Pull the questions customers actually asked in the last three months from your inbox, phone notes, reviews and DMs, group them into topics, and write one short, self-contained answer per question with the price, conditions and exceptions spelled out. Add a list of questions the bot must hand to a person, then test it on 40 real questions before launch.
A website FAQ page and a chatbot FAQ do different jobs. On a page, readers see the entries above and below, so a vague "prices vary" gets away with it. A chatbot usually retrieves one entry, or a fragment of one, and answers from that alone. Every answer has to stand on its own and carry its own conditions, and most launch-week wrong answers come from entries that relied on context the bot never saw.
Mining the questions your customers already ask
Don't start by brainstorming what customers might ask. Start with what they did ask. Gather three months of real questions from these places, which takes most small businesses two to three hours:
- Email and contact-form messages. Search the inbox for a question mark, or for "how much", "do you", "can I" and "when". Copy the question sentences only.
- Social DMs and comments. Scroll back through Instagram, Facebook and WhatsApp. These skew towards price and availability, which is exactly what bots get asked most.
- Phone calls. You probably have no record, so keep a tally sheet by the phone for one week. A line per call is enough: "does the service include new brake pads?"
- Counter questions. Ask whoever serves customers to jot down the five questions they answer most. They'll know them without looking.
- Reviews. Complaints often reveal the question nobody answered in advance: "I didn't realise parts were extra."
Strip names, phone numbers, email addresses and order details before you paste anything into an AI tool. Then let the assistant do the sorting, which is the dull part:
Below are customer questions sent to my [type of business] over the
last three months, one per line, with personal details removed.
1. Group them into topics. Give each topic a short name.
2. Under each topic, list the distinct questions (merge near-duplicates
and keep up to four different phrasings of each).
3. Count how many raw questions fell into each topic.
4. Sort topics from most to least asked.
5. Flag any question that needs a judgement call rather than a fixed
answer (diagnosis, custom quotes, complaints, safety).
Don't write answers yet.
[paste questions]For a bike shop, an illustrative slice of the reply looks like this:
1. Service prices (77 questions)
- How much is a service? / What's a tune-up cost? / Price for a
basic service? / How much to get my bike looked at?
- Is it cheaper if I bring two bikes?
2. Turnaround (55 questions)
- How long does a service take? / Can I have it back today?
- Do you do while-you-wait repairs?
3. E-bikes (41 questions)
- Do you service e-bikes? / Can you fix my e-bike motor?
- How much is an e-bike service? [FLAG: motor question needs
a fixed yes/no from the owner]
4. Worth repairing? (32 questions)
- Is my old bike worth fixing? [FLAG: judgement call]Two things to fix in output like this. First, "How much to get my bike looked at?" has been filed under service prices, but in this shop a look-over is free, so it needs its own entry. Second, "Do you service e-bikes?" and "Can you fix my e-bike motor?" sit together, yet one answer is yes and the other is no. Assistants merge questions that share words even when the answers differ, so read every group with that in mind.
How many entries a small business actually needs
Most small service businesses end up with somewhere between 40 and 120 entries. Fewer than 40 usually means answers are too broad; more than 150 usually means you've started writing entries for questions nobody asks.
Use a coverage check rather than a target number. Take 50 recent questions you didn't use while building the list. If your entries answer at least 40 of them, or correctly mark them for a person, you have enough to launch. The long tail of odd questions should go to a human anyway, and trying to anticipate all of it is how FAQ projects stall for months.
In the bike shop, as an illustration, 36 of the 50 held-back questions matched an entry and 5 more fell under the handover list: 41, so enough to launch. The nine misses are worth a glance before deciding what to do with them. Four asked about bike fitting, which the shop does offer but had forgotten to write up, so that became a new entry. Three asked about hiring a bike, which it doesn't do, so they went into a "We don't" line. The last two were one-offs (a question about a 1970s frame and a request to store a bike for a month) and stayed with a person.
Writing answers a bot can't misread
This is where the work is. Here's a typical website-style entry from a bicycle repair shop, and a version written for a bot:
Before: "Services start from $45. See our price list for more details. E-bikes may cost more."
