How do you keep AI on-brand in front of customers?
Written lesson · about 3 minutes to read
The short answer
Give the tool four things: a short voice guide, real examples of your best replies, a list of approved facts such as hours, prices and policies, and a list of topics it must hand to a person. Then read a sample of its real conversations every week and correct what it gets wrong.
Three things to remember
- 01Write a short voice guide and give the tool real examples of your best replies.
- 02Give it a list of approved facts so it does not invent prices, hours or policies.
- 03Name the topics it must hand to a person, and read a sample of conversations weekly.
Start with your voice
Write a voice guide of one page or less. Say how you speak: plain words, short sentences, first names, warm but brief. List the words you use, and the words you never use, such as stiff phrases you would not say to a customer's face.
Then add real examples. Pick ten replies you are proud of, strip out the customer details and paste them in. Examples teach the tool more than adjectives do. Tell it to match their length, greeting and sign-off. Keep the guide and examples in one shared place, so anyone setting up a new tool or chat can paste them in.
Give it facts and limits
Write an approved list of facts: your hours, the area you cover, what you charge or how you quote, your warranty, your cancellation rules and how booking works. Tell the tool to use only those facts and to say when it does not know. A date should come from your calendar, or be left as a marked blank for a person to fill in. Update the list the day a price or policy changes.
Decide which topics always go to a person: complaints, refunds, anything about safety, a formal quote, an angry customer, anyone who says they want to cancel or asks for a human. An apology can be drafted. A promise belongs to a person.
Read real conversations every week
Set aside twenty minutes each week to read a sample of what the tool actually said to customers. Look for three things: a wrong fact, a wrong tone, and a moment when it should have handed over and did not.
Fix the cause as well as the reply. That may mean adding a fact to the list, replacing a weak example or adding a topic to the hand-over list. Keep a note of each change so you can see the tool improving.
Where to start this week
Write the voice guide today, and gather your ten best replies. Draft the approved-facts list from your website and your price sheet, then ask a colleague what you left out.
Before customers see it, send it ten typical customer messages and read the replies as if you were the customer. When you switch it on, make sure a customer can always reach a person, and that the chat says it is automated. Put the weekly read in your calendar now, or it will not happen.
More in part 9
- 1How do you check an AI answer before you use it?Written
- 2Where should humans stay in the loop?
- 3What are guardrails?
- 4How do you keep AI on-brand in front of customers?WrittenYou are here
- 5What is AI bias, and can it hurt your business?Written
- 6What happens when an agent gets it wrong?
- 7What is an audit trail, and why will you need one?
- 8What happens when the model updates and your results change?
- 9What is a reward model?
- 10What are AI evals?
- 11What is a golden set?
- 12What is LLM-as-a-judge?
- 13What is calibration (and why confidence scores matter)?
- 14What is shadow testing?
- 15What is red teaming?
- 16What is AI observability?
Next step
Where does this fit your business?
Six quick questions about your business, then your AI plan: what to start with, what a person still checks, and the lessons that go with it. Or bring it straight to a free thirty-minute call.