Technical illustration: May the AI Agent Do That? Permissions Instead of Blind Trust

AI security · approx. 4 min

May the AI Agent Do That? Permissions Instead of Blind Trust

What may an AI agent do, which data may it read, and when is human approval essential?

Short answer: Just because an AI agent can technically perform an action does not mean it should be permitted to do so. Businesses should distinguish between reading, writing and executing actions, with human approval particularly for payments, data disclosure and deletion.

Technically possible does not automatically mean authorised

AI agents can increasingly do more than answer questions. They can gather information, prepare tasks and — if given suitable access — make changes in other applications. This is valuable for businesses as long as the technical powers match the agreed assignment.

Consider an assistant that identifies new orders in a mailbox and compiles missing information for sales. It may need read access to selected emails and documents. That does not justify also allowing it to delete orders, authorise payments or modify every customer's data.

An identity of the AI agent's own

Since February 2026, the US National Institute of Standards and Technology's National Cybersecurity Center of Excellence (NCCoE) has addressed software and AI-agent identities and authorisation in a concept paper. Joint international guidance issued in May 2026 on safely adopting agentic AI also recommends limited permissions, monitoring and human oversight.

In practical terms, an agent should preferably have its own technical identity rather than operate through a shared employee account. This makes permissions, audit logs and responsibilities easier to attribute.

A simple permissions matrix provides clarity

Five questions should be answered for every automated process:

  1. Who is responsible? A named person or role authorises the permitted use.
  2. Which data may the agent read? Only what it genuinely needs to complete the task.
  3. Which data may it modify? Write access requires separate scrutiny.
  4. Which tools may it use? Perhaps a quotation template, but not every available interface by default.
  5. When is human approval necessary? Especially before payments, disclosure of data, deletion or other consequential actions.

What does this mean for your business?

Start with one clearly bounded process. Have the AI agent propose results rather than immediately execute binding actions. Then determine which individual actions can be reliably automated.

Permissions, logging and approval should be part of the workflow from the start, rather than added as an afterthought.

Want to introduce AI automation with clear limits on what it can do? I can assist with process analysis and technical implementation.

Discuss a project ↗

Sources and further reading