You don't retrain the AI model itself; you give the chatbot a knowledge source it searches before every answer. Rewrite your FAQs, policies and prices as short single-topic entries, state each price as a rule with its conditions and a date, load them into the bot's knowledge settings, restrict it to those sources, then test with 50 real customer questions.
Most "trained" chatbots that misquote prices aren't short of content; they have too much of it, and it disagrees. An old PDF brochure says one fee, the website says another, and a two-year-old news post mentions a launch offer. The bot finds all three and picks one. Deleting and reconciling content does more for accuracy than adding it.
What "training" actually means inside chatbot tools
When a chatbot platform says it trains on your data, it nearly always means retrieval: your content is split into chunks, stored in a searchable index, and when a customer asks something the tool finds the most relevant chunks and hands them to a language model with an instruction to answer from them. The model's underlying knowledge doesn't change. That's good news, because you can fix a wrong answer by editing a paragraph rather than by any technical work.
It also explains the failure patterns. If the right chunk isn't found, the bot answers from something nearby. If two chunks contradict each other, it may blend them. If the price lives in an image, the index never saw it. The tools differ mainly in which sources they accept and how they keep them fresh:
| Tool | Sources it takes | Freshness and limits worth knowing |
|---|---|---|
| Intercom Fin | Help-centre articles, short "snippets", uploaded PDF or DOCX files, synced websites and other knowledge bases | Uploaded documents must be text-based (images and multi-column layouts aren't read; plain tables are), up to 45 MB each and 100 per workspace; website content re-syncs on a schedule rather than instantly |
| Chatbase | Files, website crawls, typed Q&A pairs and text | Automatic re-sync of sources starts on the Standard plan ($150 a month); on cheaper plans you retrain by hand after edits |
| Tidio Lyro | Website content and your own Q&A entries | AI conversations are metered separately from the base plan, so testing uses up allowance |
| HubSpot Customer Agent | HubSpot knowledge-base articles, website and landing pages, public URLs and uploaded files | Knowledge-base articles re-sync when you edit them; other sources re-sync weekly. Only on Professional and Enterprise tiers |
| Shopify Inbox AI agent | Your store's policies and product information | Can use web search as a secondary source, so test it on your own policy questions before trusting it |
Whichever tool you use, check two settings before loading anything: whether the bot is allowed to fall back on general knowledge or web search when your content has no answer, and what it says when it can't find one. You want the fallback off and a clear handover message on.
Audit what you already publish before writing anything new
Take a six-person surveying firm as an illustration. It wanted a website chatbot because roughly 40% of its enquiries were some version of "how much is a survey on a three-bed house?". Before loading a word, the office manager searched for every place a fee appeared. She found five:
- The fees page on the website, updated in spring.
- A PDF price list linked from an old page, still showing last year's fees.
- The email template staff paste into replies, which quoted a different band for properties over a certain value.
- A "spring offer" news post, long expired but still live.
- The booking form's help text, which said "from" a figure that no longer existed.
A crawl of that website would have taught the bot four different prices. The first job was deleting the PDF link and the news post, fixing the template and the form, and deciding that one written price entry would be the single source for the bot. That took about an hour and prevented most of the errors that would otherwise have shown up in testing.
Run the same search on your own material: website pages, PDFs, email templates, quote templates, terms and conditions, and anything pinned in a social profile. List every price and policy statement, mark the current one, and remove or fix the rest. If you're building the question list from scratch, building the FAQ your chatbot needs covers mining your inbox for the questions people actually ask.
Split long FAQ pages into one-question entries
Retrieval works best when each chunk answers one question completely. A long FAQ page that meanders through several topics in one paragraph gets chopped at arbitrary points, and half an answer is worse than none. Here is a before-and-after from the same surveying firm.
Before, one FAQ paragraph read:
We cover most property types and our surveyors are all qualified. Reports usually take a few days and we can often visit within the week, although it's busier in spring. Prices depend on the property, and we're happy to talk you through the options, including whether you need a full building survey.
That paragraph mentions five topics and answers none. After the rewrite, it became four separate entries, each starting with the question in the customer's words:
- How soon can you visit? Usually within 5 to 8 working days of booking. At busy times it can be up to 12. The bot must not promise a specific date; it offers to check availability.
- How long until I get the report? Normally 3 to 5 working days after the visit. Larger or older properties can take longer; the surveyor confirms on the day.
