. None of it is written down, and all of it decides the week. Write it down as rules, one afternoon, with the planner. That afternoon is the project.

Why it is worth doing anyway

A shop whose scheduling knowledge lives in one head is a shop one resignation from chaos. Getting the rules out makes the tool possible and the shop safer, and the second is worth more than the first. The tool then proposes from the rules, and the planner corrects the proposal instead of building from nothing.

The real week

Who is in, from the roster. Which machines are down, from maintenance. Which jobs are waiting on material, from purchasing. The setups, chained so the changeovers are short: the same family back to back, the light colour before the dark. The due dates weighted by what the planner said matters. From that, a sequence for each machine, with the load shown and the jobs at risk flagged.

What the planner changes

The order of two jobs because the customer called. The machine for a job because the operator on it is better. The overtime that the tool did not know was available. Each change is made on the draft in seconds, and the tool re-sequences the rest. What used to be an evening of building is a morning of adjusting.

What a good proposal shows

Its reasoning: why this job before that one, which constraint bound, what would happen if a machine came back. A tool that shows a sequence without reasons is a tool the planner cannot trust or correct, and it gets ignored by the second week. Ask to see the why in the demo.

Reading the change

From the MES, the machine monitoring, the operator's note or the planner's own message: a machine down, a job over its time, a material shortage, a rush order. The tool re-sequences from the point of change, shows what moved, and lists the jobs now at risk of their date, with the options: overtime, another machine, a call to the customer.

Dispatch, not control

The tool proposes and the planner publishes; the floor sees the published sequence on the screen or the sheet. The tool never changes a machine, a program or a parameter, and never sends work to the floor without a person releasing it. That line is where this practice stands, and it is also where the floor's trust in the schedule comes from.

Telling the floor

A schedule that changes without the floor knowing is worse than one that does not change. The published sequence goes to the cell the moment it is released, with what changed and why, in a sentence. The operator who understands why the job moved runs the new sequence; the one who was surprised runs the old one.

Results as they happen

Whether the measurement comes from a gauge, a coordinate measuring machine or a vision station, it is a number with a time on it. An AI watching those numbers sees the drift before a part is out of tolerance: the dimension creeping toward the limit over a shift, the one fixture that runs hot after lunch. That is the oldest idea in quality, statistical control, with the watching done continuously instead of at the end of the week.

From the data to the report

First-article reports, certificates of conformance, the customer's monthly quality pack, the corrective action write-up. All of it is the same data arranged the way a particular reader wants it. An AI drafts those from the records, in your format, and a person reviews and signs. The hours saved are real, and the reports get better, because they are written from what was measured rather than from memory the week after.

Root cause, with a person

When something goes wrong, the AI can pull together everything it knows about the lot: which machine, which operator, which material batch, what the inspection said at each step. That is the evidence. The reasoning about it is still a person's job, with the AI as the one who fetched the files. Tools that promise to find root cause on their own are, for now, tools that find correlations; treat what they say as a lead, not a finding.

What still gets signed

Deviations. Concessions to the customer. The release of a lot. The close of a corrective action. A quality system that a customer or an auditor trusts is one where a named person took those decisions, and an AI that drafts the paperwork for them changes nothing about that. Write down which decisions stay human before the tool arrives, and the audit conversation stays simple.

Where it plugs in

The quality system's records, the inspection equipment's outputs, the MES for the job and lot context, and the ERP for the customer and the order. Read first, always. The AI reads and drafts; it does not change a result or release a lot. The measurement side of this, gauges and coordinate measuring, has its own page.

Every number as it lands

From the gauge, the CMM, the operator's entry: each result read against the limits and against the run so far. The tool flags a trend before it is a non-conformance, a step change after a tool change, a pattern that lines up with a shift or a material lot. Flagged to the inspector and the operator, in a sentence, with the chart, in the moment it can still be acted on.

What the inspector decides

Whether to stop. Whether the drift is real or the gauge. Whether the pattern on nights is the operator, the temperature or the coolant. The tool sees the numbers; the inspector sees the shop, and the decision to stop a machine is a person's. The tool's job is to make sure the person sees it at part forty-one.

What it needs

Results recorded as they are taken, digitally, against the part and the feature. A shop that writes results on a traveller and types them in on Friday has nothing for the tool to read until Friday. The first project is often the gauge that sends its reading rather than displays it, and the page on gauge data is about that.

The report, drafted

The first-article report in the customer's template. The certificate of conformance with the lot, the material cert and the results. The inspection report for the shipment. The non-conformance report with the results, the disposition options and the history of that feature. Each drafted from the data in the quality system, in the format that customer requires, with every number pointing at the record it came from.

What still gets read

Every number against its source, once, which the tool makes quick by showing the source. The disposition on a non-conformance, which is a decision. The statement of conformance, which is a promise. The engineer signs what they have read, and the reading is faster because the assembling is gone.

What must never be drafted

A result that was not measured. A certificate for a lot the data does not cover. The tool drafts from what is in the system and leaves a marked blank where the system has nothing, and a marked blank on a certificate is a stop, not a nuisance.

The evidence, lined up

Every occurrence of the defect against everything the systems know about it: the machine, the operator, the shift, the material lot, the tool change, the maintenance done, the temperature if it is logged, the day it started. Cross-tabulated and charted, with the correlations that stand out named as correlations. A page the engineer used to build over four days, in a morning.

Why it must not name the cause

A correlation is not a cause, and a model presenting one as the other in a confident voice is how a plant replaces the wrong bearing. The tool says: the defect appears on nights, after the lot change, on machine three. The engineer walks to machine three on nights with the lot in hand, and finds the fixture that shifts when the coolant warms. That step does not happen at a desk.

Recording it

The cause, the evidence, the action and the check that it worked, on the non-conformance, in plain words, so that the next time this defect appears the engineer starts from the last investigation rather than from nothing. The tool drafts the record from the investigation; the engineer signs it, and the plant keeps the memory.

Writing the inspection program

Programming a coordinate measuring machine for a new part has always been slow, skilled work: which features, in what order, with what probe, from which datums. Software now drafts that program from the model and the drawing's tolerances, and the newer tools use AI to pick the sequence and probe paths. A metrologist reviews it and runs it on the first part. That turns a day into an hour for routine parts, and it is one of the clearest wins on the floor.

Gauge data as it happens

A hand gauge with a wireless output, a fixture gauge on the line, a vision station: each produces numbers that used to go on a sheet. Collected as they happen, they show drift before it becomes scrap, and an AI watching them does the watching without getting bored. The value is entirely in the collection. If the readings still go on paper, no tool can see them.

Vision measurement

Camera-based measurement has moved from presence and dimensions to catching the defect nobody could write a rule for: the scratch, the smear, the slightly wrong colour. It needs examples, hundreds of good and bad parts from your line under your lighting, and it needs the lighting and the fixture to stay as they were when it learned. Budget for the examples and for keeping the setup honest, and it earns its keep.

Calibration stays human

Traceability, the chain that connects your gauge to a national standard, is a paper trail with named people on it. Nothing about AI changes that, and a tool that claims to make calibration automatic is describing reminders and records, which are useful, not the calibration itself. Keep the chain, keep the schedule, and let the AI draft the records and flag the gauge that is due.

Where it plugs in

The measurement software on the coordinate measuring machine, the gauge collection system, the vision station's outputs, and the quality system where the results should end up. Read first. The AI reads results, drafts programs and reports, and flags drift. A person approves the program, signs the report and owns the calibration.

From the drawing to the plan

Each toleranced feature listed with its datum reference, its tolerance and the drawing zone it came from. Grouped by the datum scheme, sequenced sensibly, with the critical ones flagged from the drawing's own marks. If the model carries the tolerances, they are read from the model; if the drawing is a PDF, the text is read and the geometry is not, and the tool says which. The programmer starts from a checked list rather than a blank screen.

What it cannot judge

Which probe. The approach that will not crash. Whether the fixture hides a face. Whether the feature is even measurable on this machine. Those are the programmer's, from experience of the machine and the part, and the tool's draft stops where they begin. A plan is not a program, and the difference is the metrologist's whole trade.

Proving it

On a part, with a known result, before it measures anything that ships. A program drafted faster is still a program that has to be proved, and the tool does not shorten that step; it shortens the day before it. The record of the proof goes with the program, which the page on calibration and sign-off describes.

The cable

Most gauges made in the last twenty years can send their reading: a cable, a wireless dongle, a small box. The reading lands in the record against the part, the feature, the gauge and the time, with no transcription. The CMM already does this; the hand gauges usually do not, and they are where the transcription errors live. This is a small project with a large effect, and it is where a metrology AI project honestly begins.

What it makes possible

The drift seen at part forty-one, from the quality pages. The gauge that reads differently on nights, visible because every reading carries the gauge and the shift. The report drafted from data instead of from a traveller. The calibration history that shows a gauge going before it goes. None of it exists without the reading arriving as data.

