An AI agent that answers customers and does not invent answers

Almost every business has tried a bot by now. Plenty of them switched it off within a fortnight. What separates a bot customers actually talk to from one they spend the whole conversation trying to get past is almost never how clever it is. It is where the answer comes from.

A bot answering out of general knowledge will sound excellent and be confidently wrong. The customer cannot tell the difference, because both answers read the same. They find out when they arrive at the counter, or when the order turns up different from what they were told. At that point it is no longer a technical glitch. It is a broken promise.

Sounding good is not the measure

It is easy to be impressed by a bot that writes clean, well-mannered sentences. That is the easy part. The only measure a customer cares about is whether what they were told is true, and whether they can act on it without checking again somewhere else.

So the first question about any agent is not how well it writes. It is what it is standing on. A good agent is really just a mouth for knowledge the business already has. If that knowledge is not in one place and current, no amount of polish will save the answer.

“I don't know” is a legitimate answer

When the knowledge base has no answer, the system says so. It does not assemble something that sounds real. That reads like a weakness, and in practice it is the thing that makes the agent trustworthy at all. A missing answer costs you one conversation. A confident wrong answer costs you a customer, and sometimes a public review as well.

It also does not end there. An unanswered question is not a failure that vanishes. It is a signal. It points at the exact hole in what the business knows, and at what is worth adding so there is an answer next time. The same question coming from ten different customers is the most useful thing you can learn in a week.

Who decides what enters the knowledge base

This is where most bots break. Either nobody updates them and they quietly go stale, or they swallow everything written in a conversation by themselves, including the thing a rep got wrong on a Tuesday.

Here the default is that a person approves. A new piece of knowledge is proposed, someone on the team reads it and decides. If you want to move faster you can switch on autopilot: above a confidence level you set it goes in by itself, and below that level it keeps waiting for approval. You move the threshold as far as you trust the results. That is control, not caution.

An agent improves from real conversations

Nobody gets an agent right on the first attempt, because you cannot guess in advance how real customers will phrase things. A scan over the conversations that actually took place produces two separate things: knowledge worth adding to the base, and corrections to the behaviour of the agent itself.

If it answered a recurring question badly, that shows. If it never understood a particular way of asking, that shows too. The change is made in the agent editor, which keeps a version history and publishes straight to the bot. You can see what changed and when, and you can go back. This is what the agents look like.

What to check before you put an agent in front of customers

• Where it answers from. With no knowledge base behind it, it will fill the gaps by inventing
• What it does when it has nothing. Whether it is even able to say so
• Who approves new knowledge, and where the switch that changes that lives
• Whether you can read its conversations and correct it, or whether it is a sealed box
• Whether your data is separated from every other business on the same product

That last one looks like a technical question and is not. Each customer's data is completely separate, and the separation is enforced by the system itself rather than by a setting someone could forget. The full detail is on the security page.

And it has to know what it is talking about

An agent sitting on an empty knowledge base will say “I don't know” to everything, which is honest and useless. So it leans on the systems the business already runs: conversations from every channel give it the customer's history, your catalogue and stock let it recommend something that can actually be bought today, your invoicing tells it who is already a paying customer.

Don't see your system? If it has an API, it connects.