Your AI chatbot replied to a customer. How do you know what it actually did?

Most businesses see their chatbot from the outside. The customer's message comes in, the bot's reply goes out, and the chat looks like any other. When the customer writes back "your bot told me delivery is free" or "it kept asking for my name", the team has three ways to find out what happened, and none of them is good.

• Reread the chat and guess. The reply is there, but not where it came from. Was it a step the owner wrote, an answer from the FAQ list, or something the model improvised? The wording alone rarely tells you.
• Replay it yourself. Someone opens the bot's test screen and types the same question. If the answer differs, which is common once the customer's earlier messages are missing, you have learned nothing about the real chat.
• Ask the vendor. Logs exist somewhere, but they are technical, they expire, and they are not in front of the person who has to answer the customer now.

The scale makes this worse, not better. Meta reported in May 2022 that one billion people message a business each week on WhatsApp, Messenger and Instagram Direct, and in Malaysia 98.0% of people were online at the end of 2025. A bot that handles even a modest share of that volume produces hundreds of replies a day that nobody reads at the time. The one you need to explain is always one of those.

The record does not need to be technical. It needs to answer the two questions a team member actually asks when a customer complains, and it needs to sit where that person already is: in the conversation.

1. Which steps produced this reply? A chatbot built as a flow moves through steps the owner named: a welcome, a question about the service, a product lookup, a handoff. Showing those names under the reply, in order, tells you at a glance whether the customer was in the booking part or the pricing part, and whether the bot skipped something.
2. Which approved answer did it use? When the reply came from a knowledge base, name the entry. Not the whole list the bot had available, only the ones this reply actually used. That is the difference between "it might have used one of these forty FAQs" and "it answered from Do you deliver outside KL?".
3. When is there nothing to show? A reply that came purely from the flow should say so by showing steps and no sources. Silence about sources is a fact too.
4. Is it there later? The record has to stay with the message, so the person who reads the chat tomorrow sees the same thing as the person who watched it arrive.

Keep the deeper material, such as the full list of candidate answers and the model's reasoning, for the people who tune the bot. Front-line staff need the steps and the source, not a debugging screen.

In Mampu AI, every reply a Flow Agent sends in a live conversation carries a small line under its last message, such as 3 steps · 1 FAQ. Click it and a panel opens with two things: Steps, the flow steps the agent walked for that reply, by the names you gave them in the designer, for example Start → Opening hours → Ask name; and FAQs used, the questions from your AI Knowledge the answer drew on. Only the FAQs the reply actually used are listed, and a reply that answered from the flow alone shows the step count on its own. A step you have since removed from the flow reads Removed step, so an old chat still makes sense after you edit the bot.

The line arrives with the reply for everyone who has the conversation open, and it stays with the message, so the record is the same whether you read the chat now or next month. The agent badge at the top of the conversation counts the steps for the whole chat; hover over it to read the full path.

The same two things appear under every test reply in the designer's Playground, as Nodes visited and FAQs used, for everyone on your team. People who tune the bot can go deeper there, but the steps and the sources are what the front line sees. Alongside, the AI Knowledge Hub shows how many chats used each FAQ in the last 30 days, so the entries nobody needs are easy to find.