What it costs

Less than the tool that reads the data, usually. A cable or a dongle per gauge, an afternoon per bench, and the habit of using the connected gauge rather than the one in the drawer. The habit is the expensive part, and it is the part a supervisor has to own.

The camera that measures

Calibrated optics, controlled lighting, a fixed part position, and software that finds edges and computes dimensions with a known uncertainty. Proved against a reference part, checked on a schedule, and reported like any other gauge. This is real, mature, and it measures. Some of it uses learned models to find the feature; the measurement is still geometry.

The model that describes

Looks at an image and produces words: the surface appears scratched, the weld looks complete, there is something in the corner. Useful for sorting photos, flagging the obviously wrong part for a person, drafting the note. Not a measurement, not calibrated, not repeatable to a tolerance, and a number it gives you is a guess. The lesson on drawings and photos is the longer version.

Which you are being sold

Ask for the measurement uncertainty and the proving procedure. A vision system vendor has both on a sheet. A model-describes-a-photo vendor changes the subject. Ask what happens to the part the system is unsure about: the honest answer is that it goes to a person, and a good system flags its own doubt. Both tools have a place; the place is not the same, and the price should not be either.

From the fault to the work order

An operator reports a fault by radio, sticky note or shout. Someone types a work order later, or does not. An AI that takes the report in plain words, at the machine, and drafts the work order with the machine, the symptom, the likely parts and the last time this happened, closes the gap where most maintenance history is lost. The tech corrects and accepts it in seconds. A year later the plant has a history it never had.

Preventive schedules that follow use

A schedule on the calendar services the machine that ran two hours a week and the one that ran two shifts a day on the same day. Preventive maintenance driven by hours, cycles or output, read from the machine, is a plain improvement, and the maintenance system you may already have can do it once the data flows. The AI layer helps with the exceptions: the machine whose actual behaviour says the interval is wrong.

Condition data and early warning

Vibration, temperature, current draw, the sound a bearing makes before it goes. Sensors that capture those, on the machines that matter, let an AI notice the change before the failure. This is where prediction is real, and it needs two things most plants do not have yet: the sensors and a history of what they said before each failure. Without the history there is nothing to learn from. Condition monitoring with alarms is the honest first step, and it earns the history.

The knowledge in the tech's head

Every plant has one person who knows why that press trips on a humid day. When they leave, the plant learns it again the expensive way. An AI that can answer a tech's question from the machine's manual, its own history and the notes the senior tech left is worth more than most prediction projects, and it costs less. Capturing the notes is the work.

Where it plugs in

The maintenance system, where the work orders and the schedule live. The machines, through whatever monitoring exists. The ERP for the spare parts. Read first: the AI drafts work orders and flags conditions, and a person raises the order, schedules the job and closes it. The system never takes a machine out of service on its own.

The first sentence

Said into a phone at the machine, or typed, or a photo of the display. The tool identifies the asset from the location or the tag, writes the fault in plain words, pulls the last three times this machine reported something similar with what fixed it, lists the parts those repairs used and whether they are in stock, and drafts the work order with a priority from rules you set. The maintenance lead assigns it in a glance.

What the tech gets

The work order on their phone with the history attached: last time it was the sensor, the time before it was the belt, here is the manual page, here is the part number and the bin. The tech starts from knowing, which is the difference between an hour and an afternoon, and the knowledge came from the record, not from the one person who remembers.

What it never touches

The machine. The tool reads and drafts; it does not reset, adjust, start or stop anything, and it does not decide the isolation or the safety procedure. Those are the tech's and the controls engineer's, and the tool's job ends at the work order and the manual page.

From use, not the calendar

Hours run, cycles counted, from the MES, the machine, or the operator's log. The fault history: what fails, how often, after how much use. From that, a proposed schedule per machine: this one early, that one later, and a list of tasks that in two years have never found a problem, which are candidates for a longer interval. The lead reviews the proposals monthly, not the calendar daily.

What the lead decides

That the machine on the critical job gets serviced early whatever the numbers say. That the task the tool wants to stretch is the one that once saved a spindle. That the data on the old press is wrong because the counter is broken. The tool proposes; the lead knows the plant, and the override is recorded.

What the data has to be

Real. Hours and cycles that are actually recorded, faults that are actually logged, work orders that are closed with what was found. Most plants have half of this, and the first quarter is spent getting the other half, which is the page on the tech's knowledge. Without it, a use-based schedule is a calendar with extra steps.

At the close of the work order

The tech says what they found, what they did, what they replaced, what they would check first next time, and the trick. The tool turns it into a structured record on the work order: cause, action, parts, notes, in plain words, tagged to the asset and the fault. Forty seconds, at the machine, with the phone. The tech reads it once and taps done.

What it should capture

The cause, not just the fix. The thing that looked like the cause and was not. The part that was wrong in the manual. The order to do things in. Anything a new tech would ask the old one. A prompt that asks for those four things gets them; a blank notes field gets the word fixed.

What it is worth

The next fault on that machine starts with the last three fixes on the tech's phone, which the page on the work order describes. The new tech learns from the record. The day the senior tech leaves, the machines keep their memory. Start now; the value compounds and it cannot be recovered later.

Status, from the system

Where is my order. Has it shipped. What was the lot number. When is the next delivery. An AI with read access to the ERP and the shipping records answers those in seconds, in an email or on the phone with a person listening in, and gets them right because it is reading, not remembering. That alone removes a large share of the interruptions from the people who should be quoting and scheduling.

The inbox, sorted

A shared inbox where orders, complaints, supplier questions and spam arrive together is where messages die. An AI that reads each one, decides what it is, and routes it to the right queue with a summary is unglamorous and transforms the day. It should never delete, and it should say when it is unsure. A misrouted complaint is a small cost; a lost order is not.

Replies in your voice

For the routine reply, the AI drafts and a person sends. Give it your real past replies to learn from, and it will sound like you rather than like a call centre. Give it the rules: what you never promise, which customers get which terms, when to escalate. The draft is worth having because it is quick to correct; the send button stays with a person until the drafts have been right for a long time.

What a person still sends

The apology when something went wrong. The conversation about a late payment. Anything that changes a price or a promise. The message to the customer who has been with you for twenty years. Decide those up front, write them down, and keep them human on purpose. A customer can tell, and the ones who can tell are the ones worth keeping.

Where it plugs in

The ERP for orders and shipments, the mailbox, the phone system if it records, and the customer records wherever they live. Read first. The AI answers status questions from the records, drafts the rest, and never sends on its own until you decide it may, one message type at a time.

The fact

The order found from the reference or the sender. The stage read from the schedule or the MES: in production, at inspection, packed. The shipment read from the carrier. One sentence, plainly, in your voice, with the reference. Not an estimate; a state. If the system does not know, the tool says so and hands to a person rather than guessing.

What it must not promise

A date the schedule does not hold. A fix by Friday. Anything that commits capacity. The tool reports what is; a person promises what will be. The list of things it may not say is written before it takes its first message, and it hands over on any of them.

Before they ask

When a due date moves in the schedule, the customer gets a drafted notice with the new date and the reason, released by a person because it promises. When it ships, the tracking goes out without anyone being asked. The customers who used to call twice a week stop calling, and the person on the desk notices the quiet first.

On arrival

Classified: status question, order change, complaint, quote request, invoice query, supplier, noise. Matched to the customer and the order from the reference and the sender. Routed to the right person or queue. Urgency judged by rules you set and by the words: the customer who says cancel is urgent whatever the subject line says.

Drafted where the answer is known

Status, from the system. The invoice copy, attached. The delivery address, confirmed. Each drafted in your voice for the person to send in a click, or, after a month of watching, sent by the tool for the categories you release. The complaint and the cancellation are never in that set.

What the person does

Reads the six. Answers the complaint properly, the same day, because the other hundred and ninety-four did not need them. Reads a sample of the drafts each week. The inbox stops being a backlog and starts being a queue, and the number to watch is the age of the oldest unanswered message.

The examples and the rules

Ten replies you are proud of, stripped of customer details, as the pattern. The standing instructions: the length, the greeting, the sign-off, plain words, no exclamation marks, and the list of what never to say without a person. From those, the tool drafts a reply to each message that reads like your best day. The lesson on one good example is the whole method.

Keeping it honest

The draft says only what the system says. A date comes from the schedule or is a marked blank. A price comes from the list or is a marked blank. An apology is fine; a promise is a person's. The draft shows where each fact came from, and the reviewer reads the facts, not the prose.

What a person writes

The reply to the customer who is leaving. The apology for the thing that went badly wrong. The message to the customer who has been with you twenty years. A tool that drafts those is a tool that has misunderstood what a reply is for. Write them by hand, and let the tool draft the other two hundred.

