I want a company AI system to help answer a question people ask all the time: “Why are we doing this?”
The answer may be in a document, but it’s often spread across several conversations. Someone proposed a plan, another person raised a concern, and the team changed its mind. A search result containing the original plan may leave out the part that matters.
Keep the reason and the conditions
Suppose a team delays a launch because an integration isn’t ready. A month later, the integration is finished. If the AI retrieves only the message about delaying the launch, it could give someone the wrong impression of the current plan.
I’d want it to show the original decision, the reason for it and the later update. Who agreed to the change? Is anything else still blocking the launch?
I want to store the decision and its source references, then record later changes. That lets the system explain both the current plan and how the team arrived at it.
Check who can see the answer and its sources
Permissions get more complicated when someone shares an answer. A document might be available to the person asking the question but restricted from the team channel where they want to post it.
Even a citation can reveal something private, such as a document title or mailbox identity. Checking the answer’s wording won’t catch that.
In my company-brain implementation, I put checks before information entered the model’s context and before an output was delivered to its audience. Those are separate checks because the audience can change.
Let people check and correct the answer
If an answer includes the source, I can read it and decide whether I agree. I might notice that the document is outdated, that a later conversation changed the plan or that the AI misunderstood what someone meant.
I want a correction to be available the next time someone asks. Otherwise, people have to explain the same thing repeatedly while the system keeps returning the old answer.
That also affects how I’d test it. I’d ask a question, change a relevant decision, then ask again. Does it use the update? Can it still explain the earlier decision when someone asks about the history?
Start with questions people already ask
I’d begin with a few recurring questions: why a project was delayed, who approved a change, or what a team agreed to deliver. For each one, I’d check whether the answer uses the right sources, respects permissions and distinguishes the current decision from an old one.
Then I’d watch someone use it. Do they spend less time finding and checking the answer? Can they make the decision they came to make?
I would measure the time someone spends finding and checking an answer, then observe whether they can use it to make the decision.
The company-brain implementation · Permissions when an answer is shared · How I handle state and verification