Futur Labs
← BlogOctober 28, 2026Updated November 10, 2026

AI Data Permissions: A Buyer’s Access-Control Checklist

Scope AI data access with a permission matrix, separate user and app identities, revocation tests, derived-data controls and supplier evidence questions.

One open file drawer beside two locked drawers, with a key resting on the stone worktop.
Date
October 28, 2026
Category
AI Implementation
Reading
8 min
Bryce Choquer
Bryce Choquer
Founder, Futur Labs

AI data permissions define which records a workflow can read, what it can change and who may see its outputs. Before connecting an assistant to company files, specify the identity used for each action and test denied access as carefully as successful retrieval. A connected repository is not permission for every user to see its contents.

This guide provides a technical buying checklist for internal workflows. It does not establish that a product meets your privacy, contractual or regulatory obligations. Those decisions depend on the actual data, deployment, agreements and applicable requirements.

Separate reading from an approved writeback

For an integration that changes business records, define the action boundary as well as the documents it can read. This proposed sequence helps scope that boundary.

  1. Authorize one task, its identity, source records and permitted operations.
  2. Read the allowed records with stable IDs and source versions.
  3. Prepare the proposed field changes and show the before/after values.
  4. Have an authorized owner approve that specific change; recheck access and source freshness before executing it.
  5. Apply the permitted operation through controlled application code and record the provider response.
  6. Verify the resulting state where the API supports it. Keep ambiguous writes and failed confirmations visible for reconciliation.

Denied access stops the action. Missing IDs, stale versions and conflicting records return to the relevant owner. Retry rules depend on the provider; a timeout does not prove that a write failed or that repeating it is safe. Keep access approval, proposed content and execution results separately inspectable. For document-only workflows, use the report-generation guide.

What should a permission brief contain?

Start with a bounded task: for example, drafting a project update from documents in one approved project workspace. The assistant may prepare a draft; a project owner controls release. This is a proposed workflow, not a description of a client deployment.

For each action, name the record owner, eligible users, application identity, allowed operation and destination. Separate document access from the right to disclose a summary. Permission to read internal notes does not automatically authorize sending them to a client.

Boundary Decision to record Evidence to request
Source Which project, folder and record types are in scope? Approved inventory with exclusions
Identity Does this step act for a user or as an application? Identity and permission configuration for each connector
Operation May it read, create a draft, update or delete? Operation-specific allowed and denied tests
Retrieval How is the requesting user's access checked? Test demonstrating excluded records do not enter model context
Output Who can view the draft and its evidence links? Audience restrictions and export behavior
Copies Where are excerpts, indexes, logs and backups stored? Data-flow and retention inventory
Revocation What happens when access is removed? Demonstrated behavior and documented refresh limits

Use identifiers rather than vague instructions such as “only approved files.” An inventory needs a maintained rule for additions, moves, renamed folders and archived projects. Otherwise a correct initial setup can drift as the repository changes.

How do user access and application access differ?

For Microsoft Graph, delegated access acts for a signed-in user; effective access depends on both the granted permissions and the user's access. App-only access uses the application's identity without a signed-in user. Do not assume the second model inherits the current reader's restrictions. See Microsoft's permissions overview.

Ask the supplier to show the mode used by every connector and background job. A user-facing chat interface can still depend on a background indexer with broader privileges. “Users sign in” does not answer how indexing, caching or scheduled work is authorized.

Where supported, Microsoft Graph's Selected permissions can narrow access to particular SharePoint or OneDrive resources. Consent alone is insufficient: a resource assignment and an appropriate token are also required. Granular assignments can affect permission inheritance, so the administrator must assess that consequence. These are Microsoft-specific controls, not a universal connector feature. See Selected permissions documentation.

Have the administrator verify the exact resource, operation and identity combination. A screenshot of a consent screen is insufficient evidence that the application can access only the intended records.

How should permissions survive retrieval and summarization?

Treat authorization as an application responsibility. Do not retrieve a broad set of confidential material and ask the language model to hide unauthorized passages. The proposed design should exclude ineligible material before it enters the model context and restrict every later copy.