- Which survey do I need? A short comparison of the survey levels, with the rule of thumb the firm actually uses (older, altered or unusual properties usually need the more detailed survey).
- Are your surveyors qualified? One factual line naming the professional body membership the firm really holds.
Each entry is 40 to 120 words, stands on its own, and has a first line that looks like a real question. That first line matters, because customers' questions get matched against it.
Writing prices as rules the bot can't bend
Prices cause more chatbot trouble than any other content, because a price sentence written for humans leaves out the conditions humans infer. "Surveys from $450" to a person means "the cheapest survey on the smallest property"; to a bot it can become "a survey costs $450". Write each price as a structured entry with its conditions, exclusions and a date. A filled-in example:
PRICE ENTRY: Condition survey fees
Valid from: 1 April. Review by: 30 September.
Applies to: residential properties only.
Fee bands (property value, fee including tax):
- Up to $250,000: $450
- $250,001 to $500,000: $575
- $500,001 to $750,000: $700
- Over $750,000: quote on request
Included: visit, written report, a 15-minute call with the surveyor afterwards.
Not included: drainage or specialist tests, travel over 30 miles.
What changes the price: properties over 150 years old, extensions without
paperwork, or more than 5 bedrooms may move to the next band.
Bot rules:
- Quote the band only after the customer gives an approximate property value.
- Always say "the surveyor confirms the fee when booking".
- Never quote for commercial property; offer a call instead.The figures above are illustrative; the shape is the point. The "valid from" and "review by" lines stop a stale price living for years, and the bot rules turn a paragraph into behaviour. For trades that price after a visit, the entry changes shape but keeps the same parts. A painter and decorator might write: typical range per room, what drives it (ceiling height, wallpaper removal, condition of the plaster), a firm rule that the bot never gives a fixed price without a visit, and a sentence offering to book one. A locksmith might have one fixed figure (the daytime call-out charge), one conditional figure (the out-of-hours charge and the hours it applies), and a rule that the bot never estimates arrival times, since that depends on where the vans are.
Policies: cancellations, guarantees and what the bot must hand over
Policies need the same treatment with one extra part: a line that says when the bot stops answering and passes the conversation to a person. Customers asking about policies are often halfway into a dispute, and a bot that quotes the refund rule confidently at an angry customer creates more work, not less. A filled-in cancellation entry for the surveying firm:
POLICY ENTRY: Cancelling or moving a survey
Valid from: 1 April.
- Free to cancel or move up to 2 working days before the visit.
- Cancelled with less notice: a $95 fee covers the reserved slot.
- If the sale falls through before the visit, we waive the fee once.
- If the surveyor can't get access because nobody attends,
the full fee applies.
Bot rules:
- Explain the policy plainly if asked in general.
- If the customer says they've been charged, or wants to dispute a fee,
do not argue or explain further: say a member of the team will reply
within one working day and collect their name, email and booking reference.Do the same for guarantees, complaints, payment terms and anything involving safety. Some topics shouldn't have an answer entry at all: legal questions, anything about a specific past job, and structural or safety advice the firm would only give after inspection. For those, write a short entry whose only job is to route. The detail on routing rules is in when a chatbot should hand over to a human.
The instruction block that sits above your content
Most tools have a field called something like instructions, guidance or persona. This is where you set the rules that apply to every answer. Keep it short; the content does the heavy lifting. An example you can adapt:
You answer questions for a residential surveying firm's website visitors.
Answer only from the knowledge provided. If it doesn't cover the question,
say: "I don't have that information, but the team can help" and offer
to take a name and email for a reply within one working day.
Prices: quote only fee bands written in the price entries, always with
their conditions. Never calculate a discount, combine offers, or quote
a price for work not listed.
Dates: never promise a visit date or report date. Give the usual range
and offer to check availability.
Tone: plain, friendly, brief. Three sentences or fewer unless the
customer asks for detail. Use the customer's own words for the property.If your bot also has to refuse things firmly, such as discounts or services you don't offer, the approach in chatbot guardrails that stop the AI promising what you don't offer goes further.
