AI terms in plain English.

A

AI agent

An AI agent is a system that can take steps toward a goal instead of only answering a question: it decides what to do next, uses tools or software, checks the result and carries on. The lesson

AI bias

AI bias is when a tool repeats unfair patterns from the examples it learned from, so some people get worse results, in hiring, pricing or how well it understands their accent or name. The lesson

AI evals

AI evals are structured tests that measure how well an AI system does a specific job, by running real example inputs and scoring the answers against what is correct. The lesson

AI gateway

An AI gateway is a control layer between your apps and the AI models they use, so every request passes through one place that applies your rules, routes the work, logs use and tracks spend. The lesson

AI hallucination

An AI hallucination is an answer that is false, invented or unsupported but sounds confident, such as a fake fact, wrong number, made-up policy or non-existent source. The lesson

AI harness

An AI harness is the structure built around an AI model: it supplies the instructions and business information, connects the model to approved tools, and enforces rules, permissions and approvals. The lesson

AI image generation

AI image generation makes a picture from a written description, and it is useful for backgrounds, simple graphics, concept sketches and images for social posts. The lesson

AI memory

AI memory is not one thing: it is the mix of what the model can see in the current chat, what it fetches from your documents, and what a tool saves between conversations. The lesson

AI observability

AI observability means being able to see how your AI system is actually behaving: what goes in, what comes out, and how quickly and reliably it runs. The lesson

AI skills

An AI skill is a focused package that gives an AI system one useful job, made up of instructions, the business context it needs, the tools it may use and the output you want back. The lesson

AI video

AI video comes in three kinds: editing help on your own footage, video generated from a description, and talking avatars. The lesson

AI workflow

An AI workflow is a repeatable process built around an AI model: something starts it, such as an email or a form, the workflow gathers the data, the AI classifies, drafts or extracts, and the result is reviewed, sent or stored. The lesson

API

An API is a set of rules that lets one piece of software ask another for data or an action and get a reply, so nobody has to copy and paste between tools. The lesson

Audit trail

An audit trail is a reliable, time-ordered record of what happened in a workflow, when it happened, and who or what did it, including inputs, outputs, approvals and the model version used. The lesson

Automation tools like Zapier, Make and n8n

Zapier, Make and n8n connect your apps with little or no code: when something happens in one, such as a new enquiry, they act in another, such as adding a job. The lesson

B

Batch processing

Batch processing collects work into a queue and runs it together on a schedule, instead of handling each request the moment it arrives. The lesson

Biggest AI myths holding small businesses back

The biggest myths are that AI is only for big companies, costs a fortune to start, will replace your people, needs perfect data first and is always right. The lesson

C

Calibration

Calibration means an AI system's confidence score matches how often it is actually right: if it says 90% confident, it should be correct about nine times in ten. The lesson

Chain of thought

Chain of thought means an AI works through a problem in smaller steps before giving its final answer, instead of jumping straight to a reply. The lesson

Chunking

Chunking means splitting a large document into smaller sections so an AI system that searches your documents can retrieve just the part that answers a question. The lesson

Cloud AI platforms

Cloud AI platforms are services that let you use advanced AI models over the internet, running on the provider's servers, from companies such as OpenAI, Amazon Web Services, Google Cloud, Microsoft Azure and Anthropic. The lesson

Cobot

A cobot, or collaborative robot, is a robot built to work alongside people in a shared work area, usually on repetitive jobs such as handling parts, tending machines, packaging and light assembly. The lesson

Computer use

Computer use lets an AI work through the screen of a piece of software as a person would: it reads the page, clicks buttons, types into fields and checks whether the task worked. The lesson

Context window

A context window is the total amount of text a model can consider in one request: your prompt, instructions, chat history, documents and tool results all share it. The lesson

Custom instructions and projects

Custom instructions are standing notes an assistant reads every time you use it: who you are, what the business does, how you write and what to avoid. The lesson

D

Data retention

Data retention is deciding what information to keep, for how long, and when to archive or delete it, so you avoid both needless exposure and losing records you still need. The lesson

Digital twin

A digital twin is a live digital model of a machine, cell, line or workflow that stays current through real operating data from sensors, controls and business systems. The lesson

Document parsing (OCR)

Document parsing is software that reads scanned files, photos and PDFs: OCR (optical character recognition) turns the printed words into text, and parsing then picks out useful parts such as names, dates, totals and tables. The lesson

E

Embeddings

An embedding is a list of numbers, produced by an AI model, that represents the meaning of text, an image or other data so software can compare meanings. The lesson

F

Fine-tuning

Fine-tuning means training an existing AI model further on a set of examples so it reliably repeats a behaviour, such as a set format, a tone or a labelling task. The lesson

Foundation model

A foundation model is a general-purpose AI model trained on huge amounts of data such as text, images and code, so it can do many tasks like answering questions, summarising and drafting. The lesson

G

Generative AI

Generative AI is software that produces new text, images, audio or code by following patterns it learned from huge numbers of examples. The lesson

Golden set

A golden set is a small, saved group of real example tasks, each with a known good answer, that you re-run whenever something changes to see whether quality got better or worse. The lesson

Guardrails

Guardrails are the rules, limits and approvals that control what an AI system may say, see and do, so it can be helpful without unlimited freedom. The lesson

I

Inference

Inference is the moment a trained AI model takes your request and produces an answer, using patterns it already learned rather than learning anything new. The lesson

Integration, really

An integration links two or more business tools so information passes between them automatically, for example a website enquiry landing in your CRM, instead of being retyped by hand. The lesson