The question at the machine

What do I do when this light comes on. Which way does this part go in the fixture. What was the setting last time we ran this job. An operator asks, in plain words, at a screen by the machine, and gets an answer drawn from the approved procedure, the machine's manual and the notes your best people have left, with the source shown. Not from the internet, and not from a guess. If the answer is not in your documents, it says so and tells them who to call.

Only your documents

The rule that makes this safe is that the AI answers from what you have approved and nothing else. A general assistant that will happily invent a torque value is a hazard on a floor. One that can only quote your own procedures, and shows which one, is a colleague. Setting that boundary is a configuration decision, and it is the first thing to check in any tool you are shown.

Capturing what walks out the door

The senior setter's notes. The photo of the fixture set up right. The reason the press trips on humid days. Give people a simple way to add a note at the machine, in words or a picture, and let the AI file it against the machine and the job. In a year the plant has a body of knowledge that used to live in three people, and the three people get asked less.

Why it is a good first project

Low risk, because it changes no record and runs no machine. Fast, because the documents already exist. Visible, because operators use it on day one and say so. And it builds the habit that every other floor AI needs: writing things down where a system can read them. Start here and the harder projects find the ground already prepared.

Where it plugs in

The approved-document store in the quality system, the machine manuals, the engineering data for the drawings and instructions, and the maintenance history for the faults. Read-only by design. A person approves any document before the AI may quote it, and the notes people add go through the same gate before they become the answer.

The question and the page

Typed or spoken at the machine: what is the torque for this fastener, which fixture for this part, what did we do last time the spindle alarm came up. The tool searches your documents, the manuals, the instructions, the closed work orders, and answers in a sentence with the passage and the page it came from. The operator reads the page, not the tool's paraphrase, and carries on.

Why it is the best first floor project

It reads and does not control. It touches no machine. A wrong answer is caught by the operator looking at the page shown. It is measurable: questions answered, minutes not lost, calls to the day shift that stop. And it teaches the whole plant what these tools are before anything with more risk arrives. The manufacturing pages say start in the office; this is the one floor project that can go first.

What it refuses

Anything about isolation, guarding, lifting or safety procedure, which it hands to the written procedure and the supervisor, never paraphrased. Anything that would change a parameter. Anything the documents do not contain, where it says so instead of guessing. The refusals are written into it before it takes its first question, and a sample of questions is read monthly for the ones it should have refused.

The dangerous answer

General, fluent, plausible and from somewhere else. The torque from a similar machine. The setup from a different revision. The procedure from a manual for a model you do not have. It sounds exactly like the right answer, and the operator has no way to tell. The lesson on what it cannot know is the background; this page is what to do about it on the floor.

The setup

The tool is connected to the folders that hold your truth, from the page on what lives where, and to nothing else. Its standing instruction says: answer only from the documents provided; show the passage; if the answer is not in them, say so and name the person to ask. That instruction is tested with questions you know are not in the documents, and the tool passes when it says it does not know.

Showing the page

Every answer carries the document, the page and the passage, on the screen, so the operator reads the source. A tool that gives the answer without the page is asking to be trusted; a tool that gives the page is asking to be checked, and the second is the one that belongs on a floor.

At the machine, in their words

During a setup, at a changeover, after a fault: the senior operator says what they are doing and why, into a phone, in forty seconds. The tool turns it into a note against the machine and the part: the trick, the order, the thing to watch, the thing the manual gets wrong. Reviewed by the operator once, and filed where the tool reads. Twenty years, a few notes a week, and the plant has a document nobody could have sat down to write.

What to ask for

Not everything; the things a new operator would ask. What do you check before you start. What does it sound like when it is wrong. What do you do that the instruction does not say. What did you learn the hard way. A prompt that asks those four gets the knowledge; an open invitation to document gets nothing, because nobody documents.

The day they leave

The new operator asks the tool, and the tool answers in the old operator's words, with the note and the date. That is the return, and it cannot be bought later at any price. Start with the person closest to retiring, this month.

The office first

Quoting from your own history. Orders read off emails instead of re-typed. Supplier invoices matched. Status questions answered from the system. Each of those is a week-one win with a person still signing, and together they free the people who should be running the plant to run it. The floor projects are real too, and they go better when the office data is already flowing.

Then the floor

Inspection results watched for drift. Faults turned into work orders at the machine. Procedures and troubleshooting answered at the station from your own documents. Scheduling that sees the real state of the floor. All of it depends on floor data being collected, which is the same investment every time, so it pays more than once. What stays with the controller stays with the controller: an AI dispatches and monitors, it does not run the machine.

The systems it plugs into

The ERP, the MES if you have one, the quality system, the maintenance system, the scheduling tool, the engineering data. Most of the value in this whole subject comes from an AI reading those and drafting into them, and the makers of those systems are now shipping the way in. Ask each vendor the one question: can an AI read and act on our data, and what does it cost.

What to be careful of

A tool bought before the process it serves can be written down. Write access granted before the reading has earned trust. A demo on the vendor's data rather than yours. And any claim that AI programs your robots or writes your controller logic; the defensible claim is narrower and still worth a great deal.

How we help

An assessment across the business that ranks the opportunities by value, readiness and risk, in your words. The readiness check on your systems. A first project scoped small and reversible. Vendor and tool comparison where what fits comes first, partner or not. And, when you need it, help choosing or rescuing the ERP or MES underneath it all.

Why the office

The office already runs on documents and systems, which is what these tools read. A wrong draft is caught by a person at a desk and costs a minute. The result is measurable in hours and days: the quote turnaround, the order lead time, the expediting calls. And nothing on the floor has to change to try it, which means nothing on the floor can be broken by it.

The four usual first projects

Quotes drafted from the request and your own history. Orders read into the ERP without re-keying. Supplier invoices matched to receipts. The floor's questions to the office answered from the documents: the drawing revision, the spec, the customer's packaging requirement. Each has its own page here. Pick the one whose hours you can count.

What the first project proves

That the data is good enough, or where it is not. That someone inside will own it. That the team will use it. That the number moved. Those four are what the second project needs, and an office project produces them in a quarter. A floor project produces them in a year, if it produces them.

The three that pay

The question at the machine, answered from your documents: the setup sheet, the drawing, the last time this fault happened. The maintenance work order raised from the fault and the history. The inspection result read as it happens and the drift flagged before the batch is scrap. Each reads; none controls. Each has a page in the applications column.

The line

The tool does not set a parameter, stop a machine, override a guard or make a safety decision. Controls stay with controls, safety with safety, and the people who own them are in the room before the tool is on the floor. A vendor whose demo shows the tool adjusting the machine is showing you a project this practice does not do.

Why floor projects stall

The data is not where the demo assumed. The senior operator whose knowledge the tool needs is not asked. The floor is not told why. The IT and controls people find out from the vendor. Each is a people problem, and each is why the office project comes first: by the time you reach the floor, the plant knows what these tools are and who owns them.

What each holds

The ERP: customers, orders, prices, stock, purchasing, accounts. The MES: what is running, on what, by whom, against which order. The quality system: results, non-conformances, the trail. The maintenance system: assets, faults, work orders, history. A tool that helps in the office reads the first; on the floor it reads the rest. Ask which your tool reads, and how.

When the system is wrong

The price list with last year's prices. The stock count that is fiction. The MES nobody logs into. The tool reads what is there and the answer inherits the error, in a fluent voice. The first month of any AI project in a plant is usually spent discovering which system is wrong and fixing that. It pays before the tool does.

The other tab

Choosing an ERP or MES, rescuing a rollout, getting more from the one you run: that is Software & robotics, and it is where a plant that has outgrown its systems should go before it buys AI to sit on top of them. The AI pages assume the systems are there. That tab is for when they are not.

Customer data

Drawings, specs and prices belong to the customer, and many customers say where they may go. A consumer AI app on an engineer's phone is not one of those places. Approved tools, business plans, training use off, and a list of which customers' material may go in at all. Read before the first drawing is pasted, not after the customer asks.

The floor

Operators who learn about the tool from a vendor's badge on a screen assume it is there to replace them, and they are not wrong to wonder. Tell them what it reads, what it does not touch, and what it is for, before it arrives. The senior operator whose knowledge it needs should be the first person shown it, not the last.

The number

Quote turnaround in days. Expediting calls per week. Scrap found at final inspection. Write the baseline down before the tool arrives, and the target, and the date you will look. Without it, the project ends in an argument about whether it worked, and the page called how do I know it worked exists because that argument is common.

The quiet decision

The tool suggests a supplier and the buyer clicks. It suggests a schedule and the planner publishes. After a month, the tool is deciding and the person is signing. Decide which decisions stay with a person, tell the tool, and read a sample. The lesson on the reviewer who stops reading is about exactly this.