Test with 50 questions your customers really sent
Pull 50 questions from your inbox, contact form and past chats. Include awkward ones: typos, two questions in one message, questions about services you don't offer, and a couple of annoyed customers. Run each through the bot and score it:
| Score | Meaning | What to change |
|---|---|---|
| Right | Correct, complete, right tone | Nothing |
| Right but thin | Correct but missed a condition or next step | Add the missing condition to the entry |
| Wrong source | Answered from the wrong entry | Sharpen the entry's first line; split mixed entries |
| Wrong | Stated something false | Find the conflicting content and remove it |
| Should have handed over | Answered when it should have routed | Add or tighten a handover rule |
An illustrative test from the surveying firm. The question: "how much for a survey on a 4 bed 1930s semi, we're offering about 480k". The bot's first answer:
A condition survey for a property up to $500,000 is $575. Would you like to book?
That's "right but thin". The number is correct, but it missed two conditions from the entry: that a property of that age may need the more detailed survey, and that the surveyor confirms the fee at booking. The fix wasn't in the instruction block; it was moving the age condition higher in the price entry and adding "for properties built before 1950, mention the detailed survey" as a bot rule. The retest answer gave the band, flagged the age point in one sentence and offered a call. On the firm's first run, 31 of 50 were right, 9 thin, 4 wrong source, 3 wrong and 3 should have handed over. All three wrong answers traced back to one old PDF that was still loaded as a source. Two rounds of fixes took it to 44 right and none wrong. A fuller pre-launch plan is in testing a customer chatbot before it goes live.
Questions that trip up well-written entries
Even with clean content, a handful of question shapes cause most of the remaining errors. Test each one deliberately, because they rarely appear in a tidy test list.
- Two questions in one message. "How much is a survey and can you come Friday?" Bots often answer the first half and ignore the second. Add an instruction line: if a message asks more than one thing, answer each in order.
- Comparisons you never wrote down. "Is that cheaper than the other firm?" Your content can't answer it, so the only safe reply is your own price and what's included. A rule against commenting on competitors belongs in the instruction block.
- Haggling. "Can you do it for $400?" Without a rule, some bots reply "I can't confirm discounts, but the team may be flexible", which reads as a yes. Write the exact sentence you want: prices are as listed, and the surveyor can discuss the scope of the report.
- Out-of-scope work that sounds close. A surveying firm asked about "a valuation for a divorce settlement" or a kitchen fitter asked about "just fitting a sink" gets a confident answer stitched from nearby entries unless there's an entry saying what you don't do and who to ask instead.
- Stale context in the question. "Is the spring offer still on?" If an old offer lives anywhere the bot can read, it will find it. This is the audit step again, and the reason deleting beats adding.
Keeping prices and policies current after launch
The bot is only as current as its last sync, and most price errors after launch come from a change made in one place but not another. Four habits prevent them:
- One owner per entry type. Whoever changes prices also updates the price entry the same day. Put it in the price-change checklist next to updating the website and the quote template.
- Know your sync behaviour. If the bot reads your website, find out how often it re-reads it. HubSpot's Customer Agent re-reads its knowledge-base articles as soon as they're edited but other sources only weekly, Intercom's help pages describe scheduled re-syncs for website content, and Chatbase only re-syncs automatically on Standard and above. A price living on a synced web page can therefore be wrong in the bot for days after you change it. After a price change, trigger a manual resync where the tool allows it, then ask the bot the price question to confirm.
- Review dates on everything. The "review by" line in each entry becomes a calendar reminder. When a date passes, check the entry, even if nothing changed.
- Read the unanswered questions weekly. Most tools log questions the bot couldn't answer. Ten minutes a week on that list tells you which entries to add next, and which topics customers ask about that you'd never have guessed.
A realistic slip to watch for: a kitchen fitter raised its design-visit charge but only edited the website page. The bot, on a plan without automatic re-sync, kept quoting the old figure for three weeks, and two customers held the firm to it. A single line in the price-change checklist would have caught it. For deciding whether the bot is earning its keep once it's live, measuring whether your AI chatbot is actually working covers the numbers to track.
How long this takes and what it costs a small firm
For the surveying firm, the effort broke down roughly like this: one hour auditing and deleting conflicting content, four hours writing about 35 entries (fees, policies, process questions, areas served and a dozen handover entries), an hour on the instruction block and settings, and two sessions of testing at about 90 minutes each. Call it nine hours spread over two weeks, most of it by the person who already answers enquiries.
