How much should an AI agent be allowed to do in internal IT?

November 16, 2026 5 answers
Asked by RAVLOK-3m

I'm looking at using an agent to research tickets and prepare routine IT workflows, with a person approving system changes. Where would you draw the line on access and autonomy, and what would be a good place to start?

Answers

2e!cavrix_831

I'd start with read-only ticket research and have the agent propose one action for a human to approve or correct. Onboarding, offboarding and license changes can still benefit from having the preparation done for you.

Make approval a tool-level requirement, not just a rule in the prompt. A casual message in a conversation shouldn't be enough to authorize a system change. Keep an audit trail and think through rollback before expanding what the agent can do.

RAVLOK-3m:

That separation between a recommendation and an actual tool authorization makes sense. A simple yes in chat could be pretty ambiguous.

Reply to this answer
_p28~sovren.7

Approval isn't the only boundary. Read access can expose sensitive information too, and an agent that can write to a high-impact system is still a risk even if approval is expected.

Define permissions for each integration and give the agent only the access it needs. If you're using MCP, role-based access control is something you need to design into the implementation, not assume the protocol provides for you.

Reply to this answer
_y19+Tarka.6

One useful job is checking dependencies before a change. A group, license or policy can be connected to things the ticket doesn't mention, so have the agent surface those relationships alongside its proposed fix. The human still needs to decide whether the change is safe.

Reply to this answer
_n74~pelvo.5

For a first project, try answering straightforward how-to questions or drafting responses to common tickets. That keeps the workflow familiar for users while giving you a chance to review the quality before letting responses go out automatically.

Some people will appreciate a quick answer. Others won't be happy getting a generic AI response when they expected someone to look at their problem, so leave a clear path to a human.

Rumiq:

Solved tickets could also help build reusable guidance for common issues, though that guidance still needs review before the agent relies on it.

Reply to this answer
_x47!mavren.3

I'd be wary of treating time spent reviewing approvals as no work. The person approving changes still has to catch bad recommendations, and an agent doesn't solve the coverage problem if only one person knows how to run internal IT.

Reply to this answer

Add your answer

Answers are reviewed before publication.

Related Discussions