Orders and documents, once

Customer orders arrive as emails, portal downloads and spreadsheets. Shipments generate pick lists, packing slips, labels, advance notices and carrier bookings. Most of it is the same order typed again. An AI reads the order in and drafts every document out, with a person releasing them and the exceptions routed to a person first. It is the plainest win in the building.

The count comes first

No AI fixes a bad count, and a smarter tool makes a wrong count wrong faster. Scanning at receipt, pick and ship, and a cycle count people actually do, come before any of this. Once the system's number is the rack's number, reorder points that follow demand and lead time, suggested by the tool and decided by the buyer, are a real improvement.

Carriers and routes, proposed

Which carrier, which service, which truck, which order of stops. Tools that propose those against cost, time and the customer's window are mature, and an AI layer makes it easy to ask why and to change the answer. The dispatcher decides, because the dispatcher knows the customer who will not take a delivery after three and the driver who knows the site.

Status, answered

Where is my order is most of the phone traffic in a distribution office. An AI reading the warehouse and carrier systems answers it in seconds, by email or with a person on the line, and gets it right. That alone changes what the customer service desk spends its day on.

Where it plugs in

The warehouse management system, the order entry in the ERP, the carrier accounts, the shared mailbox. Read first: the AI reads and drafts, a person releases. Writing to stock levels comes last, if ever. Robots in the warehouse are a separate decision and a robotics one; we do those too, and we keep them separate on purpose.

Every format, one record

The PDF purchase order, the customer's spreadsheet, the portal download, the order dictated on the phone: each read for the customer, the ship-to, the lines, the quantities, the dates, the reference, and drafted into the system as an order awaiting release. Matched to the price list and the customer's terms. The person on the desk releases orders instead of typing them.

The paperwork, from the order

Packing lists, labels, the bill of lading, the commercial invoice, the customs declaration where one is needed: each drafted from the order and the product master, in the format the carrier or the border wants. Assembled in a minute, checked in a minute, printed. The shipment that used to wait on paperwork does not.

What gets checked

The order that does not match the price list. The quantity that is ten times the usual. The ship-to that is new. The customs value. Each is flagged and a person looks, which is the whole desk's job once the typing is gone. And a sample of clean orders is read weekly, because a tool that reads a hundred a day is wrong once a week somewhere.

Where it drifts

Receipts booked in against the wrong line. Picks that took from the next bin. Returns put back unprocessed. Damage nobody adjusted. The tool watches the transactions and the counts that do happen, and points at the products whose movements no longer add up. Those are the ones to count today, not the aisle whose turn it is.

Counting where it matters

A daily list of the twenty products most likely to be wrong, weighted by value and by how soon they will be promised. The count entered on a phone at the shelf. The variance reconciled against the transactions to find the habit that caused it. Over a quarter, the list gets shorter, because the habits get fixed.

Why it comes first

Availability shown to customers from a wrong count is a promise you cannot keep. Reorder points from a wrong count buy what you have and miss what you need. The route built for a pick that is not there is a wasted walk. Fix the count, then everything built on it is worth building. Skip it, and the tools are confidently wrong at speed.

What gets proposed

For each shipment: the carriers and services that meet the customer's date and requirements, ranked by landed cost, with the transit time and last month's on-time rate for that lane beside each. For a multi-drop day, the route with the stops in an order that fits the windows. The desk sees three options with reasons, not a rate card and a map.

What the desk knows

That one carrier's driver on that route is reliable and the rate card does not know it. That a customer's dock closes at two whatever their profile says. That the cheap service loses pallets in the wrong month. The tool proposes; the desk picks, and the picks teach the tool which lane is really which.

What the tools miss

Accessorial charges that appear on the invoice and not the quote. Residential deliveries flagged as commercial. The dimensional weight nobody measured. Each is fixed by feeding the invoice back in, which the previous pages make possible: the invoice read, matched to the shipment, the surprise flagged.

The sentence

The order identified from the reference or the email. The status read from the order system, the pick from the warehouse, the tracking from the carrier. One sentence: shipped Tuesday on the two-day service, out for delivery, signed for at eleven. At any hour, on any channel, in your voice. The desk answers the questions that are not this one.

Before they ask

When the carrier's scan says delayed, the customer gets the new date from you before they check. When a pick is short, the customer gets the choice: wait, split, substitute. The tool watches the exceptions and drafts the notice; a person releases the ones that promise anything. Customers remember the company that told them first.

What still needs a person

The lost shipment. The damaged one. The customer who is really calling to say they are moving their business. The tool recognises the shape of those from the words and hands over with the whole history attached, so the person starts from knowing.

Proposals from your own work

A new request comes in. The AI finds the three most similar engagements the firm has done, drafts the proposal in the firm's format from them, and flags what is different this time. The partner edits and sends. What used to take an evening takes an hour, and the proposal is better because it is grounded in what the firm actually did rather than what someone remembered.

Documents in, summaries out

The client sends four hundred pages. An AI reads them, summarizes them, and answers questions about them with the page cited. That is the single largest time saving in most firms, and it is safe as long as the person acting on the summary can see the source. It is not safe when the summary is treated as the document.

Time and intake

Time captured from the calendar and the documents worked on, drafted into the billing system for a person to correct. New-client intake read off the forms and the emails into the practice system. Both are dull, both leak money, both are exactly the shape of work that an AI does well with a person checking.

Confidentiality, first

Client documents must not become anyone's training data, and must not leave the systems the firm controls. That is a contract question and a configuration question before it is a tool question, and it is the first thing to settle. Some tools pass it easily; some cannot; the answer decides the shortlist before any feature does.

What stays with a person

The advice. The signature. The conversation about the fee. Anything a regulator or a court would ask who decided. Write those down, keep them human, and let the AI carry the rest. We help the firm find where the hours go, pick the tools that respect the confidentiality line, and start with one workflow that is small and reversible.

What it draws on

Your won proposals, your scope language, your pricing patterns, your case studies. Given a brief, a call note or an email from the prospect, it finds the three most similar past proposals and drafts from them: the approach, the phases, the team, the terms you always use. In your voice, because your voice is what it read.

What it must not invent

A fee. A named person's availability. A result you did not achieve. A client name you may not cite. Each of those is a stop point, and the tool leaves a marked blank for a partner rather than a plausible number. Read the lesson on made-up facts before letting it near a case study.

What a partner still does

Decides whether to bid. Sets the fee. Reads the draft for the sentence that is nearly right and wrong in a way only someone who has done the work would catch. That reading takes half an hour on a draft and a day on a blank page, and the half hour is where the judgement goes.

Where it plugs in

The document store, the CRM, and the templates. The proposal library needs one afternoon of curation first: the wins, tagged by service. Then the tool has something to learn from that you would want it to learn.

What it does well

A first pass over a long document: what it is, what it covers, the parties, the dates, the unusual terms. A comparison of two versions. A search across a hundred documents for the one that mentions the thing. A list of every date and deadline in a file. Each of these takes a person hours and the tool minutes, and each is a starting point rather than an answer.

Where it reads badly

The middle of a long document, which it skims. Clauses that mean something only against another clause forty pages away. Anything where the absence of a term is the point. The lesson on the lost middle is the one to read here: work in sections, ask specific questions, and never take a summary of a sixty-page agreement as the reading.

A summary you can rely on

Ask for the answer with the passage quoted and the page cited. Ask the same question of each section rather than the whole. Ask what is missing that you would expect to see. Then the person who signs reads the passages, not the document, and the time saved is real without the risk being.

Time, drafted

At the end of the day, the tool looks at the calendar, the emails sent, the documents opened and edited, and drafts the time entries: client, matter, a plain description, a duration. The fee earner corrects and confirms in five minutes rather than reconstructing the day on Friday. Firms that do this find hours that were being worked and never written down.

Intake, asked once

Before the first meeting, the prospect answers the questions the firm always asks, in a conversation rather than a form, and the tool turns the answers into the intake record and a one-page brief for the person taking the meeting. The conflict check runs on the names before anyone has spent an hour.

What stays with the person

The narrative on the bill, which a client reads. The judgement about what is billable. Anything about a client that the tool inferred rather than was told. Time entries go out under a name; the name reads them.

Where it plugs in

The practice management system, the calendar, the email, the document store. Time capture first: it is contained, measurable, and pays for itself in the first month for most firms.

The plan and the terms

Business plans with training use off, retention stated, and a contract that says so. Never the free consumer app on someone's personal login, which is exactly what half the firm is using on a Saturday. Read the vendor's data terms yourself; they are two pages, and they are the part of the product a professional is obliged to understand.

What never goes in

Anything a client has told you may not leave the firm. Anything covered by a client's own restriction on where data is processed. Names, where a placeholder would do. The habit of stripping identifying details before pasting is small, and it makes most of the other questions disappear.

