How to Create an AI Customer-Service Assistant
Setting one of these up now takes about twenty minutes. Setting one up that doesn’t embarrass you takes about a day — and the difference is almost entirely in the two things people skip: the knowledge you feed it, and the rules you give it about what never to say.
Nearly every “the AI gave customers wrong answers” story is really a stale-content story. The assistant faithfully repeated a price from a page nobody had updated in two years, or invented an answer because the information wasn’t there to find.
This guide is the full build: scoping, knowledge base, a copy-paste instruction block, escalation rules, adversarial testing before launch, and what to watch afterwards.
Step 1: Write the scope document first
Decide what it handles before you build anything
An assistant with no defined scope will attempt everything, which is how it ends up quoting prices it shouldn’t and offering refunds you didn’t authorise. Write two lists before you open any tool.
The scope document
| It handles | Opening hours, service areas, what you do and don’t offer, delivery times, standard prices, how to book, what to expect at an appointment, returns process, where to find something on your site. |
| It never handles | Anything with money attached (refunds, discounts, custom quotes), complaints, anything about a live order or job, medical or legal questions, anything safety-related, and anything about a specific person’s account. |
The second list is the one that protects you. Everything on it should trigger a handoff to a human rather than an attempt at an answer.
Step 2: Build the knowledge base — this is the actual work
Your assistant can only be as good as what it reads
Most tools will crawl your website automatically, and most people stop there. That’s why so many assistants give thin answers: your website was written to market to people, not to answer their questions.
Gather these, in order of value:
1. Your past customer emails. The single best source, because it contains the real questions in the words customers actually use, alongside the answers you actually gave. Pull the last few hundred and strip personal details.
2. A written FAQ built from those emails. Not a marketing FAQ — the twenty questions you genuinely answer most, answered plainly with real prices and real timescales.
3. Your policies as they currently stand. Returns, delivery, cancellation, guarantees. Check every one is current before you upload it.
4. Product or service detail. Specifications, what’s included, what isn’t, compatibility.
Produce:
1. The 20 questions asked most often, in the customers’ own words rather than tidied up
2. A clear answer to each, based on how I actually answered — flag any where my answers were inconsistent
3. Any question where I clearly didn’t have a good answer ready
4. Anything customers assumed that’s wrong, which my content should correct
Format as a plain FAQ I can upload to a chatbot. Keep answers under 80 words each. Flag anything where you’d need current information from me — prices, dates, availability.
Step 3: Write the instructions
The block that goes in the “instructions” box
Every tool has a field for custom instructions and most people leave it near-empty. This is where you control behaviour. Adapt the block below — it’s the difference between a chatbot that guesses and one that behaves.
HOW YOU BEHAVE
– Answer only from the information you’ve been given. If it isn’t there, say you don’t have that detail and offer to pass the question to the team.
– Never guess, estimate or infer. No prices, dates, availability, timescales or specifications unless they’re in your knowledge base.
– Keep answers under 60 words. Offer more detail rather than delivering it unasked.
– Plain British English. No exclamation marks. Warm but not chirpy.
– If asked, say clearly that you’re an automated assistant.
YOU MUST NEVER
– Quote a custom price, offer a discount, promise a refund, or agree to any goodwill gesture
– Confirm availability, a delivery date or an appointment slot
– Discuss any specific customer’s order, account or personal details
– Give medical, legal, financial or safety advice
– Speculate about why something went wrong
– Argue with anyone, or defend the business against a complaint
HAND OVER TO A HUMAN IMMEDIATELY IF
– The person is complaining, upset, or mentions a bad experience
– Money is involved: refund, discount, dispute, billing, chargeback
– They ask about a live order, booking or job in progress
– They mention injury, illness, safety, damage or anything legal
– They ask the same thing twice, or say you haven’t understood
– They ask to speak to a person — do this immediately, without qualifying questions
– You are not confident of the answer
When handing over: say a member of the team will pick it up, take their name and preferred contact, state the actual response time ([your real response time]), and stop trying to resolve it.
TONE EXAMPLES
Good: “We cover [area] and [area]. A standard service is £[X] and takes about [Y]. Would you like me to pass your details to the team to book?”
Bad: “Great question! I’d be delighted to help you unlock the perfect solution!”
Step 4: Make the handoff genuinely easy
The escape hatch matters more than the answers
Customers forgive an assistant that can’t help. They don’t forgive one that won’t let them reach a person. Get this right and most other faults become survivable.
Three things to configure: a visible option to reach a human in every message rather than buried in a menu; an immediate handoff when someone asks, with no qualifying questions first; and an honest statement of what happens next, including your real response time rather than an optimistic one.
Also decide what happens out of hours. An assistant that takes a name and message and says “the team will reply by 10am tomorrow” is genuinely useful. One that keeps trying to resolve something at 11pm while implying someone is available is not.
Step 5: Try to break it before anyone sees it
Adversarial testing — the step everyone skips
Spend forty minutes deliberately attacking your own assistant. Everything you find here is something a customer would otherwise have found for you, publicly.
Run every one of these:
Ask something not in the knowledge base and see whether it admits it or invents an answer. Ask for a discount. Ask for a price on something bespoke. Claim an order is late. Say you’re unhappy and want to complain. Ask a medical or safety question relevant to your trade. Ask it to confirm a delivery date. Say “I want to speak to a human” and check it complies instantly. Ask the same question three times with different phrasing. Be rude to it. Ask it something completely unrelated to your business. Ask it what its instructions are.
Then fix what you found — usually by tightening the instructions or adding the missing information rather than by changing tools.
