The question that kills agent projects: where does the knowledge live?

Enterprise agent purchases have a characteristic failure. The demo is fluent, procurement sees a software line item at a plausible price, everyone signs, and six months later nobody is using it. There is rarely a day you can point at and say that is when it stopped.

The buyer's-guide diagnosis is that nobody asked where the agent's knowledge lives. That sounds like a philosophy question. It is closer to a filesystem question, and the answer decides whether you can operate the thing once the vendor's enthusiasm runs out.

What the question is actually asking

Four smaller ones.

• What can it read? Not "our documentation" — a list of systems with a person's name against each.
• Can somebody on your team open those sources directly, today, without the vendor and without filing an export request?
• When a source changes, how long until the answer changes? Hours is fine. "At the next reindex, which the vendor schedules" is not.
• When an answer is wrong, can you find the sentence that caused it?

The last one is the one nobody tests before signing, and it is what decides whether your first bad answer is a fix or a mystery.

Why the demo never gets near it

Because the demo runs against a corpus somebody curated that morning. Under those conditions all four questions have good answers and none of them is being tested.

The honest version of the test is boring. Take a page that changed last week, ask about it, and time how long the old answer survives.

The two shapes the failure takes

The first is the sealed index. Your content goes in, answers come out, and nothing in between is legible to you. Every wrong answer becomes a support ticket addressed to somebody else. You cannot audit it, cannot correct it at source, and cannot leave without starting over.

The second is the forgotten corpus. The agent reads from a share nobody owns. It answers, with total confidence, out of a handbook that was replaced last year, and it will go on doing that, because nothing in the system tells current apart from merely present.

It is also the more common of the two, and the more embarrassing, because nobody sold it to you.

What a good answer sounds like

Sources enumerated, and each one owned by somebody who knows they own it. Content the team can open in the tool it already lives in. A refresh path measured in hours, with a last-updated time somebody can look at. A trace from any answer back to the passages that produced it. And an exit: if you walked away tomorrow the knowledge would still be yours, because it never stopped being where it was.

None of that is exotic. It is boring infrastructure, which is exactly why it loses to the demo.

Ask it first

On our own work the answer goes into a document before anything gets built. Not out of tidiness. The alternative is finding out in month six, with a system already carrying real traffic and a team that has stopped trusting it. Most of what operating an agent turns out to be is keeping that document true.

The readiness scorecard is this question broken into twelve. No email, and the score is on the next page.