The one page

Approved tools and accounts. What may go in and what may not. How to strip a document first. Who to ask. Where the settings are checked and by whom. Pinned where the tools are used, read by every new joiner, reviewed twice a year. This is the document a regulator or a client asks for, and having it is most of the answer.

Takeoffs and estimates

Reading a drawing set to count and measure is slow, and software for it has existed for years. The newer tools use AI to do the first pass and to pull your own past bids for the pricing. The estimator reviews the takeoff and owns the number. Bidding more jobs in the same hours is the win, and it is a large one for a small contractor.

Change orders that get written

A foreman says it in the truck and nobody writes it down, and the extra work gets done for free. An AI that takes the note in plain words, on a phone, at the site, drafts the change order with the job, the date, the reason and the likely cost, and sends it to the office for a person to price and issue, recovers money that was already earned.

Field reports from photos and notes

Daily reports, safety records, progress photos with a sentence attached. Drafted by an AI from what the crew already captured, checked by the super, filed against the job. The report gets written on the days it used to be skipped, which is most days.

Submittals and the paper trail

Submittals, requests for information, approvals, the log of who said what when. An AI that reads the documents and keeps the log current, and answers where a submittal is, takes a job's worth of chasing off the project manager. It does not decide anything; it makes sure the decision is on file.

Where it plugs in

The estimating software, the job-cost and accounting system, the project management tool, the phones in the field. Read first: the AI drafts, a person issues. Job cost is the number the business lives on; nothing writes to it without a person. We help find where the hours and the margin go, and pick tools that fit a contractor's size rather than a national builder's.

The takeoff, drafted

From the drawings, a first-pass count of the things you always count: fixtures, outlets, linear feet, doors, square footage by finish. Not perfect, and it does not need to be; it needs to be a list the estimator corrects rather than a blank page they fill. Read the earlier lesson on what a model cannot measure before trusting a dimension it did not read as text.

The estimate, from your history

Priced against your own unit costs and your own last five similar jobs, with the places where you went over budget last time flagged. The tool can say: the last three jobs like this ran twelve percent over on electrical. That is the sentence an estimator wants at eleven at night and never has.

What stays with the estimator

The number. The read on the client. The knowledge that this site has a slope the drawings do not show, and that the general contractor pays slow. The tool builds the spreadsheet faster; the estimator decides whether to want the job and at what margin.

Where it plugs in

The estimating software, the drawing files, the cost history in the accounts, and whatever spreadsheet the real unit costs live in. The last one is usually the problem: get it true first, and the tool has something worth reading.

On the ladder

The foreman says it: client wants the outlet on the north wall instead, about six feet of extra run, one more box. Takes a photo. The tool drafts the change order: the description, the quantities against your unit costs, the schedule impact if there is one, in your format. The foreman reads it, the client gets it on their phone, and the crew moves when it is signed.

Why it changes the number

Not because the tool prices better; because the change order exists. Most give-aways are not generosity, they are friction. Remove the friction and the billing follows. Contractors who do this consistently find the change order line on the year's accounts and are surprised by it.

What the site lead decides

Whether to charge. Whether it is really a change or was always in scope. Whether the client relationship is worth the small one. The tool makes the document; the person makes the call, and the document makes the call cheap to make.

End of day, two minutes

The foreman talks: who was on site, the weather, what got done, what got delivered, what went wrong, who visited. The day's photos go in. The tool writes the report to the contract's format, dates and times it, places the photos, and flags anything that sounds like a delay or a safety issue for the project manager. The foreman reads and signs. Two minutes, every day.

The dispute, later

A daily report that says the drywall delivery was two days late, with a photo of the empty bay and a timestamp, is the claim. A report that was not written is a memory against theirs. The value of this page is not in the report on the day; it is in the folder of them at the end of the job.

What still gets checked

Anything that names a person or a subcontractor and what they did or did not do. Anything that will be read by a lawyer. The foreman signs it, which means the foreman reads it, which the two-minute version makes possible.

Submittals, assembled

For each section of the spec, the tool lists what has to be submitted, pulls the product data from the supplier documents you have, and drafts the package in the format the architect wants. The project manager reviews and sends. What took an afternoon per section takes an hour, and the schedule of submittals is generated rather than remembered.

RFIs, drafted and half answered

Before an RFI goes out, the tool searches the spec and drawings for the answer. Often it is there, in a section nobody had read. When it is not, the RFI drafts with the drawing reference, the spec section and the photo, so the answer comes back faster because the question was asked properly.

The trail itself

Every submittal, RFI, response and change, indexed and searchable, so that when the question is asked again in month nine, the answer from month two is found in seconds. That is the quiet value: the project's memory in a place that is not one person's inbox.

Where it plugs in

The project management platform, the document folders, the email. Read first; drafts only. A submittal goes out under a person's name, and that person has read it.

The call, booked

A homeowner calls or messages. An AI that answers, gathers the address, the equipment and the symptom, checks the calendar and books the slot, with a person able to step in at any point, means the call gets booked at seven in the evening instead of going to a competitor. It should say plainly that it is automated and hand over the moment the caller asks.

Dispatch, proposed

Which tech, in which order, with what parts. Tools that propose the day's route against skills, location and priority are mature, and the AI layer makes them easy to ask and easy to override. The dispatcher decides, because the dispatcher knows the customer who needs the senior tech and the tech who is having a bad week.

The quote at the job

The tech describes the fix in plain words, or photographs the nameplate, and the AI drafts the quote from your own price book and the last similar jobs, on the tablet, before the tech leaves the driveway. The tech checks and presents it. Faster quotes close more work, and quotes built from your prices stay on margin.

Parts and the truck

Truck stock that is counted, replenished from the jobs completed, and suggested from the calls booked for tomorrow, so the part is on the van before the tech knows they need it. It depends on the count being real, which depends on techs scanning what they use. The tool is the easy part.

Where it plugs in

The field service software that runs the calls and invoices, the phone system, the price book, the supplier accounts. Read first: the AI books, proposes and drafts; a person dispatches, presents and invoices. We help a contractor see where the hours go, use the software they already pay for, and add AI where it earns its keep rather than where it is sold hardest.

What it can do at seven

Answer at once, say plainly that it is an automated assistant, and gather what a dispatcher would: the address, the equipment, the symptom, whether anyone is at risk. Check the calendar. Offer the first real slot, or the emergency line if the symptom warrants it. Confirm by text. The homeowner has a time and a name before the second number rings anywhere.

What it must never do

Quote a price. Promise a technician by a time nobody has confirmed. Diagnose. Tell a caller with a gas smell to wait for the morning. Those are the stop points, written into the tool before it takes its first call, and each one hands to a person or to a script you wrote.

The hand-off that works

For an emergency, it reaches the on-call technician by the route you set and stays on with the caller until it has a confirmed answer. For everything else, the dispatcher arrives at seven in the morning to a list of booked calls with the details already in the system, not a voicemail queue to work through. That hour is where the money is.

Where it plugs in

The phone system, the field service software that holds the calendar and the customer record, and the on-call rota. Read first: it books into the calendar and writes the customer note; a person dispatches. Start with after hours only, listen to a week of recordings, then decide whether it answers in the day too.

What gets proposed

A day's route for each truck, against skills, location, priority and the parts on board. The reshuffle when a job overruns. The question the dispatcher used to answer by scrolling: who could take a call in the north end at two? The tool answers it in a sentence and shows its reasoning. That is the part worth paying for.

What the dispatcher still knows

That the customer at number nine will only have the tech who fixed it last time. That the new hire should not be sent to the commercial job alone. That one tech is having a bad week. None of that is in the system, and it decides more of the day than the map does. So the tool proposes and a person moves the pieces, and the pieces stay movable.

Where the tools get it wrong

Optimising for miles when the customer wanted the morning. Treating a two-hour estimate as a fact. Sending the right skill to the wrong parts. Every one of these is fixed by the dispatcher looking at the draft before it goes to the trucks, which takes five minutes, against the forty it used to take to build.

What the tech does

Says the fix in plain words, or photographs the nameplate and the failed part. The tool reads it, looks up your price book, pulls the last three similar jobs, and drafts the quote with the options you always offer: repair, replace, the maintenance plan. The tech reads it, changes what needs changing, and presents it. Ten minutes becomes two.

Where the margin lives

In the quote being built from your prices, not the tech's memory. In the labour being estimated from your own history of that job, not a guess. In the second option being offered every time, because the tool never forgets to. Shops that do this see the average ticket move, and the movement is not from charging more; it is from charging what the book says.

What stays with the tech

The number the customer hears. The read on whether they want the cheap fix or the right one. The judgement that the nameplate is wrong about the model. The tool drafts a document; the tech makes a sale, and the two are not the same skill.

Where it plugs in

The price book, the field service software's quote module, the job history, and the tablet in the truck. The one rule: the price comes from the book at the moment of the quote, never from anything the tool remembered. When the book changes, the next quote changes.

