Stork Delivery
Give your AI search a permission model
An internal AI search tool should respect the same access boundaries as the documents and records behind it.
The problem: an assistant can make restricted information easier to find
A company assistant can make information easier to find. That is useful until it makes restricted information easier to find too.
Before connecting an internal search tool to shared drives, customer records, or support systems, decide which information each user is allowed to retrieve. The model should receive only the material that the current user is authorized to use for the task.
A practical approach
Carry identity through the request
Knowing that someone can sign in to the assistant is not enough. The application needs to connect that identity to relevant permissions in the source systems.
For example, a project manager may be allowed to read documents for one customer while another manager is assigned to a different account. A search index that mixes all customer documents without retaining access information makes that distinction difficult to enforce.
Decide where permission checks occur and how they fail. If the application cannot establish authorization, it should not guess.
Plan for changes in access
Permissions change when people move teams, projects end, or a folder becomes restricted. Your design needs to account for those changes after documents have been indexed.
Ask how an access change reaches the retrieval layer. Check caches, stored conversations, previews, and exported results as well as the original document. Removing access to the source may not remove information already copied somewhere else.
The right implementation depends on the sensitivity of the information and the source systems. The requirement is simpler: the team should be able to explain how access is granted, checked, and revoked.
Test for separation
Create test users with different access rights and ask the same questions through each account. Include questions that cross project boundaries or ask directly for restricted material.
Check the results, citations, document titles, and error messages. A response can reveal sensitive information even when it does not quote the document itself.
Keep these cases in the evaluation process so future indexing or application changes do not quietly weaken the boundary.
Separate knowledge from authority
Reading an order does not imply permission to refund it. If the assistant will take actions, define those permissions separately and require appropriate approval.
This is part of the surrounding application, not a sentence added to a prompt. Useful AI connected to company data depends on those ordinary software controls working reliably before the answer reaches the user.
The result: search results that respect the current user
Use two test users assigned to different customer accounts and ask each for both accounts' information. Then revoke access to a test document and repeat the checks after the defined propagation window. The release should meet the agreed boundary across answers, titles, citations, and caches. Start with one bounded collection so failures can be traced before more sources are connected.