Your first website chatbot in the UK: what to teach it before it talks

A visitor asked if you could come tomorrow, and the chat said yes. You're already booked that morning. They've told their partner it's confirmed.

You'll write ten answers in your own words before anyone else sees that chat.

A website chatbot in the UK is only as useful as those answers. It can greet people at any hour. It can take a name and a question.

It should already know what you do, where you work, how a booking starts, and what you won't guess. If the sheet is blank, the bot fills the gap. That's how a polite tool becomes a wrong quote.

The wider choice, chat or a form, is in chatbot or contact form. Write the answers first. Then let the bot use them.

Teach facts you can stand behind

A lead capture chatbot has one job on day one. It helps a visitor work out whether you can help, and it takes a clear message when a person needs to take over. Clever is optional. Right about your business is the job.

Write the answers on one sheet, in the words you'd use on the phone. Short sentences. The same words you'd want a new starter to use.

AdviserAlly describes its chatbot as replying in your business's own words. That only works if you've supplied the words.

Read each answer aloud. If you hesitate, it isn't ready. If a colleague would phrase the price rule differently, agree the sentence before the bot says either version. Two "official" prices is how you end up apologising on the phone.

The NCSC's note on large language models says these models can get things wrong and state incorrect facts. They can be biased. They can sound sure when they aren't. Where your sheet is silent, the line is "I don't have that. I'll ask a person to reply." The fuller NCSC warnings are in the FAQ.

Ten minutes: read your homepage aloud and write the one sentence that says what you do.

Ten questions for your website chatbot in the UK

These are the ten questions your bot must answer well before it talks to customers. Put the finished sheet somewhere the whole team can edit. When a fact changes, change the sheet the same day.

Change "Not written" to "Checked" only after someone who didn't write the sheet has asked the question and the reply matches. Ten checked rows is the bar for day one. Extra questions can wait.

You'll be tempted to add an eleventh row about parking. Leave it until the ten are true.

Tone is a separate job. Once the facts are right, chatbot replies that sound like you is where you tune the voice. Facts first. Voice second.

Ten minutes: ask question 3 out loud and write the price rule you use, with no figure you haven't published.

What the bot must refuse to guess

Write a short stop-list and put it on the same sheet.

Prices you have not published. Appointment slots you have not checked. Promises about a regulated product.

Complaints. Anything a visitor describes as urgent or unsafe. Legal conclusions.

Medical conclusions. The contents of a previous customer's job.

On each of those, the reply is a handover. "A person needs to answer that. Leave your name and number, and when they'd like a call." Put that sentence on the sheet, word for word, so the bot doesn't soften it into a promise.

The NCSC point underneath is simple. A model can be wrong. You write the sentences, and the bot repeats them.

If you're also weighing live chat or an AI receptionist, the labels are untangled in chatbot, live chat and AI receptionist. Day one is the same either way. Somebody still writes what the business will stand behind.

Ten minutes: list the three topics you will never let the bot answer, and paste the handover sentence under them.

Tell people what happens to their details

A chatbot that asks for a name and a phone number is collecting personal data. GOV.UK's guide for businesses says you follow the data protection rules if you store or use customer details, and you keep them secure, accurate and up to date. When the bot asks for a name, it also says who you are and how you'll use it.

The ICO's chapter on transparency in AI says you give that privacy information at the moment you collect the details, if you collect them directly. Question 8 on the sheet is those few sentences, plus a link to the full notice. Write the retention period you use.

If you haven't decided it, say "the period is in our privacy notice" rather than inventing one. The rights to see, correct or delete, and the ICO's list of what the notice should cover, are in the FAQ.

The bot doesn't have to finish a legal request itself. It does have to show a person, and that path has to work.

Ten minutes: write the three sentences for question 8 and put the link to your privacy notice under them.

Cookies and the chat box

Chat tools often store a session on the visitor's device so the thread survives a refresh. Some also set analytics or marketing cookies. You need to know which yours does.

The ICO's guide to cookies and similar technologies says you tell people, explain the purpose, and get consent before non-essential storage. If you can't say what the widget stores, you don't switch it on. You can't explain what you don't know. The consent rule, the essential-cookie exception, and what counts as a clear explanation are in the FAQ.

Before you launch, ask the supplier, or look in the browser tools you already use. Then do one of these.

• If the storage is essential to the chat they opened, say that plainly in your cookie information.
• If the tool also measures marketing, or tracks the visitor across sites, wait for consent before those cookies are set.
• If you still can't tell, leave the widget off.

A banner that says you care about privacy, then sets a pile of trackers, is a familiar little performance. Don't copy it.

That check sits beside question 8. Privacy lines cover the name and number. Cookie lines cover the device. Visitors meet both in the same minute.

Ten minutes: open your site, start a chat, and note every cookie the widget sets before anyone has agreed.

Keep client secrets out of public tools

Teaching the bot isn't the same as pasting yesterday's customer emails into a public chat website "to see what it would say". You'll get a tidy paragraph. You'll also have handed over someone else's details.

The NCSC article says you shouldn't put sensitive information into a public model. A client's finances, a health comment on a booking form, a complaint, a staff issue, or an address paired with a time the house will be empty, stay out of that box. The storage warning is in the FAQ.

Write the bot's answers from your own sheet, inside the tool you've chosen for the website. If you use any AI to prepare a reply, start from the question ("how should we explain a call-out fee?") and leave the customer's details out. The customer detail stays on your own list.

The same NCSC piece says these tools produce convincing text. A visitor can't hear the hesitation you'd have had on the phone. Your sheet is that hesitation.

Ten minutes: look at the last prompt you typed into a public tool and delete it if a customer's name was in there.

A test you run before the switch

Sit down with one other person if you have one. If you work alone, leave the sheet overnight and test it the next morning. Ask all ten questions in ordinary language, not in the headings from the table.

"Do you cover Leeds?" is a fair way to ask question 2. "Can you come tomorrow?" is a fair way to ask questions 4 and 5.

Pass only if:

• The reply matches the sheet.
• The reply refuses the stop-list topics and offers a person.
• The privacy sentences appear when the bot asks for a name or number.
• A made-up detail, such as a discount you do not offer, never appears.

Then ask one messy question: "My neighbour said you were cheap and could come tonight and also do the paperwork for a pension." The right reply splits that into what you do and what you will not answer in chat. If the bot tries to please the whole sentence, it is not ready.

Fix the sheet. Test again. Launch when the ten rows say "Checked".

After launch, read the chats. AdviserAlly says the built-in assistant keeps track of every conversation so you know who to get back to. Use that list.

A question the bot couldn't answer can become row eleven later. A question it answered wrongly is a same-day correction.

The better morning is quiet in a specific way. Overnight chats match the sheet. The one the bot couldn't answer is a named message, with a phone number, waiting for you. Nothing on the list is a price you don't charge.

Your chatbot, once taught, is ten answers, a stop-list, a privacy sentence at the moment you ask for details, a cookie check, and a rule that client secrets stay out of public tools. The plain-English account of what this kind of CRM does sits next to this list. The bot handles the first reply. You make the calls that need a person.

Ten minutes: ask the bot "can you come tomorrow?" and "what do you charge?" and keep it offline until both replies match the sheet.