Tool costs depend on how the vendor meters usage. Chatbase's Hobby plan is $40 a month for 700 message credits, and how many credits one reply uses depends on the model you pick, so a busy month with a premium model can outgrow it quickly. Intercom's Fin charges $0.99 per resolved conversation on top of the helpdesk seats. HubSpot's Customer Agent uses 50 credits per resolved conversation, about $0.50 at HubSpot's credit price, and needs a Professional tier. Testing also consumes allowance on metered plans, so run your 50-question test on a trial where possible. The cost that matters more is the ongoing half hour a week keeping entries current; without it, the bot drifts out of date and the accuracy you tested for disappears.
Loading your own content into a chatbot: more questions
Can I just point the chatbot at my website and let it crawl?
You can, and it's a reasonable first draft, but crawled pages carry navigation text, old news posts, footers and marketing copy the bot will treat as fact. Crawl only the pages you'd stand behind, exclude old offers and news, and add a set of short written entries for prices and policies. When a crawled page and a written entry disagree, fix the page, because the bot may pick either.
Should the chatbot quote exact prices at all?
Quote exact prices only where the price really is fixed, such as a set call-out fee or a fixed survey fee band. For anything that depends on a site visit, measurements or materials, have the bot give the published range, name what moves the price, and offer to book the visit. That keeps customers informed without the bot committing you to a figure you'd have to honour.
How many entries does a small business chatbot need?
Fewer than people expect. Most small firms get 80-90% of their chat questions from 25 to 60 topics: prices, availability, areas covered, timescales, cancellations, guarantees, payment and what happens next. Start with the 30 most frequent questions from your inbox, test, and add entries only when the conversation logs show a real gap.
Is my content used to train the vendor's AI model?
It depends on the vendor and plan, so read the data-processing terms rather than the marketing page. Most chatbot platforms store your content to search it and pass relevant passages to a model when answering, which is different from training the model on it. Ask the vendor directly whether your content or customer conversations are used for model training and whether you can opt out.
Further reads
- How Much Does a Website AI Chatbot Cost, and Will It Pay Off? — Check the chatbot will pay for the time you put in.
- How to Stop an AI Chatbot Giving Customers Wrong Answers — Fix the wrong answers that slip through after launch.
- How to Build a Company Knowledge Base AI Can Answer From — Build one knowledge base every AI tool can draw on.
- What to Ask an AI Chatbot Vendor Before You Sign Up — What to ask before you sign, including data terms.
- Who Is Liable When Your AI Chatbot Gets It Wrong? — Who carries the risk when the bot misquotes a price.
- Can You Build an AI Chatbot for Your Business Without Coding? — The no-code tools that take these entries directly.
- What to Do When AI Gets Something Wrong With a Customer — The first hour after an AI error reaches a customer: whether to honour what it said, apology wording for email and phone, and how to stop it happening again.
- 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.
- 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.
- Your First 30 Days of AI Customer Support for an Online Shop — Week by week: collect real questions, let AI draft while people send, automate tracking, delivery and returns, then review 30 conversations.
- How to Take Custom Cake Orders With an AI Enquiry Assistant — The details a custom cake quote needs, pricing rules an assistant can quote from, a ready-to-edit prompt and the test enquiries to run first.
- Can a B&B Use an AI Chatbot and Keep the Personal Touch? — Which guest messages a B&B chatbot should answer, which ones only you should, and how the bot can make your own replies more personal, not less.
- Should an AI Chatbot Answer Allergen Questions for Your Restaurant? — Which allergen questions a restaurant chatbot may answer, which must go to a person, and how to stop it guessing from general knowledge.
- AI for Letting Agents: Handle Tenant Queries Without Extra Staff — Count a month of tenant messages, sort them into three bands, give AI a property pack to answer from, and hard-wire emergencies and disputes to a person.
- Answering Pupil and Parent Enquiries With AI at a Driving School — Build an answer sheet from your last 50 enquiries, put AI on web chat or WhatsApp, and hand test dates, complaints and progress questions to a person.
- AI Chatbots for Nursery Enquiries and Visit Bookings — What a nursery chatbot should answer, what it must hand to the manager, how show-round booking works, and the two mistakes that cost parents' trust.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Intercom Fin help pages on content types and uploaded documents; Chatbase pricing page; Tidio pricing page; HubSpot Knowledge Base on customer agent content sources; HubSpot credits pricing; Shopify Inbox documentation.