To train an AI chatbot on your business knowledge, give it four things: your catalogue, your frequent questions, your written policies and examples from past chats, then correct its answers as real customers use it. The agent only knows what you give it, so the work is mostly tidying information you already have. This guide shows what to prepare and how to keep it current. For the wider picture, start with our guide to AI WhatsApp agents.
What the agent needs to know
Think of the agent as a new staff member on their first day. They are quick and polite, but they know nothing about your business until someone tells them. Start with the information customers ask about most.
| Area | What to provide | Where it usually lives |
|---|---|---|
| Products or services | Names, descriptions, sizes, options, what is included, what is not | Online store, menu, price list, brochure |
| Frequent questions | The 20 to 50 questions customers ask most, with your preferred answers | Staff memory, old chats, website FAQ |
| Policies | Delivery, collection, returns, cancellations, payment methods | Website footer pages, receipts, staff notes |
| Business details | Locations, opening hours, holiday hours, contact routes | Google Business Profile, website |
| Limits | Topics the agent must hand to staff | Usually nowhere yet, so write them now |
Write the information in plain sentences, the way you would explain it to a new colleague. Avoid internal codes, and spell out what abbreviations such as “COD” or “SKU” mean, because customers will use everyday words.
Do not wait until everything is perfect. A focused set of accurate information beats a large pile of old documents with contradictions.
Turning old chats into FAQs
Your WhatsApp history is the most honest record of what customers really ask. A simple way to mine it:
- Export or scroll through a month of chats. Choose a normal month, not a holiday peak.
- List every distinct question. Group similar wording: “do you deliver to Arabian Ranches” and “delivery to Ranches?” are the same question.
- Count how often each appears. The top 20 usually cover most enquiries.
- Write one clear answer per question. Use the answer your strongest staff member would give, in two or three sentences.
- Note the variations. Keep a few real phrasings next to each question, so the agent recognises them.
Example FAQ entry:
Question: Do you deliver on Fridays?
Answer: Yes, we deliver on Fridays between 10am and 8pm. Orders placed before 6pm on Thursday can be delivered on Friday morning.
Also asked as: “friday delivery?”, “open on friday”, “can I get it Friday”
Remove personal details such as names, phone numbers and addresses before you share chat extracts with anyone, including the agent’s knowledge.
Policies: delivery, returns, opening hours
Policies are where wrong answers cost the most, because a customer may hold you to what the agent said. Write each policy as short, unambiguous statements.
- Delivery: areas covered, cut-off times, delivery days, what happens if nobody is home.
- Collection: where, when and what the customer must bring.
- Returns and exchanges: the time limit, condition of items, what cannot be returned and who approves exceptions.
- Cancellations: until when an order or appointment request can be changed.
- Payment: accepted methods and when payment is due.
- Opening hours: normal hours, public holidays and Ramadan hours, each with dates.
Add an instruction for grey areas, for example: “For any return outside these rules, say a team member will review it and hand the chat over.” This stops the agent improvising a policy you never agreed.
Where a policy has conditions, give an example. “Returns within 7 days, unopened” is clearer to the agent and the customer when followed by “for example, an unopened box bought on Monday can be returned until the following Monday”.
Correcting answers
Even with good preparation, the agent will sometimes answer badly. The question is how quickly you notice and fix it.
A practical correction loop:
- Spot the error. From a weekly chat review, a staff member’s note or a customer complaint.
- Find the cause. Was the information missing, out of date, contradictory or unclear?
- Fix the source. Update the FAQ, policy or product detail rather than writing a one-off patch.
- Write the correction plainly. For example: “Correction: we do not deliver to Hatta. Offer collection from Al Quoz instead.”
- Test it. Ask the question again in a few different ways and check the new answer.
Keep a short log of corrections with dates. It shows which areas cause most trouble and helps the next person who maintains the agent.
Keeping knowledge current
Knowledge goes stale quietly. A product is discontinued, a delivery fee changes or holiday hours are announced, and nobody tells the agent.
Simple habits that prevent this:
- Name an owner. One person is responsible for the agent’s knowledge, with a backup.
- Link updates to existing routines. When you change the website or store, update the agent the same day.
- Diary seasonal changes. Ramadan, Eid, National Day and summer hours each need a dated update and a reminder to revert afterwards.
- Review monthly. Read the full FAQ list once a month and remove anything outdated.
- Check after staff changes. When the person who knew a product line leaves, read that part of the knowledge again and fill any gaps they used to cover from memory.
- Retire old documents. If an old price list is still in the knowledge, the agent may quote it. Remove it when a new one arrives.
How Dub Ai helps
With Dub Ai’s knowledge library, you send your documents, price lists and FAQs, and the assistant answers from your own information. It also learns from the corrections you give it on WhatsApp. If you would rather not prepare everything yourself, our team offers knowledge base setup: we gather your documents, policies and FAQs and organise them for the assistant. The FAQ builder turns the questions customers ask most into clear answers for your website and the assistant. Our help article on how the AI’s knowledge works explains the details.
Frequently asked questions
How much information does the agent need?
Less than most owners expect at the start. Your main products or services, the 20 to 50 most common questions, your key policies and your opening hours are enough to begin. Accuracy matters more than volume. Add more as real chats reveal gaps, and remove anything outdated so the agent never has to choose between two versions of the truth.
How do I fix a wrong answer?
Find why the answer was wrong, then fix the source: the FAQ, policy or product detail it relied on. Send a short, plain correction, for example “We close at 9pm on Fridays, not 10pm.” Then ask the same question in a few ways to check the new answer. Logging each correction helps you see recurring weak spots.