What does an AI agent do when it does not know the answer?

It is not one situation, it is three

"It does not know the answer" lumps together three cases that are resolved differently, and confusing them is where the worst errors come from.

The first: the information exists in the business but was never loaded. The agent cannot know it, and the right move is to say so and hand off. The second: the information does not exist yet —a promotion that was never defined, a service that is not offered—. There a denial is correct, because the agent knows it does not exist. The third, and the most delicate: the agent has a source it could check, and does not check it.

Handing off is not failing

There is a temptation to measure an agent by how many conversations it closes on its own. It is the wrong metric, because the way to raise it is to have the agent answer anyway when it should not.

An agent that hands off a share of conversations with the case already summarized —what the customer needs, what they were told, what is missing— saves the team more work than one that closes almost all of them and answers badly in a good number. The first delivers finished work; the second delivers work to undo, plus a conversation that has to be recovered.

The error nobody reviews: denying without checking

We learned this one reading real conversations, and it changed how we verify agents.

When you audit answers, you look for data: an odd price, an impossible date, a number that does not add up. An answer like "no, we do not handle that service" has nothing to check, so it passes the filter and reads as prudence. But if the agent had a way to verify it and did not, it has just asserted something false with complete confidence.

For the customer the damage is the same as an invented fact, with one aggravating factor: they have no reason to doubt it. An odd price invites a follow-up question; a "we do not have that" ends the conversation.

The practical conclusion is that a review which only looks at answers containing numbers is not looking at half the problem. Denials have to be audited too.

What gets defined so this works

Much of this is not a box you tick: it is defined when the agent is set up, and it depends on the industry. In healthcare the rules are hard —it does not diagnose, does not prescribe, does not interpret test results— and in a law firm too: it does not give legal opinions.

What each business defines is where its own limit sits and what happens next.

• Which topics are never answered, even when the agent looks capable of it.
• Who each type of case is handed to, and with which details already collected.
• What is said in the meantime, so the customer is not left hanging.
• Which situations break the normal flow, such as an emergency.