How You Find Out Your AI Is Wrong

A bad AI answer doesn't throw an exception. Nothing turns red. No alert fires.

What happens is quieter. Someone asks a question, gets an answer that's confidently wrong or weirdly incomplete, thinks huh, and goes and asks a colleague instead. They don't file anything. They just stop trusting it a little.

Do that forty times across a department and your deployment is dying, and every dashboard you own still says green.

This is the failure mode nobody demos.

It's worse in a curated deployment

We argued on Monday that curating what the AI can see beats inheriting your whole tenant's permissions. We still think so. But it moves the risk rather than deleting it.

If you point a chat at the wrong sources — or at the right sources that have gone stale, or at four of the five libraries that matter — nothing announces it. Permissions are enforced correctly. The retrieval works. The model answers fluently from what it was given.

It's just answering from the wrong material, and the only person who can tell is the one reading the answer.

So the question for a curated deployment isn't "did we scope it safely." It's "how do we find out when we scoped it badly."

The smallest useful instrument

Every answer in Findable carries a thumbs up and a thumbs down, with an optional comment box. Both are on by default; an administrator can switch either off, and a chat's owner controls whether feedback appears on that chat at all.

That's it. That's the whole interface, and that's rather the point — the moment you ask a busy analyst to fill in a form to report a bad answer, you stop hearing about bad answers.

The field that actually matters

Here's the part that makes it an instrument rather than a satisfaction score.

Each piece of feedback records which data source the answer came from — name, type and id — alongside the conversation and the model settings in play at the time. It's stored against the person's own identity, taken from their authenticated session rather than anything the browser claims.

So you're not collecting opinions. You're collecting a map of which sources produce answers people trust.

Five thumbs-down in a week, all pointing at the same library, is not a model problem and not a user-training problem. It's a curation problem, and it names its own fix: that source is stale, or wrong, or shouldn't be attached to that chat. You'd have had no way to see that from logs, because every one of those queries succeeded.

Who gets to look

Three scopes, which matter more than they sound:

• You see your own feedback.
• A chat's owners and contributors see everything on their chat — because they're the people who chose its sources and can actually fix them.
• Administrators see the tenant, filterable by date, chat, source type, keyword and direction, with positive and negative counts per chat.

The middle one is the useful one. Feedback that only reaches a central admin team becomes a quarterly report nobody acts on. Feedback that reaches the person who configured the chat becomes a change that afternoon.

Users can delete their own feedback, which is the correct answer to the obvious privacy question.

What this isn't

Two honest limits, because overselling this would be funny given the subject.

It is not analyzed for you by default. There's a setting for that and it ships off. Out of the box, this is an instrument that tells you where to look — somebody still has to look. In our experience that's a fifteen-minute job once a week, and it's the highest-value fifteen minutes in running an AI deployment.

And it is not an audit trail. Different thing, different mechanism, and we'd rather say so than let the two blur together.

The question worth asking a vendor

Most enterprise AI tools ship no feedback path at all. You get a chat box and a hope.

So: when this gives someone a wrong answer, how do I find out?

If the answer is "users can contact support," what that means is you'll find out at renewal.

Findable runs inside your own Microsoft tenant: Azure AI Search for retrieval, Azure AI Foundry or Azure OpenAI for models, your identity, your storage, and only the sources you point it at. Request a demo.