Why my bot gave a wrong reply, and how to fix it

Updated 2026-07-26 · 8 min read

A customer asked something ordinary and your agent said something wrong. Before you change anything, it helps to know where that sentence came from.

Every reply is assembled from your own settings

Each time a message arrives, the system builds a fresh instruction sheet for that one reply, in a fixed order:

1. Voice and language — the language you chose, and it wins. A Banglish agent answers in Banglish even when the customer writes in English.
2. Persona — agent name, shop name, tone, how much detail to give, and your own written instructions.
3. Behaviour rules — including the hard one: if the answer is not explicitly in your catalogue, your FAQs or your instructions, never guess. What it does instead is set by your escalation level, which out of the box hands the question to you.
4. Your catalogue — every saved product, with price, description and negotiation ladder.
5. Your custom scripts, then your FAQs, your sensitive topics, your payment methods.

There is no general knowledge about your shop hiding anywhere else.

The bot is not deciding what to say. It is repeating what your settings say, in your voice — so a wrong reply is nearly always a wrong setting, not a wrong bot.

It invented a product, or quoted a price you never set

The catalogue is built from your saved products and nothing else. There is no draft state and no on/off switch — a product is live the moment you save it, and deleting one removes it from the agent's knowledge completely. So the cause is usually duller than it looks: the item was never saved, or it was deleted, or it was saved without a price. Ask for something that is not there and the agent has nothing to work from, so it either escalates or improvises.

Fix: open Products, add the item, give it a real base price.

Once your catalogue passes about 50 products — or roughly 6,000 tokens of product text — the system stops sending the whole list on every message and sends the slice that matches the customer's wording, with a floor of at least five products and ten FAQs so it can never go blank. It matches on the product's name and its description, so Item-77-blue with an empty description gives a customer typing "নীল shirt" nothing to catch on. Name products the way customers say them, and write one real line of description.

For prices specifically there is a guard, and it is off until you switch it on: Agent setup → Safety → Hallucination Guardrail. Once on, every taka figure in an outgoing reply is checked against the numbers already in your settings — base prices, floors, both negotiation steps, variant prices, your COD charge, and any figure written into a product description, an FAQ answer or your instructions — plus any number the customer themselves said. If a number is not in that set, the entire reply is thrown away and replaced with "Bhai/Apu, ektu wait korun, ami detail check kore confirm kori." Blunt, but a customer waiting two minutes is cheaper than a customer holding you to a price you never offered.

It fumbled a question you have answered a hundred times

FAQs are matched on trigger phrases — the comma-separated list you type when you create one. The answer is only reachable if the customer's message resembles one of them.

Most missed FAQs are missed because the trigger list is written in shop language and customers type in customer language. Open Conversations, read five real messages, and paste that exact wording in:

Bad triggers:  delivery information, shipping policy

Good triggers: delivery charge koto, dhakar baire koto, delivery koto din,
               kobe pabo, ctg te dibe, kotodin lagbe, delivery charge

Include the misspellings. Customers type fast on a phone and the agent is told to read by sound, not spelling, but the FAQ trigger list still has to give it something to catch.

FAQs are loaded in the order you sort them, and at least ten always reach the agent — a thin match on a large set gets topped up from the top of that list. So drag your price, delivery and payment answers to the top.

It answered from the right FAQ but changed your words

That is a setting, not a fault. Agent setup → Persona → FAQ Adherence, at the bottom of that tab, has five levels:

Use Verbatim for anything where the wording is the point — return policy, warranty terms, an account number, a legal line. Leave the rest on Balanced, or the agent starts sounding like a form letter.

Images and links attached to an FAQ still send at every level. Anything in curly braces { } inside an answer is a private instruction to the agent — it obeys it and strips the braces from the reply. That overrides even Verbatim, which is why a Verbatim FAQ can come out shorter than you wrote it.

It does not sound like your shop

Three settings, in the order they matter:

• Language (Persona tab). Banglish, Bengali script, or English. This overrides the customer — it is the shop's voice, not a mirror.
• Tone (Persona tab). Friendly, formal, or street-smart. Street-smart sounds like a market seller; formal sounds like a bank.
• Detailing (its own tab). Minimal answers only what was asked, in one or two lines. Persuasive gives benefit, price, savings, warranty and a soft close. A shop complaining that "the bot says too little" is usually on Minimal.

Then there is Additional Instructions, under Sales playbook on the Persona tab — the anchor for everything else. Empty it and the agent does not become neutral; it falls back to generic guidance derived from your shop profile, and gets blander. Write two or three real lines there.

Test the fix before a customer sees it

Both tools live on Agent setup → Test chat.

Test only in the Response Sandbox and you are testing a config your live Messenger bot has never seen. A banner warns you when the two have drifted apart — save before you trust the result.

Run the same four questions every time: dam koto, delivery charge koto, koto din lagbe, cash on delivery ache? That is what real buyers ask, in that order — the reasoning behind answering all four unprompted is in why your Facebook page gets messages but not orders.

When to stop fixing settings and take the conversation

Open Conversations, pick the thread, and type in the reply box — it is labelled Pauses bot for a reason. Your own message pauses the agent on that conversation for 30 minutes, or until three more customer messages arrive, whichever comes first. Both numbers live in Settings (Resume after (Minutes) and Resume after (Messages)), and a Resume Bot button hands it back early.

The agent also pauses itself. When it rates its own reply low-confidence — or hits anything your escalation level covers, which by default includes any question it cannot answer — it stops, marks the thread Attention, and messages you on Telegram with the customer's actual question. Reply to that Telegram message and your answer goes straight to the customer. If nobody is paired to Telegram, nothing is pushed anywhere and the flag in the dashboard is your only signal — pairing takes two minutes and is covered in how to get order alerts on Telegram.

Take over yourself when it is a complaint about an order already placed, when money has already moved, or when the agent has been wrong twice in the same conversation. A third wrong reply costs more than the ten minutes.

This week

• Filter Conversations by Bot paused and read what made it stop. That list is your FAQ backlog.
• Copy five real customer messages into the trigger phrases of the FAQs they should have matched.
• Set your returns and warranty FAQs to Verbatim, leave the rest on Balanced.
• Turn on the Hallucination Guardrail if you sell at fixed prices.

What the agent handles once it is behaving is on the features page; how many pages and conversations your plan covers is on pricing.