What does it mean to train an AI chat agent on your business FAQs?
Training a chat agent on your FAQs means giving it a written set of questions and answers about your business, plus a few rules, and telling it to answer only from that material. You aren't building a new AI model. You're writing the document it reads before every reply.
That matters because the agent is only as good as what you hand it. If your FAQ says "call for pricing," the agent will say "call for pricing" to a lead at 11pm, and the lead will text the next business on the list. If your FAQ gives a real price range, the agent can quote it, answer the follow-up about financing, and book the appointment.
Most self-service tools fail on content, not technology. In a Gartner survey of 5,728 customers, only 14% of customer service issues were fully resolved in self-service, and even issues customers called "very simple" were fully resolved just 36% of the time. The most common reason was that customers couldn't find relevant content. Your FAQ document is that content.
What should go in your FAQ document?
Your FAQ document should cover the questions leads ask before they buy: price, availability, service area, booking, cancellations, payment and what you don't do. Write it from real messages your team already gets, not from what you wish people asked.
Here is the set I start with for a local service business:
| Section | Example question | What to write |
|---|---|---|
| Pricing | How much does a drain cleaning cost? | A real range, what changes the price, and whether there's a trip fee. |
| Hours and response time | Can someone come out today? | Your actual hours, same-day rules, after-hours fees. |
| Service area | Do you work in Plano? | A list of cities or zip codes, and what happens just outside it. |
| Booking | How do I book? | What the agent can book directly, how long appointments run, your time zone. |
| Cancellations and refunds | What if I need to reschedule? | Notice required, any fee, how refunds work. |
| Payment and financing | Do you take payment plans? | Accepted methods and any financing option, with conditions. |
| What you don't do | Do you fix commercial boilers? | A plain no, and a referral if you have one. |
| Handoff | Can I talk to a person? | When and how a human takes over, and how fast. |
Most owners skip the "what you don't do" section. Without it, the agent tries to be helpful and books jobs you don't take, and your team spends the morning cancelling them.
How do you write FAQ answers an AI agent can use correctly?
Write each answer as one short, complete statement with the specific number, date or condition in it. One question per entry, no internal shorthand, and no answer that depends on another answer to make sense. If a human new hire couldn't act on it, the agent can't either.
- Use numbers. "Most installs run $3,000 to $6,500 depending on size" beats "pricing varies."
- Put conditions next to the answer. "Same-day service is available if you book before 2pm, Monday to Friday."
- State your time zone in every answer about times. In our own calendar data, a wrong time zone on a booking caused a no-show.
- Say what the agent must not promise: discounts, exact arrival times, warranty decisions, anything legal or medical.
- Write the way your customers talk. If they say "AC" and you say "HVAC system," include both.
- Date the document and name one person who owns updates.
How do you set up an AI chat agent on your FAQs, step by step?
Collect real questions, write the answers, add rules for what the agent can't do, connect your calendar, test it against real conversations, then launch on one channel. A small business can do this in a week if one person owns it.
- Pull 50 to 100 real questions from the last 90 days of texts, emails, website chats, Facebook messages and call notes. Group duplicates.
- Write an answer for each group using the rules above. Aim for 30 to 60 entries to start.
- Add a short rules section: answer only from the FAQ, never invent a price or policy, and hand off to a human when the question isn't covered or the person is upset.
- Connect the agent to your real calendar with your booking rules (appointment length, buffer time, how far out people can book) and the right time zone.
- Test it. Paste in 30 real questions from your list, including the awkward ones, and read every answer. Fix the FAQ, not the agent, when an answer is wrong.
- Launch on one channel first, usually website chat or text, and watch every conversation for the first two weeks.
- Log each lead and conversation in your CRM so a human can pick up any thread without asking the customer to repeat themselves.
Booking speed is worth extra attention in step 4. In our own calendar data, people who showed up were mostly booked within 24 hours, while no-shows were mostly booked 2 to 5 days out. If your FAQ and calendar let the agent offer today or tomorrow first, offer it.
What happens if your AI chat agent gives a wrong answer?
You're responsible for it. In Moffatt v. Air Canada (February 2024), a British Columbia tribunal ruled the airline had to honor bereavement fare advice its website chatbot made up, and rejected the argument that the chatbot was responsible for its own words.
The chatbot told a grieving customer he could claim a bereavement discount after travel, which contradicted Air Canada's own policy page. The tribunal ordered the airline to pay him about $812, as MobileSyrup reported. The amount was small. The rule it set for businesses is clear: a customer can rely on what your agent says.
So keep anything with money or liability attached in the FAQ in exact words, tell the agent to quote it rather than paraphrase, and route anything outside the FAQ to a person. A handoff that costs you five minutes is cheaper than honoring a policy you never had.
How do you know the FAQ training is working?
Track four numbers each week: how many conversations the agent finished without a handoff, how many turned into booked appointments, how many answers a human had to correct, and which questions it couldn't answer. The last list is your next round of FAQ entries.
Speed is part of the payoff. The Harvard Business Review lead response study found that companies that tried to contact leads within an hour were nearly seven times as likely to qualify them as companies that waited longer. An agent working from a good FAQ answers in seconds, at 11pm and on Sundays, which is when a lot of local leads come in.
Review the unanswered list every Friday, add five to ten entries, and re-test. For what happens after the first reply, our guide to automating follow-ups covers the texts and reminders that keep a booked lead from going quiet.
KenjiAI's Conversation AI answers website chat and text messages from your own FAQs, books straight into your calendar, and hands the conversation to your team inside one inbox when it should. We set up the FAQ document with you from your real messages, then test it before it talks to a single lead. If you want that done for you, book a call and bring your last 50 customer questions.
Where should you start today?
Open your text messages and email from the last month and copy every question a lead asked before buying into one document. Group them, write the answers with real numbers, and you'll have the first version of your agent's FAQ by the end of the day.