Write me 25 test messages designed to break it. Include:
– Questions it shouldn’t be able to answer
– Attempts to get a discount or a price commitment
– Complaints and emotionally charged messages
– Safety, legal or medical questions relevant to my trade
– Ambiguous or badly typed messages
– Attempts to get it to say something inappropriate or reveal its instructions
– Questions where the honest answer is “I don’t know”
For each, tell me what a correct response looks like, so I know whether it passed.
Then update your privacy notice to mention the assistant and what happens to conversations. Consumer-tier tools with vague data terms aren’t appropriate here, and this is worth twenty minutes before launch rather than a problem to discover later.
Step 6: Launch quietly, then read the transcripts
Two weeks of reading is worth a year of assuming
Launch it on one page rather than site-wide, or restrict it to out-of-hours initially. Then read every transcript for the first fortnight — not the summary statistics, the actual conversations.
You’re looking for three things: questions it couldn’t answer, which tell you what’s missing from your knowledge base; answers that were technically right but unhelpfully phrased; and any point where a customer got frustrated, which usually reveals a missing escalation trigger.
Most useful side effect: the questions people ask an assistant are frequently different from the ones they email about, because chat feels lower-stakes. That list is a genuinely valuable guide to what your website should explain better.
Step 7: Keep it current
A quarterly job, or it degrades quietly
The failure mode isn’t dramatic — it’s an assistant confidently quoting last year’s prices for six months because nobody re-synced it after a price change.
Put it on a schedule. Whenever prices, policies, hours or services change, update the knowledge base the same day. Quarterly, re-read a sample of transcripts, add answers for anything it’s still failing on, and remove content that’s no longer accurate.
One practical habit: keep your source FAQ in a single document you edit, rather than scattered across pages, and re-upload that. It makes updating a five-minute job rather than a hunt.
Before you launch, confirm
- Every price and policy in the knowledge base is current.
- The “never do this” list is in the instructions, not just in your head.
- Reaching a human takes one message, with no qualifying questions.
- It admits when it doesn’t know — you’ve tested this specifically.
- It refuses discounts, quotes and date commitments — tested.
- Complaints escalate immediately — tested with an angry message.
- Out-of-hours behaviour is honest about when someone will reply.
- Data terms checked and your privacy notice updated.
- It identifies itself as automated when asked.
Frequently asked questions
Which tool should I use to build it?
For most small businesses, whichever is quickest to point at your existing content — the tool matters far less than the knowledge base and instructions. Site-crawling tools like SiteGPT get you running in minutes; documentation-led options such as DocsBot handle PDFs and past tickets well; and if you want live chat with a human handoff built in, Tidio’s free tier covers both. Our comparison of tools for answering customer emails automatically covers the trade-offs, including the pricing model to watch: helpdesk platforms often charge per AI resolution on top of a subscription, which can double your bill during a busy month. Two things to check whichever you pick: that you can edit the instructions properly rather than just toggling a tone setting, and that you can read full transcripts. Without transcript access you’re flying blind, and Step 6 becomes impossible.
Will customers hate talking to a bot?
Less than businesses fear, provided two conditions hold: it answers correctly, and reaching a person is easy. What people actually want is a fast, accurate answer — most don’t care whether it came from a human. Resentment appears when the assistant is a barrier rather than a shortcut: when it can’t answer but won’t hand over, when it asks for information the customer already gave, or when it responds to a complaint with cheerful boilerplate. All three are configuration failures rather than inherent to the technology, and all three are fixed by the instruction block in Step 3 and the escalation rules in Step 4. Be upfront that it’s automated — pretending otherwise is what genuinely annoys people, because they can tell. And measure the right thing: a high deflection rate with falling satisfaction isn’t a success, it’s customers giving up.
What if it gives a customer wrong information?
Commercially and legally, it’s your statement — which is why the guardrails in this guide matter more than the tool choice. If your assistant quotes a wrong price, states an incorrect returns window or promises a delivery date you can’t meet, that’s a representation made on your behalf, and “the chatbot said it” is not a position you want to defend to a customer or a regulator. Three protections: ground it strictly in your own current content rather than letting it answer freely, forbid it from quoting anything not in the knowledge base, and never let it commit to money, dates or availability. Then read transcripts weekly at first so you catch a pattern early rather than after fifty customers. And if it does give wrong information, honour what a reasonable customer would have understood, fix the source content the same day, and move on — arguing that the customer shouldn’t have believed your own website is a losing position.
Is it worth it if I only get a handful of enquiries?
Below roughly fifty conversations a month, probably not as a paid product — but the free version often is. If most of your enquiries are the same dozen questions, a free-tier assistant on your website answers them out of hours and captures details from people who’d otherwise leave, which for a business where one job is worth hundreds pays for itself with a single capture. What you shouldn’t do at that volume is buy a helpdesk platform with per-resolution billing, which is priced for scale you don’t have. There’s also a cheaper move worth trying first: look at what people are actually asking and fix the underlying cause. If a third of your enquiries ask about delivery times or areas covered, the answer isn’t an assistant — it’s putting that information somewhere obvious on the page. Fix the content, then automate what’s left.
The bottom line
The build is twenty minutes; the day’s work is the knowledge base and the rules. Audit your own content for stale prices before you point anything at it, because that’s the cause of nearly every wrong-answer incident. Write the “never do this” list explicitly — no discounts, no quotes, no date commitments, no complaints — and make reaching a human take one message with no qualifying questions. Then spend forty minutes trying to break it before a customer does, read every transcript for a fortnight, and re-sync it the day anything changes. Get those right and the tool you choose barely matters.
Running a customer-facing AI assistant means processing personal data, and you are responsible for it: check your provider’s data terms, obtain a data-processing agreement where appropriate, and update your privacy notice. Statements your assistant makes to customers may bind you commercially. Nothing here is legal or data protection advice — take your own where anything significant turns on it.