After: "A standard service costs $45 and covers gear and brake adjustment, wheel truing, a safety check and cleaning the drivetrain. Parts are extra. If we find a worn part we phone you with the price before fitting it, and we don't fit anything without your go-ahead. Drop the bike off before 10am on a weekday and it's usually ready the next working day; from March to June allow three working days. A standard service for an e-bike costs $65. We don't repair e-bike motors or batteries."
The second version is longer, but a bot can answer six different questions from it without guessing. The rules behind it:
- One question per entry, with two to four other ways customers phrase it.
- Answer in the first sentence. The price, the yes or the no comes first; explanation follows.
- Write numbers in full. "$45", "three working days", "before 10am". Never "a small charge" or "a few days".
- Put conditions inside the entry. If the answer changes for e-bikes, weekends or new customers, say so in the same entry.
- No references outward. "As above", "see our terms" and "this" pointing at another entry all break when the bot retrieves the entry on its own.
- Say what you don't do. Bots fill silence with helpful guesses. "We don't repair motors" prevents the bot cheerfully booking one in.
- Write in the voice you want the bot to use. Bots copy the tone of their source material, so a stiff FAQ produces a stiff bot.
Letting an assistant draft entries from your notes
Writing the entries is quicker if you dictate or jot rough notes and ask an assistant to put them in the format shown further down. The instruction that matters most is the last line:
Turn my notes into one chatbot FAQ entry with these fields:
Question, Also asked as, Answer, Conditions, We don't, Hand over if.
Answer first, full numbers, friendly and plain.
Use ONLY facts in my notes. If a field has no facts, write "?".
Notes: e-bike service 65 dollars, same as normal service plus
motor diagnostic readout and firmware check if brand allows.
no motor or battery repairs. spring = 3 days.An illustrative draft came back with most fields right, plus this in the answer: "All e-bike work comes with our 30-day workmanship guarantee." The notes never mentioned a guarantee. The shop does have one, but it's 60 days and excludes electrical parts, so the invented line was wrong twice. Read every AI-drafted entry for sentences you didn't supply; they are usually plausible and often close to true, which is what makes them easy to miss. The "?" instruction helps: fields the model can't fill show up as gaps instead of guesses.
The entries most small businesses forget
Once the obvious price and opening-hours entries exist, check for these. Each one is a question that sounds rare but arrives every week, and each is a place where an unprepared bot improvises:
- What you don't do. Brands, services or job sizes you turn away.
- What happens if the job costs more than expected. Customers ask this before they book, and the honest answer builds trust.
- Busy-season lead times. The spring rush, the pre-holiday rush. Date the entry so it gets reviewed.
- Payment, deposits and when you charge. Which cards, whether you take cash, whether a deposit is refundable.
- Drop-off and collection outside opening hours. Can a bike be left overnight? What happens if it isn't collected?
- Guarantees on your work. How long, what's covered, what voids it.
- Bringing your own parts. Whether you'll fit them and whether you guarantee the result.
- Access and parking. Step-free entry, where to stop to unload.
- How fast you reply to messages. So the bot can set expectations when it hands over.
The handover list: questions the bot must pass to a person
An FAQ needs a second list alongside it: topics the bot recognises but doesn't answer. Write it as carefully as the answers, because it's what stops the bot inventing a diagnosis or a refund policy.
| Question type | Why the bot shouldn't answer | What it should say instead |
|---|---|---|
| Diagnosis from a description ("my gears slip, what's wrong?") | It will guess, confidently | Offer a check-up booking, or ask for a photo or video for the mechanic |
| Exact quotes for custom work | Depends on things only a person can see | Give the price range from the FAQ and take details for a proper quote |
| Safety problems ("my brakes failed") | Delay or bad advice could hurt someone | Tell them not to ride it and give the phone number for the same day |
| Complaints, refunds and disputes | Needs judgement and authority | Apologise, collect the details, promise a reply within a stated time |
| Status of an existing job | Unless connected to your job system, it can't know | Take the name and job reference and pass it to the team |
| Insurance or crash-damage claims | Wording has consequences | Hand over, and say a person will reply |
How and when the bot passes the conversation on is its own design question; when an AI chatbot should hand over to a human covers the triggers and wording.
