Giving AI Agents the Right Permissions
AI agents are now doing real work inside businesses. Here’s how to decide what they can access, what they can do, and when a person should stay in charge.
AI agents are no longer just chatbots. They now read documents, make recommendations and even take actions inside business systems. That raises an important question for business leaders: how much should these agents be allowed to do?
What Microsoft is saying
In October 2026, Microsoft warned that AI agents working inside companies should be treated as “agentic insiders”. These agents have a lot of access to business data and tools, and it is hard to predict what they will do with it. Microsoft’s advice is simple: an agent should start with no permissions and get only the access it needs for its task.
Not every task is the same
Think about an AI agent helping with purchasing. It might:
. Read supplier quotes, which is low risk.
. Recommend a supplier, which has more impact but still leaves the decision with a person.
. Approve an order, which is high risk because it spends company money.
These tasks carry very different levels of responsibility, so the way the agent is set up should reflect those differences.
Four simple rules
Limit what it can see. Give the agent access only to the information it needs.
Limit what it can do. Set clear rules on which actions it can take.
Keep people in charge of big decisions. Anything with a significant financial or customer impact should need human approval.
Keep a record. Log everything the agent does, so you can check what happened if something goes wrong.
Plan it from day one
Don’t wait until after your AI project starts to set these rules. Build them into the plan from the beginning.
This also makes it easier to show results. When the agent’s job is clearly defined, you can measure the time saved, the errors reduced and the decisions improved.
A leadership responsibility
Deciding what AI agents can and can’t do isn’t just IT’s job. It’s a leadership responsibility shared across operations, finance and technology.
Where would you start?
Which task would you trust an AI agent to handle first, and where would you still want a person to approve?