What the tool can do

Keep a live count per truck, taken down by what the jobs consumed and topped up by what was restocked. Look at tomorrow's booked calls, the equipment on each, and the parts those jobs usually need, and print the pick list for each van tonight. Flag the truck that is about to run out of the thing it uses most. None of this is clever. All of it is a chore nobody does.

What it depends on

The count being real, which means techs scanning what they use. That is a habit, not a feature, and it is the reason most truck stock projects fail. Make scanning the fastest possible motion: a code on the bin, a tap on the phone, done. Then check the count against a real van once a month and tell the crew the number. The tool is the easy part.

The arithmetic

Count the second trips. Multiply by the hours each one took, at the rate you bill. Most shops are surprised by the number, and it is the number that pays for the scanner, the bins and the tool several times over in the first year.

Privacy before anything

Patient information is governed by rules that decide which tools may touch it, where it may be stored and who may see it. That is the first conversation, before any feature, and it removes most tools from the shortlist. The ones that remain have the agreements and the controls in place; ask for them in writing and read them. We help with that reading, and we do not sell any of the tools.

The front desk

Appointment requests read from the phone and the portal and drafted into the schedule. Reminders and confirmations sent. Intake forms read into the practice system instead of typed. Eligibility checked before the visit rather than after. Each takes work off the front desk with a person still confirming, and together they change what the desk is able to do for the patient in front of it.

Billing and documents

Claims drafted from the visit record, checked for the errors that get them denied, and reviewed by a person before they go. Records requests answered from the file. Referral letters drafted for the clinician to sign. This is the same read-and-draft pattern as any office, and it is where practices recover the most money.

Not the clinical decision

Diagnosis, treatment, anything a clinician is licensed to decide, stays with the clinician. Tools that draft the clinical note for the clinician to review exist and some are good; tools that decide are a different category with different rules and are not what we advise on. Draw that line in writing and keep it.

Where it plugs in

The practice management system, the scheduling and phone systems, the billing and clearing house connections, the document store. Read first, with access limited to the minimum each task needs. A person confirms an appointment, releases a claim and signs a letter. We help a practice find the hours, settle the privacy question, and start with one workflow that is small and reversible.

The agreement

Any vendor whose tool handles patient information has to sign the agreement your rules require of a third party, and to be able to say where the data is stored, who can see it, and how long it is kept. A vendor who cannot or will not is not a vendor for a practice, however good the demo. This is the first question, and it ends most conversations early, which is the point.

The accounts to close

The consumer app on a staff member's phone that has been summarising referral letters. The free transcription tool. Every practice has some of these, used with good intentions and no agreement. Find them, close them, and give people an approved tool that does the same job, or they will go back.

The line

Front desk work, booking, reminders, billing and documents, is where the tools help and where the rules can be kept with a signed agreement and the right settings. The chart and the clinical decision are a different territory with different rules, and these pages do not go there. The next pages stay on the office side of that line.

Booking and rebooking

By phone, text or the portal: the request understood, the slot found against the schedule and the appointment type, the confirmation sent. Rebooking handled the same way. The patient who calls at six in the evening gets an appointment rather than a voicemail, and the desk arrives to a schedule already filled.

Reminders and forms

A reminder two days out with a yes or a reschedule. Intake forms sent before the visit and returned filled, so the clipboard disappears and the appointment starts on time. The no-show rate moves first, and it moves a lot; it is the number to write down before you start.

What stays with the receptionist

The upset patient. The urgent one. The question that is really a clinical question in disguise, which the tool must recognise and hand over rather than answer. Anything about a bill that is a dispute rather than a query. The tool takes the volume; the person takes the people.

Where it plugs in

The practice management system's scheduler, the phone system, the patient messaging channel, and the forms. Under the agreement from the previous page, with the settings checked. Reminders first; booking after a month of watching.

Claims, checked before they go

The claim read against the rules that usually reject it: the missing field, the mismatched code, the authorisation that was not attached. Flagged before submission rather than explained after rejection. Rejections that do come back are summarised in a sentence with the fix, so the resubmission is a correction rather than an investigation.

Letters and authorisations, drafted

The referral letter drafted from the visit record in the practice's format. The prior authorisation assembled with the documents the payer asks for. The summary for the patient in plain words. Each a draft, each reviewed, each taking a person minutes instead of the afternoon.

What the coder still signs

The code. The tool can suggest and check; the coder decides, because the coder is accountable and the tool is not. That line is the same one the clinical page draws, and for the same reason.

Why the line is here

The clinician is trained, licensed and accountable. The tool is a pattern matcher that sounds certain either way. On the office side, a wrong draft is caught by a person and costs a minute. On the clinical side, the cost of the same mistake is not a minute. Keep the tool where a mistake is cheap and a person reads everything before it matters.

The tool that crosses it

You will see it in a demo: the front desk tool that also offers to draft the clinical note, or answer the patient's question about their symptoms. The question to ask is whether the vendor has thought about the line at all. If the answer is a feature list, it has not. If the answer is a clear statement of what the tool refuses to do, it has.

Keeping it

Write down what the tool may touch: scheduling, reminders, intake, billing, documents. Write down what it may not: the chart, the clinical note, anything a patient asks about their condition. Put the second list into the tool's instructions as hand-off rules, and check a sample of conversations monthly for the ones where it should have handed off and did not.

Documents in

An application arrives with its statements, its identification, its supporting files. An AI reads them, checks that the set is complete, pulls the facts into the file in your format, and flags what is missing or inconsistent. The person who used to do that reading now starts at the interesting part. Days of turnaround become hours, and nothing has been decided yet.

The file, summarized

For the underwriter, the loan officer, the adjuster: a summary of the file with every fact linked to its source page. That is the shape that makes AI safe in a regulated office. The person can check any line in a click, and the record shows a person decided. A summary that cannot be traced to its source is a liability, and we would not recommend a tool that produces one.

Service, answered

What is my balance, where is my claim, when does my policy renew, what did I pay last year. Answered from the system in seconds, by email or with a person on the line. That is most of the inbound traffic in many offices, and it is the safest place to start because the answer is a fact the system already holds.

The decisions that stay licensed

Approval. Pricing. The claim decision. Advice to a client. Anything a regulator would ask who decided and on what. Those stay with the licensed person, in writing, before any tool is chosen. Tools that offer to make them exist; they carry rules and risks of their own, and they are not what a small office should start with.

Where it plugs in

The core system for policies or loans, the document store, the client records, the mailbox. Read first, with access limited to each task. Your regulator and your carriers or funders have views on where data may go and which tools may touch it; settle those before the shortlist. We help find the hours, read the tool agreements for what they say, and start with one workflow that is small and reversible.

On arrival

Each document identified: a bank statement, a payslip, a photo of a licence, a repair estimate. Sorted into the file. Checked against the list of what a complete file needs, with the missing items named and the request for them drafted the same hour. The retyping of names, dates, amounts and account numbers into the system done by the tool and confirmed by a person, with the crop of the page beside each figure.

What it checks

That the statements cover the months required. That the name on the payslip is the name on the application. That the dates are in range. That the estimate has a total. Simple rules, tirelessly applied, which is exactly what catches the incomplete file on day one instead of day eight.

What it cannot vouch for

That a document is genuine. That a figure read off a photograph is right without a person confirming it. Anything that would be a finding rather than a fact. The tool sorts and reads; a person verifies, and the tool's job is to make the verifying quick.

The one page

What is being asked. The facts that bear on it, each with the document and page it came from. What is unusual against the files that usually look like this. What is missing. Written to the decider's own checklist, because the checklist is what they read for anyway.

Making it trustworthy

Every fact cited to its page, so the decider checks the citation rather than re-reads the file. A standing instruction to say when a fact is not in the file rather than guess it. A monthly sample of summaries read against their files by someone senior. The lessons on made-up facts and the lost middle apply here more than anywhere on the site.

The decision stays licensed

The tool summarises; it does not recommend. Anything that shapes a decision about a person's money or claim is made by the person licensed and accountable for it, from the file, with the summary as a map. That is the rule and the reason the summary cites its pages.

What it answers

The status of the claim, the application, the payment. The documents still needed and how to send them. The appointment booked. The form sent. Each from the record, each a fact rather than an opinion, each at eight in the evening as readily as at ten in the morning.

What it hands over

Anything about whether a claim will be paid, what a rate will be, what the customer should do with their money, whether to accept an offer. The list is written before the tool takes its first message, the tool is told to hand over on any of it, and a sample of conversations is read monthly for the ones it should have handed over and did not.

Saying which is which

It says it is automated, at the start, every time. It says what it can help with. When it hands over, it says so and says when a person will be in touch. A customer who knows they are talking to a tool forgives its limits; one who finds out later does not.

