When an AI agent gives a wrong answer, the cause is almost never "the AI is bad". It is usually that the information it was given was missing, vague or contradictory. The good news: you can fix this in an afternoon, and every improvement helps every future conversation.
Write for a new employee on their first day
Imagine hiring someone who knows nothing about your shop. What would they need to answer customers on their own? That is exactly what your agent needs: what you sell, who it is for, your prices, how delivery works and what you will not do.
The five things every agent needs
- Products and prices: names, sizes or variants, current prices and whether each item is in stock.
- Delivery: areas you serve, delivery charges, and how long delivery takes.
- Policies: returns, exchanges and payment methods, in plain words.
- Tone: friendly and short, or formal? Bangla, English or both?
- Limits: what the agent must hand over to a human, such as complaints, refunds or custom requests.
Write FAQs the way customers ask them
Customers do not type "What is your return policy?" They type "ferot dewa jabe?" or "size ta chhoto hole?" Write your questions in the same mixed Bangla, Banglish and English your customers use, and add two or three phrasings of each. The agent understands the meaning, but real phrasing makes the answer more precise.
Be specific
"Delivery is fast" teaches nothing. "Inside Dhaka 1 to 2 days for 60 taka, outside Dhaka 3 to 5 days for 120 taka" lets the agent answer exactly.
Keep one source of truth
If your price list is in one place and a post says something different, the agent cannot know which is right. Update the agent's information whenever prices or policies change, and remove outdated answers rather than adding new ones next to them.
Review real conversations every week
Open a handful of recent chats. Look for three signs: a question the agent could not answer, an answer that was true but unhelpful, and a moment where a human should have taken over. Each one becomes either a new FAQ entry or a new rule. After a few weeks, most conversations need no correction at all.
Know where to configure each part
The full walkthrough is in Training Your Agent, and how the agent behaves (returning-customer memory, human escalation, pausing itself when you reply manually) is described in Smart AI Behavior. If you are just starting, our setup guide shows the whole path from sign-up to your first captured lead.