A format that pastes into any chatbot tool
Keep the master copy in a spreadsheet, one row per entry, with these fields. Most chatbot tools accept uploaded documents, question-and-answer pairs or a website page, and this format converts to any of them:
ID: SRV-01
Question: How much is a bike service?
Also asked as: service price | cost of a tune-up | how much to service my bike
Answer: [direct answer first, full numbers, voice you want]
Conditions: [what changes the answer: e-bikes, season, new customers]
We don't: [anything adjacent you don't offer]
Hand over if: [the point where a person should take over]
Owner: [who keeps this entry true]
Last checked: [date] Review by: [date]Filled in for the bike shop's most-asked entry, it looks like this:
ID: SRV-01
Question: How much is a bike service?
Also asked as: service price | cost of a tune-up | how much to service my bike
Answer: A standard service is $45. It covers gear and brake
adjustment, wheel truing, a safety check and a
drivetrain clean. Parts are extra, and we phone you
with a price before fitting any.
Conditions: E-bike standard service is $65. From March to June
allow three working days instead of one.
We don't: Repair e-bike motors or batteries.
Hand over if: The customer describes a fault ("it makes a noise")
or wants a price for a specific repair.
Owner: Workshop lead
Last checked: 1 Sep Review by: 1 DecThe "Owner" and "Review by" fields look like admin, but they're what keep the bot honest six months from now. Training an AI chatbot on your FAQs, policies and prices explains how to load content like this into the common tools and what each one does with it. If you also want the same material to serve staff, the approach in building a company knowledge base AI can answer from scales it up.
Testing the FAQ before the bot goes live
Hold back 40 real questions that you didn't look at while writing. Put each one to the bot exactly as the customer wrote it, typos included, and score the reply:
- Right: correct and complete.
- Handed over well: it recognised a handover topic and said the right thing.
- Incomplete: correct but missing a condition that matters.
- Wrong: anything false, or a promise you wouldn't keep.
Three illustrative rows from a practice round the bike shop ran on a handful of questions, before the entries were finished, show how the scoring works:
| Question as the customer typed it | What the bot said | Score | Fix |
|---|---|---|---|
| "hw much for ebike servce" | $65 for a standard e-bike service; parts extra; no motor repairs | Right | None |
| "brakes squeal really bad is that dangerous" | Suggested cleaning the rims and replacing the pads | Wrong | Add "squeal" and "grinding" to the safety handover; the bot should say not to ride and give the phone number |
| "do you do bmx" | "Yes, we service all types of bike" | Wrong | New "We don't" line: the shop doesn't stock BMX parts and refers those jobs elsewhere |
The BMX reply is the classic silence-filling guess: with no entry saying otherwise, "all types of bike" sounded helpful. The brakes reply is the more serious of the two, because it gave repair advice for what could be a safety fault.
My launch bar is zero wrong answers about prices, safety or policies, and at least 32 of the 40 either right or handed over well. Every "incomplete" or "wrong" traces back to an entry, so fix the entry rather than adding an instruction to the bot. For a fuller pre-launch routine, including trying to trick it, see how to test a customer chatbot before it goes live.
Building it for a two-mechanic bike shop
Say a small bike shop pulls three months of messages: about 410 emails and DMs, plus a week of phone tallies with 45 calls. Clustering produces 58 topics. The biggest are service prices (17% of questions), turnaround times (12%), e-bike work (9%) and whether a bike is worth repairing at all (7%). That last one goes on the handover list with a line inviting the customer to bring it in for a free look.
Writing 58 entries takes one of the owners about six hours over two evenings, using an assistant to turn rough notes into the format above and then editing each answer by hand. The first test scores 29 right, 5 handed over well, 2 incomplete and 4 wrong. All four wrong answers involve e-bikes or spring lead times, because the original entries said "may cost more" and "allow extra time". After rewriting those six entries with real numbers, the retest scores 34 right, 4 handed over well, 2 incomplete and none wrong.
