In short
Permissions decide what an AI system is allowed to read or do. Approval is a person’s decision about a proposed action at a meaningful point in the process. They work together when the application enforces the limit and the person can see the consequence.
Why it matters
A tool that reads a public guide has a different effect from one that sends a message, changes a record, or spends money. If an assistant can take several steps, a broad approval at the start may not cover what it later proposes. The reviewer needs to see the actual recipient, content, or record before deciding.
How it works
First list the resources available to the workflow. Give it only the access needed for its job, often called least privilege. NIST’s guidance on attribute-based access control explains that access can depend on properties of the user, resource, action, and environment. An AI connection still needs those checks in the application. The model’s request to use a tool is not evidence of the user’s right to the record.
Next distinguish reading, drafting, and acting. Read access to an approved document can help answer a question. Saving a draft may affect another person’s workspace. Sending it externally has a further consequence. Separate tools or settings for these actions so the application can enforce each boundary. NIST’s AI Risk Management Framework also calls for governance and controls fitted to the use case, rather than a generic declaration of trust.
An approval should show enough to make a decision: the source, proposed change, destination, and uncertainty. It should occur immediately before the consequential action, when the details are known. The person must be able to decline, change the proposal, or stop the run. A record of what was approved helps explain a later outcome. Review the failure path too. A timeout, changed source, or new recipient should not quietly reuse approval for an earlier version of the action.
Example
Illustrative example: this scenario is invented.
Maya, 34, asks for a draft reply to an inquiry about a custom gift basket. The assistant can read a public price sheet and save a private draft. It cannot send the message. Maya sees the proposed price, recipient, and wording. She changes the delivery promise because the sheet does not cover that date, then chooses whether to send. The approval applies to the corrected message in front of her, not to any later revision the assistant might create.
She checks the exact message after her edit. If the assistant later changes a sentence or recipient, she would need to see that new version before it goes out.
What it doesn’t mean
A button marked approve isn’t useful when the reviewer sees only a vague task name. A prompt telling the model to ask permission does not enforce authorization. Nor does permission to read one record imply permission to expose it in another context. The smallest useful access and a concrete decision at the consequence make the boundary real.
Try it
On paper, imagine an assistant that plans a public park visit. List three capabilities: read public hours, save a private itinerary, and message participants. Ask any assistant which inputs and approvals each needs. Then change the destination of the message and decide whether the earlier approval should still apply. No account connection is needed.
Keep going
- Tools and function calling
- The human in the loop needs a real job.
- Sandboxes and controlled environments
Sources
- Guide to Attribute Based Access Control (ABAC) Definition and Considerations, NIST, 2014. Checked October 13, 2026.
- AI RMF Core, NIST, 2023. Checked October 13, 2026.