What permissions should an AI agent have?

An agent that drafts support replies needs access to the assigned ticket and permission to save a draft. Here’s how to limit access and test it before expanding it.

6 November 2026

Suppose you want an AI agent to draft replies to support tickets. It needs to read the assigned ticket and somewhere to save its draft. Give it the same access as a support agent and it can also send messages, open other tickets or change their status. For the first release, I’d let it read the assigned ticket and save a draft.

Decide what the agent can do before you connect it

Describe the task in one sentence. Here, it’s “draft a reply to the assigned ticket”. List the information the agent needs and the one change it should be able to make. Give it a dedicated account in the help desk system with permission for those operations. Check that the service enforces which ticket the agent can access.

This is how OWASP applies the principle of least privilege to AI agents. If the agent tries to open another ticket, the help desk should reject the request. The prompt can guide the agent, but the permission check has to work even when the model ignores the instruction. OWASP guidance on tool security and least privilege

Let the person reviewing the reply send it

I’d keep sending in the team’s existing workflow. Whoever opens the draft can see the recipient, the message and any attachments. They can edit the reply and send it from their own account. If you later need to automate sending, ask someone to approve the exact message that will go out. Ask for approval again if the recipient or message changes. OWASP guidance on human approval and action integrity

Test it with three requests

Give the agent a test ticket and check that it saves an appropriate draft reply.
Ask it to read a ticket assigned to another account and check that the service denies access.
Put an instruction in a test ticket telling the agent to send the reply. Check that the message stays in drafts.

Before expanding permissions, look at where drafting already saves the team time. If you find a repetitive step worth automating, add that step on its own and test it separately.

For more practical examples of using AI at work, subscribe to AI at Work.

All articles →