Solution · Company knowledge
Company knowledge an assistant can actually rely on.
The quality of an AI workspace depends less on the model than on the material behind it: which documents count, who keeps them current, who is allowed to see them, and what happens when the answer is not there. This is the part of every workspace setup we spend the most care on.
01Source of truth
Decide what counts before anything is uploaded.
| Tier | Typical material | Rule |
|---|---|---|
| Approved reference | Current service terms, procedures, price list, policies, templates | The assistant answers from these first and names the source. |
| Background context | Company overview, service descriptions, tone examples | Used for tone and orientation, never quoted as policy. |
| Kept out | Drafts, superseded versions, personal files, sensitive records | Not uploaded or connected. Excluding them is part of the work. |
In a standard setup this is up to 20 reasonably organized documents — about 100 pages. Choosing the right twenty matters more than collecting everything.
02Ownership
Every source has a named owner and an update routine.
A knowledge set goes stale the first time a price or procedure changes and nobody updates the copy the assistant uses. So each document — or each group of documents — gets one owner on your team, and the handover writes down when they review it and how a new version replaces the old one.
Where two documents disagree, your team decides which one is right. We do not ask the assistant to resolve conflicts in your own policies.
03Uploads and connections
A copy, or a live connection?
Uploaded
A copy of the document as it was
Predictable and easy to review. Changes to the original do not flow through until someone uploads the new version, so the update routine matters.
Connected
Looked up in the source app
Reflects the current version, using each person’s own permissions in that app. Availability depends on the platform, plan, admin settings and the specific app — and the two platforms differ.
Our default
Small, approved, owned
A curated uploaded set for the shared team workspace, plus up to one read-only connection where the plan and source genuinely support it.
04Access
Two sets of permissions, not one.
A file can be restricted in its original folder and still be visible to everyone in an AI project once it is uploaded there: on both ChatGPT and Claude, members of a shared project can see the files added to it. Connected sources behave differently, because the assistant checks each person’s access in the source app.
So before anything is added, we check both: who can see it where it lives today, and who will be able to see it in the AI workspace. If those groups differ, it stays out of the shared project.
05Testing
Test what happens when the answer isn’t there.
Answer is present
“How much notice do we need to reschedule an install?”
ExpectedCorrect answer, with the document named.
Answer is missing
“Do we offer weekend emergency service?” (not in any document)
ExpectedSays it isn’t covered and suggests who to ask. No guess.
Sources conflict
Two documents give different cancellation windows
ExpectedPoints out the conflict instead of choosing one.
Should escalate
“A customer is threatening legal action — what do I tell them?”
ExpectedDirects to the named person; does not draft a position.
A standard setup uses around 20 real examples like these as its acceptance exercise. We report what passed and what did not — not an accuracy percentage.
06What this isn’t
Not a custom knowledge platform.
- Not model training. Organizing reference material and instructions does not retrain the underlying model.
- Not a custom database or search engine. We use the projects and connections your chosen platform already provides.
- Not a document clean-up service. Large archive cleanup and writing missing procedures are scoped separately, if at all.
Next step
Start with one conversation about one piece of work.
An Automation Assessment is a first conversation about fit: what your team does repeatedly, which tool you already use, and whether a workspace setup or a single workflow is the sensible first step. If neither fits, we will say so.