Why a recommendation is a decision

When the tool says the file looks fine and a busy person clicks approve, the tool decided. The lesson on the reviewer who stops reading is the whole story: after enough good recommendations, the review becomes a click. So the tool is not asked for a recommendation. It is asked for the facts, cited, and the person recommends.

The list

Whether to lend, and at what rate. Whether to pay a claim, and how much. What a client should do with their money. Whether to insure, and at what premium. Each is a licensed act, and each is on a written list the tool is told to stay away from, in its instructions and in the design of the product.

Keeping the line

Name the person who owns it. Write the list into the tool's instructions as hand-off rules. Keep the record of who decided what, so an auditor sees a person's name on every decision and a tool's name on every summary. Read a sample monthly. That is the whole discipline, and it is the same one the firm already has for junior staff.

Maintenance, triaged

A request comes in by phone, text, portal or email. An AI reads it, decides what kind it is and how urgent, drafts the work order with the unit, the issue and the likely vendor, and routes it. A burst pipe reaches a person in seconds; a dripping tap reaches the schedule. The manager corrects the ones it got wrong, and it gets fewer wrong each month.

Leases and dates

Renewal dates, rent steps, notice periods, options, the clause about the pet. Read out of the lease into the system once, watched thereafter, with a draft notice ready before the date. Most offices have that data on paper and in one person's memory. Putting it where a system can see it is the project, and the AI does the reading.

Questions, answered

When is rent due, what is my balance, where is my statement, is the pool open. Answered from the property system, by email or text, with the tone you set and a person a click away. Owners get their statements drafted from the ledger for the manager to review. The office phone rings less, and the calls that come through matter more.

What stays with the manager

Anything about a tenant's tenancy: a notice, a dispute, a decision to renew or not. Anything a housing rule would ask who decided. The conversation with an owner about money. Keep those human, in writing, before the tool arrives. The AI drafts; the manager sends, and reads what it drafted first.

Where it plugs in

The property management system, the maintenance vendor list, the mailbox and the phone, the accounting. Read first. The AI reads, routes and drafts; a person issues the work order, sends the notice and releases the statement. We help an office find where the hours go and pick tools that fit its size.

On arrival

Each message read, whatever channel it came by. The unit identified from the sender. The issue classified: plumbing, electrical, appliance, access, noise. Urgency judged by rules you set: water, gas, no heat and lockouts go to the top and to a person at once. Everything else becomes a work order with the tenant's photo attached, ready to assign.

What the manager decides

Which contractor. Whether the tenant's dripping tap is the third in that building this month, which is a different problem. Whether the request is really the landlord's to fix. The tool has sorted the pile; the manager works the sorted pile, which takes a fraction of the time and none of the retyping.

The tenant hears back

A reply within minutes: received, logged, here is the reference, here is roughly when. Then an update when it is assigned and when it is done. Most tenant frustration is silence, and the tool is never silent, which is the part tenants notice first.

Where it plugs in

The property management system's work orders and tenant records, the messaging channels, the contractor list. The emergency route is the thing to test first, at night, with a real phone.

Read once, dated forever

Each lease read on signing: the term, the rent, the review dates, the notice periods, the break clauses, the deposit, the responsibilities. Extracted into the system with the page cited, and every deadline put on the calendar with the reminder set for the notice period before it, not the day itself. The renewal conversation starts when it should, not when the tenant raises it.

Drafted on the date

When the reminder fires, the renewal offer, the increase notice or the reminder to the tenant is drafted from the lease and your template, with the right figures and the right notice period. The manager reads it and sends it. The letters that used to go out late or not at all go out on the day.

What a person still reads

The unusual clause. The lease that was negotiated and does not match the template. Anything that will end up in front of a tribunal. Figures pulled from a lease are read against the page, once, before they become a notice with legal effect.

Enquiries

Availability, the rent, what is included, pets, parking, the nearest station, when it can be viewed. Answered from the listing at once, with the viewing booked into the calendar and confirmed. The prospect who gets a viewing at nine at night is the one who turns up on Saturday.

Tenants

How to pay, when the rent is due, who to call for the boiler, what the lease says about a lodger, when the bins go. Answered from the lease and the building notes, in a sentence, and logged. The manager stops answering the same question a hundred times and starts answering the ones only a manager can.

What always goes to the manager

A complaint. A dispute about money. A request to leave early or change the lease. Anything that could become a legal matter. The tool is told this list, hands over on it, and says so. A sample of the conversations gets read monthly for the ones it should have handed over.

What the tool carries

The sorting, the logging, the reminding, the drafting, the answering of the questions with known answers. The work orders raised and chased. The dates tracked. The reports assembled for the landlord at month end. All of it done by the tool and read by the manager, which is a different week from doing it.

What stays

The tenant in arrears who has just lost a job. The neighbour dispute. The decision to serve notice. The landlord who wants something the law does not allow. The contractor who is being paid and not turning up. Each needs a person who knows the building and the people, and the tool is told to bring these to that person, not to handle them.

The manager, better

With the admin carried, the manager answers the hard message the same day instead of the same week, walks the building, and knows the tenants' names. Owners notice. Tenants notice. That is the return, and it does not show up as hours saved; it shows up as buildings kept.

The count, again

If the website says in stock and the shelf says no, every other tool makes it worse. Scanning at receipt and sale, one system of record for stock, and a count people actually do. Once that holds, reorder suggestions that follow the season and the sell-through are a real improvement, with the buyer deciding.

Listings, drafted

A new product needs a title, a description, attributes, sizes, images with alt text, in your voice, on every channel you sell through. An AI drafts all of that from the supplier's sheet and your own past listings, and a person checks it. Listing a hundred products becomes an afternoon. Keep the person on the check: the AI will confidently describe a colour the product does not come in.

Questions from the order

Where is my order, can I change the address, is this in stock in a medium. Answered from the order and stock records in seconds, by chat or email, with a person able to step in. That is most of the support volume, and answering it well is the difference between a review and a refund.

Returns, once

The return request read, the reason recorded, the label sent, the refund drafted when the item arrives and is checked. Done once from the order, with a person approving the refund. The return desk stops being the place where the day runs late, and the reasons, finally recorded, tell the buyer what not to reorder.

Where it plugs in

The commerce platform, the point of sale, the stock system, the marketplace accounts, the support inbox. Read first. The AI drafts listings, answers from the record and prepares returns; a person publishes, refunds and reorders. We help a retailer find the hours, get the count honest, and add tools where they pay.

Where the count is wrong

The tool watches sales, receipts, returns and the counts that do happen, and flags the products whose numbers no longer add up: the one that sold more than it had, the one that has not moved in a way that means it is probably mis-shelved. The cycle count list each morning is the ten products most likely to be wrong, not the aisle it happens to be Tuesday for.

Reorder points that move

A reorder point set in January is wrong by June. The tool proposes points from the actual rate of sale, the lead time, and the season, and shows its working. The buyer accepts or changes each in seconds rather than never getting round to it, which is what happens to static reorder points everywhere.

What it depends on

Receipts scanned in, sales rung through the right product, returns processed as returns. The tool reads what the system holds; if the system holds fiction, the flags are fiction too. The first month is usually spent finding the three habits that put bad numbers in, and fixing those pays before the tool does.

From the sheet to the listing

The supplier's data sheet, the photo, the category, and three listings you are proud of as the pattern. The tool drafts the title, the description, the bullet points and the attributes the marketplace wants, in your voice, for each product. What took a day per twenty takes an afternoon per four hundred, and the tone is the same across all of them.

What gets checked

Every fact that a customer will rely on: the dimensions, the material, what it fits, what is in the box, the care instructions. The tool copies these from the sheet and it can copy wrong, or the sheet can be wrong. A person reads the facts, not the prose. The prose is what the tool is for.

The mistakes that reach customers

A description that claims a feature the product does not have. A size copied from the wrong variant. A compatibility that was true of last year's model. Each is a return, a review and a refund. The check takes seconds per listing when the reviewer knows to look at the facts, and the lesson on catching it says how.

From the order

The customer identified from the email or the order number. The status read from the system and the carrier: dispatched Tuesday, out for delivery, delayed with a new date. The change made where the system allows it: the address before dispatch, the size before picking. Each answered in minutes, at any hour, in the tone you set.

From the catalogue and the policy

Does it fit, is it in stock, does it come in blue, can I return it. Answered from the listing and the returns policy, with the link. The tool must answer only from those and say when it does not know, because a confident wrong answer about a return window is a dispute.

To a person

The angry customer. The damaged item. The refund outside policy. Anything the tool was not built for. It hands over, says so, and gives a time. And the desk reads a sample of its conversations each week for the ones it handled that it should not have.

The request

Read, matched to the order, checked against the policy: in the window, in the condition, the right reason. If it passes, the label goes out in the same reply with the instructions. If it does not, it goes to a person with the policy line quoted, so the decision takes a minute and the answer to the customer is a person's.

