LAB
DevelopmentLAB

AI Answers from Company Documents: Scope Sources, Permissions and Abstention

Published:

6 min read

Connecting a model to a folder and asking it to cite sources is not enough. The system needs to know which documents are approved, which users may access them, when they changed or were withdrawn, and what to do when evidence is insufficient. An acceptable knowledge assistant answers within source and permission boundaries, abstaining or escalating when documents conflict or lack a clear basis for an answer.

This article assumes the task and initial sources have been selected. It addresses implementation and acceptance controls for retrieval, not broad AI project feasibility. A file citation does not prove an answer correct, and it does not authorize the assistant to approve a policy exception or decide an employee's or customer's entitlement.

Turn a folder into an approved source register

For each collection, define owner, purpose, audience, status, effective date and update process. Distinguish approved policy from a draft, explanatory correspondence and a superseded version. A document retained for archival reasons is not necessarily an appropriate source for a question about current policy.

Source field Decision it supports
Owner and approval status Who resolves conflict and authorizes use?
Scope and audience Which team, branch or situation does it cover?
Effective date and version Is the answer using the appropriate policy period?
Access permissions Who may retrieve and cite the content?
Update and deletion behavior How do changes reach indexes and cached outputs?

If usage rights or permission integration cannot be established, return that constraint to AI Readiness Assessment. Uploading everything to a new tool and planning to fix access later is not a substitute.

Check permission before content reaches the model

Do not retrieve an unauthorized document and ask the model not to reveal it. The access design should prevent that content entering the context used for an unauthorized user. Include snippets, titles, citations and cached information, not only full documents. A service account's ability to read a collection does not grant every assistant user the same rights.

OWASP recommends denying access by default and validating permission on each request. Apply that principle to retrieval rather than treating authentication alone as document authorization. OWASP authorization guidance.

API Integration & Development can specify how user identity, access rules and denial outcomes move between systems. Test safe failure when permission cannot be checked; do not automatically fall back to a broader identity simply to keep answers flowing.

Hypothetical example: a general travel policy and a restricted addendum

Imagine an internal assistant answering questions about travel-request procedures. Its sources include an approved general policy, an addendum restricted to a project team, and a draft policy not yet effective. This is a fictional knowledge-control example, not an interpretation of actual employment rules.

An employee outside the project asks which documents to submit. The answer should rely only on the relevant general policy. It must not reveal the restricted addendum's title or hint at a confidential exception. An authorized project member asks the same question and can receive information from permitted sources with their scope explained. The draft remains excluded from authoritative answers until its owner changes its status.

A later test revokes a test employee's access after they have viewed a previous answer. Check that restricted content is not resurfaced through conversation memory or a cached result. In another case, two apparently approved versions conflict. The assistant should not decide entitlement: it should identify that the policy owner needs to resolve the conflict. Linguistic similarity to a question does not give a document authority.

Make citations useful rather than decorative

An authorized user should be able to open the source and identify its version and the relevant passage. Test whether the cited material actually supports the answer, and whether the link resolves to the correct file rather than a changed version without explanation. Displaying a filename is not enough to establish support.

Source text may itself contain instructions or untrusted content. Treat it as material for the permitted question, not authority to change system permissions, operate tools or transmit information. The assistant remains bounded to reading and explaining; decisions and exceptions return to an authorized person.

Define abstention and freshness as normal states

Specify behavior when sources are missing, conflicting, inaccessible, stale or outside the requested scope. “There is insufficient evidence” can be the correct response, accompanied by a permitted next step. General model knowledge should not silently fill a gap in company-specific policy evidence.

Agree with the source owner how quickly changes must propagate and what happens to deleted files and stored snippets. If the connection to the source fails, decide whether clearly labeled cached information is allowed or whether those answers stop. Neither permanent freshness nor complete elimination of hallucination should be promised as a technical guarantee.

The register also needs someone responsible for resolving ownership gaps. If no one can confirm which version is authoritative, better retrieval alone cannot settle the policy question. Mark that collection unavailable for authoritative answers until its owner resolves it.

Acceptance cases worth demonstrating

These controls can be part of AI & Machine Learning Engineering, supported by defined Software Quality Assurance checks. Use authorized or synthetic sources and accounts rather than demonstrating leakage of real sensitive data to a broad audience.

  • An allowed question receives an answer supported by an approved version.
  • The same question from an unauthorized identity reveals neither restricted content nor a restricted title.
  • Revocation affects retrieval and caching within the agreed boundary.
  • Missing or conflicting sources lead to clear abstention and escalation.
  • Test updates and deletions propagate without misleading old answers remaining available.
  • Instructions inside a document cannot expand permissions or trigger external action.

Record the application version, source state and test identities used. Passing one role's examples does not establish all access combinations; disclose the tested matrix and any exclusions.

Frequently asked questions

Do citations prevent wrong answers?

No. Test claim support, retrieval limits and abstention, with appropriate domain review. A citation is evidence to inspect, not a correctness guarantee.

Can the integration use one service account?

Possibly, but that does not solve end-user authorization. The design still needs to enforce each user's limits before exposing content.

Should the entire archive be included initially?

Start with a bounded collection with clear ownership and status. A conflicting archive may require temporal rules and approval before use.

What should we provide for scoping?

Bring a redacted source inventory, owners, an access matrix, and questions that should be answered or refused. Do not share sensitive documents before agreeing purpose, environment and permission.

Reference

OWASP Authorization Cheat Sheet, reviewed 26 October 2026. The example and acceptance design are original guidance, not a guarantee against every failure.

DevelopmentLAB

Need help with this?

Explore LAB's AI & ML Engineering service.