How to teach an AI agent about your business (and test it)
Generic AI answers come from missing facts. Write down prices, policies, FAQs and tone rules, test with 20 real questions, then fix and keep it current.
To teach an AI agent about your business, write down the facts it needs in short, specific notes (your price list, policies, the 20 questions customers ask most, tone rules and a list of things it must never say), put them where the agent looks things up, and then test it with 20 real customer questions before it talks to anyone. Generic answers happen because the AI has nothing of yours to ground its reply in, so it fills the gap with general knowledge. Fix wrong answers by fixing the note behind them, and keep the notes current whenever a price or policy changes.
This is the practical process, from the first page you write to the monthly check that keeps answers right.
Why does my AI chatbot give generic answers?
Because it's answering from what it learned about the world, not from your business, so it guesses whenever a question is about you.
A language model knows a lot in general and nothing about your delivery charges. When you ground it, meaning it retrieves your own documents and answers from them, made-up answers drop sharply; researchers at Facebook AI Research showed that retrieval augmentation "significantly reduces" hallucination in conversation.
Grounding helps a lot, but it isn't perfect. Stanford researchers found that general chatbots got legal research questions wrong 58% to 82% of the time, while legal tools grounded in real case law still hallucinated on more than 17% of queries, and one on more than 34%. Two lessons for a small business: give the agent your facts, and test it before you trust it.
Getting this wrong has a price. In 2024 a Canadian tribunal ordered Air Canada to pay a customer C$650.88 in damages, plus interest and fees, after its website chatbot described a refund policy the airline didn't have. The tribunal rejected the argument that the chatbot was responsible for its own words.
What should I write down first?
Start with the 7 things customers ask about or the agent could get wrong in a way that costs you money.
| What | Examples | Why it matters |
|---|---|---|
| Price list | Every product or service, price, GST included or extra, minimum order | The most asked question, and the costliest to get wrong |
| Policies | Delivery areas and times, returns, refunds, cancellations, warranty | Wrong policy answers become promises |
| Top 20 FAQs | Copied from real emails and chats, with your best answer | Covers most of what customers ask |
| Business basics | Hours, locations, service areas, how to reach a person | Stops "I'm not sure" on easy questions |
| Tone rules | "Short, warm, no emojis, Hinglish if the customer writes Hinglish" | Sounds like you, not a template |
| Never-say rules | "Never offer discounts. Never promise a delivery date for custom orders." | Hard limits the agent must not cross |
| Hand-over rules | "Complaints, bulk orders over ₹1 lakh and anything legal go to a person" | The agent knows when to stop |
Write each note the way you'd brief a sharp new hire:
- One topic per note, with a title phrased as the question: "How much is delivery to Pune?"
- Answer first, in the first sentence. Details after.
- Numbers, not adjectives. "Free delivery above ₹999, otherwise ₹79" beats "affordable delivery".
- Dates on anything that changes. "Prices from 1 November 2026."
- No secrets. Passwords, API keys and card details never go in notes; they belong in the Security Vault.
Where does each piece go in KutrimBots?
Your agents look things up in Super Brain; your website chat answers from its own help center and business background, so customer-facing answers go in both.
| Knowledge | Where it goes | Who uses it |
|---|---|---|
| Prices, products, processes, FAQs, decisions | Super Brain, Company space (folders such as "Pricing" or "Policies") | Every agent, when it looks something up |
| Values that change (price, launch date, owner) | Super Brain facts, where the newest value wins and the old one is kept as history | Every agent |
| Lead and client details | Super Brain, Customers space, linked to the CRM lead | Agents working with that customer |
| What you sell, prices, delivery, policies, for website visitors | Chat Widget business background ("about") | The website chat agent |
| Tone, never-say and hand-over rules for the website | Chat Widget guidelines | The website chat agent |
| Answers to the top visitor questions | Chat Widget help center articles | The website chat agent and visitors on the Help tab |
The split is deliberate. The public website copy of your agent can only see the help center and business background, never your internal notes or other customers' data. Ask an agent to "fill in the Chat Widget background from what you know about our business" and it drafts the text from Super Brain for you to check. How Super Brain organises and searches memory is covered in Super Brain: shared memory for your AI team; installing the website chat is in how to add an AI chatbot to your website.
How do I test it with 20 real customer questions?
Ask 20 questions your customers really asked, mark each answer, and don't go live until at least 18 are right.
- Collect 20 real questions from your email, website chats and call notes. Use their words, typos and all.
- Mix them: 8 about prices and products, 5 about policies, 4 awkward edge cases ("Can you deliver tomorrow to Lonavala?") and 3 that should go to a person (a complaint, a bulk order, a legal question).
- Ask them where customers will, as a visitor in your website chat and in an email to the agent's address.