Running costs depend on the tool. On per-resolution pricing, 120 resolved conversations a month would cost about $119 at the $0.99 per outcome Intercom lists for its Fin agent, or about $60 with HubSpot's Customer Agent at roughly $0.50 per resolved conversation in credits, before any seat or plan fees (Customer Agent needs a seat on any Professional or Enterprise hub (usually Service Hub), which is the larger cost for a small shop). The better the FAQ, the more conversations resolve, so the bill rises as quality does. Budget for that rather than being surprised by it.
Keeping the FAQ true after launch
A chatbot FAQ goes stale faster than a web page because nobody reads it back. Build in three habits:
- Change the entry the same day you change the fact. New prices, new hours, a service you've stopped. Put "update the bot FAQ" on the same checklist as updating the website.
- Read the unanswered-questions list weekly. Most tools log questions the bot couldn't answer or handed over. Each recurring one is a candidate for a new entry.
- Rerun ten test questions monthly. Pick the highest-risk ones: prices, lead times, anything seasonal.
A typical way this goes wrong: the opening-hours entry says "Open Monday to Saturday, 9am to 5.30pm", with no mention of public holidays. On a holiday Monday the bot tells three customers the shop is open, and two of them arrive to a locked door. The entry wasn't false; it was incomplete. The fix is a line in the same entry ("We close on public holidays; our holiday dates for this year are…") and a reminder in the calendar to update it each January.
If wrong answers still slip through after that, the cause is usually an old document the bot can still see, or a website page that contradicts the FAQ. Stopping an AI chatbot giving customers wrong answers walks through tracking those down.
Further reads
- Best AI Chatbots for Small Business Websites — Which chatbot tools accept a Q&A knowledge base like this one.
- Chatbot Guardrails: Stop AI Promising What You Don't Offer — Stop the bot promising services your FAQ says you don't offer.
- AI Chatbot Disclosure: What to Tell Customers at the Start of a Chat — The opening message that tells customers they're chatting to AI.
- How to Measure Whether Your AI Chatbot Is Actually Working — The numbers that show whether the FAQ is doing its job after launch.
- What to Ask an AI Chatbot Vendor Before You Sign Up — What to ask a vendor about how their bot uses your content.
- Can You Build an AI Chatbot for Your Business Without Coding? — Turning this FAQ into a working bot without a developer.
- What Business Data Should You Start Collecting Now for AI? — Seven datasets worth capturing from today (enquiries, quotes, job actuals, questions, complaints, prices, feedback), with the fields that make them usable.
- AI Mistakes That Damage Customer Trust, and How to Avoid Them — Nine AI mistakes customers notice, why each one stings, how to prevent it, and a 20-minute monthly check that catches problems before customers do.
- Pet Grooming Software With AI: Features and Prices Compared — What the AI in MoeGo, Teddy, Groomify, DaySmart Pet and Gingr actually does, what each costs a year, and a 14-day trial script for groomers.
- Best AI Chatbots for Small Shopify Stores (2026) — Shopify Inbox now has a free AI agent. When it's enough, when Tidio or Gorgias earns its fee, and a 30-question test to pick the right one.
- AI Booking Systems for Appointment Businesses: What to Check — A 24-point checklist for choosing an AI booking system, grouped by risk, with how to verify each item and a two-hour test script to run before you sign.
- DIY or Get Help? Setting Up AI Bookings for a Small Business — Six questions that tell you whether to build AI bookings yourself, with three worked cases and a weekend plan if you go it alone.
- How Wedding Venues Use AI for Viewings, Enquiries and Follow-Ups — Use AI at three points in a venue's sales pipeline: the first reply, the notes after each viewing, and follow-ups timed around provisional holds.
- How Dog Groomers Can Use AI to Collect Pet Details Before Booking — Collect breed, coat condition, temperament and health details before the booking, and let AI turn them into a pet card and suggested slot you approve.
- How Day Spas Can Take Bookings From Instagram DMs With AI — Three ways to put AI in a spa's Instagram inbox, the knowledge sheet it needs, when to hand over to a therapist, and a 20-message test before launch.
- How Boutiques Can Answer Customer DMs Within an Hour Using AI — A three-layer system for boutique DMs: automatic acknowledgement, AI-drafted answers from a fit-and-stock sheet, and a rota that keeps replies under an hour.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Intercom pricing page (Fin per-outcome price) and HubSpot credits pricing, checked September 2026.