What if our documentation is out of date?

Then the agent confidently repeats outdated information, which is worse than having no agent. This is the one use case where the quality of your source material decides the outcome entirely. Cleaning up the twenty documents people actually rely on is the first half of the project.

How does it handle conflicting documents?

It should say there is a conflict and cite both, not silently pick one. A single confident wrong answer destroys trust in the whole system faster than any number of "I am not sure" responses.

Does every answer need a source?

Yes. Every answer links to the document it came from, so the asker can verify and so an outdated source gets noticed and fixed. Without citations you cannot tell a correct answer from a plausible one, and neither can your team.

What about documents not everyone should see?

The agent has to respect the same permissions as the underlying systems, per person asking. If it can read a folder the asker cannot, it becomes a way around your access controls. This needs designing at the start rather than added later.

How do we know it is working?

Track which questions get asked, how often the agent says it does not know, and how often people ask a colleague anyway. The list of unanswered questions is the most useful output — it tells you exactly which documentation to write next.

Use case

An AI agent for internal knowledge

The agent answers the questions your team keeps asking each other — which supplier for which part, how the return process runs, what we agreed with that customer. Every answer cites its source. The deciding factor is not the model but whether your documents are current, and usually they are not.

By Gorden WübbeAutomation, agents and search visibilityUpdated 10 August 2026

The interruption nobody measures

In a company of 10 to 100 people, a large amount of time goes on colleagues answering each other. Which supplier do we use for that part. How does the return process run. What did we agree with this customer two years ago. Every one of those answers exists somewhere — in a document, an email, a folder nobody has opened since.

The cost is invisible because it never appears as a task. It appears as the long-serving colleague being interrupted eleven times a day, and as new starters taking three months to become useful.

What the agent does

Answers in the tool your team already uses — the chat system, not a new portal that requires remembering it exists.
Cites the source of every answer, linked, so it can be verified in one click.
Says when it does not know instead of assembling something plausible from adjacent documents.
Flags conflicts where two documents disagree, rather than silently picking the one it found first.

The part that decides everything

Your documents. This is the one use case where the model is not the variable. An agent reading a three-year-old process document will state three-year-old process as current fact, in a confident tone, to someone who has no reason to doubt it. That is materially worse than having no agent, because the colleague who used to be interrupted at least knew what had changed.

The realistic sequence is: identify the twenty documents people actually rely on, bring those up to date, and point the agent at them only. Not at the whole drive. A small, current corpus beats a large, mixed one every time — and the clean-up is frequently the more valuable half of the work, whether or not you then build the agent.

Citations are not optional

Without a source link, nobody can distinguish a correct answer from a plausible one. With it, two useful things happen: the asker can verify anything that matters, and an outdated document gets spotted and fixed by whoever noticed. The agent becomes the mechanism that keeps documentation honest, rather than another place for stale information to hide.

Permissions have to be designed in

The agent must answer according to the permissions of the person asking, not its own. If it can read the folder containing salary information and a junior team member asks the right question, you have built a way around your access controls without anyone deciding to.

This is straightforward if it is handled at the start and awkward to retrofit, which is why it belongs in the first conversation rather than the launch checklist.

What it connects to

Wherever the knowledge lives: document storage, the wiki, shared drives, sometimes the ticket system. Plus the chat tool where people already ask each other, because an internal knowledge tool that lives anywhere else gets used for a fortnight.

How to tell whether it worked

Questions asked per week, and how that changes.
Share of questions the agent could not answer.
How often people ask a colleague after asking the agent.
Time for a new starter to stop needing supervision on routine questions.

The second number is the most useful thing this use case produces, and it is easy to overlook. The list of questions the agent could not answer is a precise, continuously updated list of the documentation you are missing.

What people ask about internal knowledge

What if our documentation is out of date?: Then the agent confidently repeats outdated information, which is worse than having no agent. This is the one use case where the quality of your source material decides the outcome entirely. Cleaning up the twenty documents people actually rely on is the first half of the project.
How does it handle conflicting documents?: It should say there is a conflict and cite both, not silently pick one. A single confident wrong answer destroys trust in the whole system faster than any number of "I am not sure" responses.
Does every answer need a source?: Yes. Every answer links to the document it came from, so the asker can verify and so an outdated source gets noticed and fixed. Without citations you cannot tell a correct answer from a plausible one, and neither can your team.
What about documents not everyone should see?: The agent has to respect the same permissions as the underlying systems, per person asking. If it can read a folder the asker cannot, it becomes a way around your access controls. This needs designing at the start rather than added later.
How do we know it is working?: Track which questions get asked, how often the agent says it does not know, and how often people ask a colleague anyway. The list of unanswered questions is the most useful output — it tells you exactly which documentation to write next.

Which question does your team ask most?

If two or three come to mind immediately, that is the shortlist. Thirty minutes is enough to tell whether the answers exist somewhere written down.

Book a scoping call