- Mark each answer:
| Mark | Meaning | Usual cause |
|---|---|---|
| Right | Correct and complete | |
| Incomplete | Correct but missing a detail, such as GST | The note leaves it out |
| Wrong | A real fact, but the wrong one | Old and new versions of a note disagree |
| Made up | A fact you never gave it | No note covers the question |
| Should have handed over | Answered when it should have passed to you | Missing hand-over rule |
- Score it. 18 or more right out of 20 is ready for customers, with you reading chats for the first week. Fewer than 15 means a gap in your notes, not a broken agent.
Example scenario: a home bakery in Bengaluru
The numbers are assumptions for illustration, not customer data.
The bakery sells 60 cakes and bakes through its website, where the chat agent answers visitors. Before anything was written down, the agent got 9 of 20 test questions right: it promised "delivery in 3–5 days" (the bakery delivers the same day within 8 km) and offered a 10% discount that doesn't exist.
The owner spent 2 hours writing 12 notes: a price list by category, delivery charges (free above ₹999, ₹79 below), a cancellation policy (24 hours' notice for custom cakes), 20 FAQs, tone rules and 4 never-say rules, then copied the customer-facing parts into the Chat Widget background and 10 help articles. On the second test the agent got 18 of 20 right. The two misses were a question about eggless options (now a help article) and a 200-piece wedding order it should have handed over (now a hand-over rule).
How do I fix a wrong answer?
Find the note that caused it and fix the note; correcting the agent in one chat doesn't fix the next chat.
- Made up: there's no note. Write one.
- Wrong: two notes disagree. Save the new value as a fact so the newest wins, and ask the agent to forget the old value. In Super Brain, forgetting is precise: it forgets the old delivery charge, not the whole delivery note, and you can undo it in the app for 30 days.
- Incomplete: add the missing detail to the note, in the first two sentences.
- Wrong tone or a broken rule: add it to the tone or never-say rules, and to the Chat Widget guidelines for the website.
- Right in agent chat, wrong on the website: the help center or background is missing it. Add a help article.
Tell the agent which note an answer should have come from; in KutrimBots, agents say which note an answer came from, which makes the cause easy to find.
How do I keep the agent's knowledge current?
Update the note the same day something changes, and review the top questions once a month.
- Say "remember". "Remember: from 1 December, delivery to Pune is ₹99." The agent saves it as a fact, so the old value becomes history.
- Let agents save as they work. When a lasting decision comes up in a chat, they save it to team memory. If you prefer editing files, notes edited as Markdown in File Manager are read back into Super Brain every 5 minutes.
- Check what visitors ask. The Chat Widget summary shows what visitors asked and which chats needed a person; turn the top gaps into help articles.
- Run the 20-question test monthly, and after any price change, new product or policy change.
- Let the tidy-up run. Super Brain places new notes, marks newer versions and keeps a short brief agents start from, every 10 minutes in the background.
Notes are information, not instructions. If a note ever contains something like "ignore your rules", agents don't follow it.
What can't good notes fix?
They make answers accurate; they don't give the agent judgement, authority or certainty.
- It can still be wrong. Grounded systems still make mistakes, which is why the test and the weekly read-through matter.
- Judgement calls. Exceptions, discounts and angry customers need a hand-over rule, not a better note.
- Advice you're liable for. Legal, medical, tax and financial advice shouldn't come from a chatbot; tell it to hand those to a person.
- Facts you don't have. If stock levels change hourly and live only in your head, the agent can't know them.
- Secrets. Never put passwords or card details in notes; use the Security Vault.
To try the process on your own business, start the free 7-day trial; it includes Super Brain, the Chat Widget and up to 3 agents. For the difference between a scripted chatbot and an agent that looks things up, see AI agents vs chatbots.
Frequently asked questions
How do I teach an AI agent about my company?
Write short notes on your prices, policies, top 20 FAQs, tone rules and never-say rules, save them where the agent looks things up (Super Brain in KutrimBots, plus the Chat Widget help center for website visitors), and test with 20 real customer questions.
Can I give the agent my price list as a PDF or document?
Yes. Share the file or paste it into the chat and ask the agent to turn it into Super Brain notes, one topic per note. Then check the notes, because a long document often mixes old and new prices.
Why does my AI still make things up after I added documents?
Usually because no note answers that exact question, two notes disagree, or the answer is buried deep in a long document. Test with real questions, then add or fix the note behind each wrong answer.
Does the website chatbot use the same knowledge as my agents?
Not directly. In KutrimBots, the website chat agent sees only the Chat Widget help center and business background, for safety. Ask an agent to fill those from Super Brain, and update both when something changes.
How often should I update my AI's knowledge?
The same day a price, policy or product changes, plus a monthly 20-question test. Saving changes as facts means the newest value wins and the old one is kept as history.