J

Jev

Jev is TypeSafe AI's decision model: instead of writing text, it answers questions your software asks about a piece of text with typed answers, a choice, a score or a yes-or-no, each with a probability. The lesson

K

Knowledge base

A knowledge base is an organised, current collection of your business's answers and documents that people and AI tools can search. The lesson

L

Latency

Latency is the delay between sending a request to an AI system and getting the response back: the wait you feel. The lesson

LLM-as-a-judge

LLM-as-a-judge means using one AI model (a large language model, the kind that writes text) to score another model's answers against a written rubric, such as accuracy or format. The lesson

M

MCP

MCP, or Model Context Protocol, is an open standard that gives AI apps one shared way to connect to tools and data through a small connector program called an MCP server. The lesson

MHS

MHS usually means Material Handling System: the conveyors, mobile robots, lifts, sensors, software and rules that move products through a warehouse or factory. The lesson

Model distillation

Model distillation trains a smaller AI model to copy the answers of a larger, stronger one, so you keep much of the quality with lower cost and faster responses. The lesson

Model routing

Model routing means sending each task to the AI model that suits it best, using rules or an AI classifier, instead of sending every request to the same model. The lesson

Multimodal AI

Multimodal AI is an AI system that can work with several kinds of information, such as text, images, audio and video, in one task. The lesson

P

PII

PII, or personally identifiable information, is any data that can point to a specific person, such as names, emails, phone numbers, account numbers or dates of birth, and even combinations of details that identify someone. The lesson

Predictive maintenance

Predictive maintenance uses data from your equipment, such as vibration, temperature and power use, to spot early signs of wear so you can fix problems before they cause downtime. The lesson

Prompt caching

Prompt caching lets an AI provider store the repeated opening part of a prompt, such as standing instructions or long background text, so later requests only pay to process what has changed. The lesson

Prompt engineering

Prompt engineering is the practice of giving an AI clear instructions, useful context, limits and the output you want, so it produces a better result. The lesson

Prompt injection

Prompt injection is when untrusted text, hidden in an email, document, web page or message, tells an AI system to ignore its real instructions, reveal information or take an action the business did not intend. The lesson

Q

Quantization

Quantization shrinks an AI model by storing its numbers in a smaller, lower-precision format, such as 8-bit or 4-bit, so it needs less memory and can run faster on modest hardware. The lesson

R

RAG

RAG stands for Retrieval-Augmented Generation: the AI first searches your approved documents, notes or records for the relevant passages, then uses them as context to write its answer. The lesson

Re-ranking

Re-ranking is a step that re-orders the search results an AI system has found, scoring each one against the question so the most useful sources come first. The lesson

Reasoning effort

Reasoning effort is how much deliberate thinking an AI model does before it answers: more effort means more step-by-step checking and comparing, less means a faster, lighter reply. The lesson

Red teaming

Red teaming is deliberately trying to break your AI system, using tricky, misleading or hostile inputs, to find how it can fail before real customers do. The lesson

Reward model

A reward model is an AI model that scores how good a response is, having learned from people ranking answers by qualities such as accuracy, helpfulness and clarity. The lesson

RPA

RPA, or robotic process automation, is software that works a computer screen like a person: clicking, typing and copying between systems. The lesson

S

Semantic search

Semantic search finds information by meaning instead of exact words, so a search for machine repair can also find a page about fixing broken equipment. The lesson

Shadow testing

Shadow testing runs a new AI system quietly beside your current process, on the same real inputs, while customers only ever see the live system's answers. The lesson

Speech-to-text

Speech-to-text turns spoken words from calls, meetings, videos or voice notes into written text that you can read, search, save and use in other tools. The lesson

Structured output / JSON

Structured output means an AI model replies in fixed, labelled fields, usually as JSON (a plain-text format of labelled values), such as name, email and priority, instead of free-flowing prose. The lesson

Subagents

A subagent is a small, specialised AI helper that handles one part of a bigger job, such as research, drafting, routing or checking, and hands the result back to the main agent, which combines it and decides the next step. The lesson

System prompt

A system prompt is a set of instructions, usually hidden from the user, that tells an AI its role, rules, tone and boundaries before it reads any request. The lesson

T

Temperature

Temperature is a setting that controls how predictable or varied an AI model's answers are: low gives steady, repeatable answers and high gives more creative, varied ones. The lesson

Text-to-speech

Text-to-speech turns written words, such as a script, message or web page, into spoken audio in a synthetic voice that can be played live or saved as a file. The lesson

Token

A token is a small piece of text, such as a word, part of a word or a punctuation mark, that an AI model breaks your writing into before it processes it. The lesson

Tool calling / function calling

Tool calling, also called function calling, is how an AI model asks for an outside lookup or action, such as an order's status or a calendar booking, instead of only writing text. The lesson

U

Uptime

Uptime is the percentage of time an AI tool is available and working. The lesson

V

Vector

In AI, a vector is a list of numbers that captures the meaning of something, such as a word, sentence, image or customer note. The lesson

Vector database

A vector database stores numbers that represent meaning (called embeddings), so it can find content that is similar in meaning, not just content with the same words. The lesson

Vendor lock-in

Vendor lock-in is when your tools make it hard to leave, switch or negotiate, because your data is trapped, your workflows depend on one product and switching costs too much. The lesson

W

Webhook

A webhook is an automatic message one app sends to another the moment something happens, such as a new order, a form submission or a paid invoice, carrying the details of the event. The lesson