One useful review sequence is:

  1. Authenticate the requester and resolve the relevant project membership.
  2. Apply the current eligibility rule while selecting source records.
  3. Retrieve only permitted material, preserving source identity and revision.
  4. Build a draft whose audience is no broader than the allowed output audience.
  5. Recheck authorization when opening evidence, sharing or exporting the draft.
  6. Record access decisions with sufficient diagnostic context and without unnecessary document content.

This sequence is a design recommendation. It needs implementation tests; naming it in a proposal does not prove it exists. Source access and output access can also differ. An internal reviewer might see a confidential note that must be excluded from a client-facing report even after the reviewer approves its factual accuracy.

For the rest of the report pipeline, use the AI report generation guide. That article owns fact assembly and review; this checklist owns the access boundary around those steps.

What happens to access after data is copied?

A source access-control list (ACL) does not automatically control an exported file, cached excerpt, search index or application log. Ask the supplier to list derived stores, their administrators, retention behavior and removal process. Include support tooling and diagnostic exports.

Specify separate behavior for immediate access denial and eventual copy removal. If membership changes, the workflow should deny newly unauthorized retrieval and output access according to its documented authorization design. Removing cached or indexed material can involve separate processes and delays. Agree a maximum acceptable delay for each access path before the pilot. Measure revocation from the recorded membership change through retrieval, existing draft access and export; test active sessions as well as a fresh sign-in. If the supplier cannot meet the required deadline, restrict the pilot or defer that source. Require separate evidence for copy removal; do not promise instantaneous deletion across all systems.

Avoid recording sensitive full-text prompts just because a dashboard makes them convenient to inspect. Decide what investigators need, who may inspect it and how long it remains available. Minimized logging still needs enough context to locate a failed authorization decision.

Which tests should gate a pilot?

Use approved synthetic fixtures for the first tests. Give two project groups different records, including similar filenames and an explicitly restricted item. Record the expected behavior before running the workflow.

Test Required observable behavior
User requests another project's unique phrase Neither restricted passage nor derived answer is disclosed
User has no eligible source for a question Workflow reports unavailable evidence instead of inventing an answer
Indexer has broader access than requester Requester still receives only authorized material
Membership is revoked Retrieval, draft access and export deny access within the agreed maximum delay, including active sessions; copied stores follow their separately documented removal process
Evidence URL is opened by another user Source or application enforces the new reader's access
Draft is shared outside its permitted audience Share or export is denied or routed through an explicit authorized release process
Connector loses access mid-run Affected sections stop or show an exception; previous cached access is not silently treated as current permission

A negative test must inspect the context and output paths as well as the visible response. A polite refusal is not enough if restricted text was already sent to a model or exposed in logs. Have the implementation team provide appropriately restricted evidence of the decision path.

What should buyers ask before signing?

Ask who owns the connector registration, grants and revocations; which permissions are essential; which features stop working with narrower access; and how administrators can remove access without waiting for a supplier. Request a documented exception process when the desired source does not support the required boundary.

Also ask what the provider receives, where each copy is processed or stored, how retention is configured and which agreements govern its use. Verify those facts for the selected product, tier and configuration. This guide does not infer data residency, model-training behavior or compliance from an API name.

AI integration is the relevant service discussion for permissioned connections. Bring the permission matrix and denied-access scenarios so the team can assess feasibility before introducing live confidential material.

Common questions

  • No. Sign-in establishes identity. The connector and application still need to enforce which records that identity may read and what outputs it may receive. Test both successful sign-in and denied document access.

  • It can be designed that way, but the account's access needs an explicit boundary. If it can read more than an individual user, the application must enforce the user's narrower eligibility before retrieval and disclosure. Background work needs its own authorized purpose and audience.

  • Do not assume it does. Ask about indexes, excerpts, generated drafts, logs and backups individually. The supplier should document removal behavior, retained copies and any delays for the selected configuration.

  • Read-only access reduces write capability but can still expose confidential information through answers, logs or exports. A pilot also needs source limits, output restrictions and tests of denied access. Add write permissions only for a defined action with separate acceptance tests.

Subscribe

Field notes on AI agents, ERP, and shipping software — straight to your inbox. No spam, unsubscribe anytime.

The latest in AI, in your inbox.