How do I give AI role-based access to company knowledge?
Decide it at the knowledge, not only at the tool. Each person's AI should receive the company context that person is meant to use, plus their own working context, and nothing beyond it. One shared company context layer can still hand different rules to different people, which is what lets you put the useful material in it instead of a thinner, safer copy.
A single knowledge base everyone's AI reads in full is a different product
Permissions answer what a person is allowed to reach. Good AI context also has to answer what the AI should actually use for this task. A person may be allowed to read hundreds of policies while only two of them should guide the decision in front of them.
The appealing version of company AI is one place that holds what the company knows. The problem shows up on the second reading: if every person's AI can use all of it, then the compensation policy, the unannounced roadmap, and the one customer's negotiated exception are all available to everyone who asks a question near them.
The usual response is to leave the sensitive material out. That keeps the base safe and makes it much less useful, because the parts with the most judgment in them are exactly the parts that got removed.
Four positions, all of them reasonable
One shared workspace, open to all staff
Simple, and it works in a small company where everyone genuinely is allowed to see everything.
Folder permissions
Mature, well understood, and enforced by the source system. It answers who may open a file.
A second, AI-safe copy
A curated subset the assistant is allowed to read. Safe. It also becomes a second thing to maintain the moment the first one changes.
Keep the sensitive parts out entirely
The most common answer. The honest cost: the AI ends up working from the least useful version of what the company knows.
Where per-folder permissions are the right answer
Access control on documents is not the weak part of anyone's stack. It is decades old, it works, and it should stay exactly where it is. If the question is whether a person may open a file, the source system already answers it correctly.
The gap is narrower than it first looks. Permissions describe reachability. Doing good work needs something else: which rule should be in front of this task, for this person, right now.
Three gaps between file access and useful context
| What is missing | What happens instead |
|---|---|
| A rule is not a file | The rule that matters is one paragraph inside a long document. Access is granted or denied at the document, so the choice is all of it or none of it. |
| Relevance is not permission | A person may be allowed to read fifty policies. Handing all fifty to their AI makes the one that governs this task harder to apply, not easier. |
| The AI often acts as a service account | When an assistant reads and writes through shared credentials, it is not subject to the person's own permissions, and the record does not say who it was acting for. |
One context layer, with the audience decided per piece
In HexaHQ, creating context and distributing it are separate decisions. A piece of knowledge can stay personal, go to selected people, go to the whole organization, or go to one named collaborator outside it. That decision is attached to the knowledge itself, so it travels with the rule rather than depending on where a file happens to sit.
The same boundary governs what a person's AI may use on their behalf. Their AI gets the organizational context they are meant to have, plus their own personal working context. Because the boundary is real, the good material can go into the shared layer instead of being held back.
Actions follow the person too. Each person connects their own account to a system, so a write happens with their permissions and their name on it, not through a shared identity.
This is about better work, and the boundary is what makes it possible
It is tempting to read all of this as a security feature. The reason it earns its place is the opposite: selective access is what lets the useful, specific, sometimes confidential material be in the shared layer at all. A company that can scope its context does not have to choose between an AI that is safe and an AI that knows anything.
Sales and finance, one context layer, two different answers
-
One shared layer holds the pricing rules
The approval thresholds, the exceptions, and the margin floor all live in HexaHQ, each with its own audience.
-
A representative asks for a quote
Their AI receives the discount rule for their segment. It does not receive the margin floor, which finance keeps to itself.
-
An analyst asks about the same deal
Their AI receives the margin floor and the exception history, because that is what their role is meant to have.
-
Two real actions, each correct
The representative's AI builds the quote and files the approval request in Salesforce. The analyst's AI updates the exception record in Salesforce. Each ran under its own person's account.
Questions people ask next
How do I control what company data AI can access?
At two levels. Connections decide which systems the AI can reach and under whose account, and knowledge carries its own audience, which decides which rules reach a person's AI. The second one is the part most setups are missing.
How do I share company AI context without sharing everything?
Give each piece its own audience rather than sharing a whole workspace. That way one rule can be organization-wide while the reasoning behind it stays with a small group.
Can a contractor's AI get only part of our context?
Yes. A controlled set can be shared with a named external collaborator, without giving them the rest of the organization's context.
Does the AI ever see more than the person does?
It should not, and that is the design. A person's AI works from the context that person is meant to have, and acts through that person's own connected account.
Is this just permissions with extra steps?
Permissions answer who may open a document. This answers which rule should be in front of this task, for this person. You want both, and only one of them exists in most companies today.
Can we require review before something reaches everyone?
Yes. An organization can turn on review, so sharing puts a document in a queue and an administrator decides whether it goes organization-wide.
Related pages
Give the right context to the right people and their AI.
One shared layer, with the audience decided per rule. Free to start.