The box

When it arrives, the receipt matched to the request, the item checked against the order by the person on the bench, and the refund queued. The tool has done the matching and the paperwork; the person has looked in the box, which is the one step that needs eyes.

The refund

Queued for release, not released. Money leaving the business is a stop point, always, and a person taps to release the day's refunds in one go. The customer gets the refund the day the box arrives, which is the number that changes the reviews.

Bookings and questions

A guest calls or messages during service. An AI that takes the booking, answers the question about the menu or the parking, and hands to a person the moment it is asked to, means the phone is answered when the manager cannot. It should say that it is automated, and it should never guess about an allergy; that question always goes to a person.

The schedule, drafted

Next week's rota drafted from the bookings, the forecast, the weather and who is available, for the manager to adjust. Forecast tools have existed for years; the AI layer makes the draft and explains it. The manager decides, because the manager knows who can run the pass on a Saturday and who cannot.

Invoices and inventory

Supplier invoices read in and matched to the delivery, price changes flagged, the entry drafted for the bookkeeper. Counts read from the sheet or the scan into a stock system that can then suggest the order. The food cost that used to be known at month end is known on Tuesday.

The guest gets a person

The complaint. The special occasion. The regular. The allergy. Decide up front what an AI may never handle and keep it that way in writing. The point of the tools is to give the people on the floor more time for exactly those moments, not fewer.

Where it plugs in

The booking system, the point of sale, the scheduling tool, the accounting, the supplier accounts, the phone. Read first. The AI books, drafts and matches; a person confirms, publishes the rota and approves the payment. We help an operator find the hours and pick tools that fit a single site or a small group.

The booking

Date, time, party size, the request for the window, taken by phone, message or the site, checked against the floor plan and the turn times you set, confirmed in writing, reminded the day before. Cancellations and changes go straight back into the book. The no-show rate is the number to write down first.

The questions

Is the terrace covered, do you do a set lunch, is there parking, can you do a vegan menu, are dogs allowed. Answered from what you have written down and only from that. Allergen questions get the written allergen information and a hand-off to a person, every time, because that is not a question to get nearly right.

What a person handles

The party of thirty. The complaint. The guest who wants to speak to the manager. Anything about an allergy beyond reading out the sheet. The tool hands over and says so, and the host gets those calls and not the other two hundred.

What the draft uses

The bookings and the forecast for each service. Staff availability as they have given it. The rules: maximum hours, minimum rest, who is trained to close, who is on the bar, the pairings that work. From that, a rota that covers each service to the level you set, with the cost shown. The manager sees the draft, not a blank grid.

What the manager knows

Who is quietly looking for another job. Who needs the Saturday for the money. Which two should not be on the same shift this week. The tool proposes; the manager moves the pieces and the reason stays in their head where it belongs. The point is that they move four pieces instead of placing forty.

When it changes

Someone calls in sick on Friday afternoon. The tool lists who is available, trained and under hours, and drafts the message. The manager picks and sends. Ten calls become one.

Where it plugs in

The reservation system, the scheduling software or the spreadsheet the rota lives in, and the messaging the team actually reads. Get the rules written down first; most of them have never been.

The invoice, read

Photographed at the back door, read line by line, matched to what was ordered and what was signed for, and priced against the last one from that supplier. The line where the price moved is flagged with the percentage. The short delivery is flagged. The invoice goes into the accounts without being keyed, and the chef sees the price rise the week it happens instead of at month end.

The count

Entered on a phone, walking the walk-in, by voice or by tapping the list the tool prepared from what should be there. Reconciled against deliveries and sales to show what went missing, was wasted or was never rung through. The theoretical against the actual, weekly, in a form the manager reads in two minutes.

What a person still checks

The figures read off a crumpled invoice are confirmed against the crop of the image before they become a payment. The variance that looks like theft is looked into by a person, because it is sometimes a menu change nobody told the system about. The tool finds; the person judges.

Behind the pass

Everything the previous pages describe: the booking taken at any hour, the rota drafted, the invoices read, the count reconciled, the supplier chased, the review replied to in draft for the manager to send. None of it in front of a guest. All of it taking hours off the people who would otherwise do it after service.

In front of the guest

The greeting, the recommendation, the read on the table that wants to be left alone, the apology when the kitchen is slow. A tool that tries to do these is a tool that guests remember for the wrong reason. The booking confirmation can be automated; the welcome cannot.

Why the line pays

A front of house with the admin taken off them looks up more. Regulars get remembered. Complaints get caught at the table. Reviews move. That is the return on automating the back office, and it does not show up as a saving; it shows up as a fuller room.

Booked and dispatched

The call taken and booked, with a person able to step in. The day's jobs proposed against skills, location, priority and parts, for the dispatcher to approve. Neither is new; the AI layer makes both easy to ask and easy to override, and it means the phone is answered at seven in the evening.

The report, at the site

The tech says what they found and did, photographs the equipment, and the AI drafts the service report, the parts used and the follow-up recommendation before they leave. The tech checks and signs. Reports get written on the day, invoices go out on the day, and the customer gets a document worth having.

The quote, before you leave

Follow-up work priced from your own price book and the last similar jobs, drafted on the tablet, presented at the site. Faster quotes close more work, and quotes built from your prices stay on margin. The tech presents; the office can review before it is accepted if you want that gate.

History, on the spot

What was done to this unit last time, by whom, with what parts, and what the manual says about that fault code. Answered at the equipment from your own records and documents, not from the internet. That is the knowledge the senior tech carries, made available to every tech, and it is one of the best first projects in the industry.

Where it plugs in

The field service software, the phone system, the price book, the parts and supplier systems, the accounting. Read first. The AI books, proposes, drafts and answers; a person dispatches, signs and invoices. We help a service business see where the hours go, use the software it already pays for, and add AI where it earns its keep.

The ask, answered

By phone, text, web form or the message that arrives at ten at night. The tool reads it, asks the questions your booker would ask, finds a slot that fits the route already planned for that area, and confirms it in writing. It says it is automated, and it hands to a person the moment the customer asks for one or says something it was not built for.

The day before

A confirmation text with the window and the tech's first name. A reply of yes, or a reschedule that goes straight back into the calendar. The no-show rate is the thing to watch here; it moves more than any other number when this is done well, and it is mostly a matter of the reminder being sent at all.

What the tech gets

The address, the gate code, the dog, the last visit's notes and the photo the customer sent, all on the phone before the knock. Pulled from the record by the tool, not typed up by the office. Small, and the difference between a tech who looks like they know the house and one who does not.

Where it plugs in

The phone and messaging, the field service software's calendar and customer records, and the map. Start with the confirmations and reminders: lowest risk, fastest payback. Let it book new jobs after a month of watching it handle the ones already booked.

At the site

The tech talks for forty seconds: what they found, what they did, what they replaced, what they told the customer, what to watch. Takes four photos. The tool turns that into the report in your format, with the photos placed, the parts listed against the price book, and the follow-up flagged. The tech reads it on the phone, fixes a word, taps done. In the truck, the invoice is already drafted.

What changes

Invoices go out the same day, because the report exists. Disputes shrink, because there is a timestamped account with photos, written when it happened. The tech's evening comes back. And every report is the good tech's report, because the format is the format and the tool never skips the section on what to watch.

What a person still checks

The parts and the quantities, because those become money. Anything the tech said the customer agreed to. A report that says the customer approved a replacement had better be right, and that is a ten-second read by the tech before done, not a habit to skip.

The one-paragraph brief

On the way to the job, the tech gets a paragraph: who the customer is, what has been done at the address, what was replaced and when, what was promised, what the last tech noted to watch. Drawn from the record and the last reports, written plainly, and read at the light before the turn. The visit starts with the tech already knowing.

What it should surface

Promises, first. Anything a tech wrote that a customer might hold you to. Then the equipment and its age. Then the pattern: the third call about the same thing is a different conversation from the first, and the tool should say so. Then the small things: the gate code, the dog, the customer's name spelled as they spell it.

Where it comes from

The field service software's customer and job records, and the reports written at the site. Which is why the report page comes before this one: a brief is only as good as what was written down. Do the reports first, and this page is what you get for free a few months later.

What it reads

The model and serial off a nameplate, straight into the job and the parts lookup. The reading off a gauge or a meter. The label on the failed part. Text in an image is the thing these models do best, and it removes the retyping that every tech hates and gets wrong.

What it assembles

The before-and-after pair, labelled and attached to the job. The warranty claim, drafted with the photos the manufacturer asks for and the serial already filled in. The customer's copy, with the pictures that show what was done. A folder that used to take the office an hour becomes a draft waiting for a glance.

What it cannot judge

Whether the weld is sound. Whether the crack is structural. How big anything is. Any question that needs the eye of the trade stays with the trade. The tool describes and files; the person decides, and the file makes the decision defensible later.