The internal case for AI agent governance

The short answer: why AI assistants in Microsoft 365 reach so much

AI assistants in Microsoft 365 can reach everything the people behind them can reach, including files that were shared too widely years ago. That exposure is not new. What is new is that it now answers questions.

People build agents in Microsoft 365 that run at their own access
Audit exposure, remediation and a stalled AI program are the costs
Good means an inventory, approval before exposure and continuous review

What is happening

People across the business are building AI agents in Microsoft 365

People across the business are building AI agents, in Microsoft 365, without going through IT. Each one runs at the access of whoever built it.

That access was granted for a person doing a job, not for a standing capability that answers anyone who asks it. Nobody approved it for this purpose, because nobody was asked.

When the person who built an agent changes team or leaves, the agent keeps running. Gartner expects the average Fortune 500 enterprise to be running more than 150,000 agents by 2028, up from fewer than 15 in 2025.

What it costs

Three costs, in the order they usually arrive

Audit exposure

Under the EU AI Act, regulators now ask what AI systems the organization runs, what they reach, and who approved them. Those questions are answerable or they are not.

Remediation

Fixing access across thousands of workspaces after an incident typically costs more than finding it beforehand, and it lands on the team least able to absorb it.

A stalled AI program

When nobody can say what an agent reaches, the safe answer to every rollout request is no, and the investment stalls while the risk stays.

What good looks like

AI agent governance checklist the approver can hold you to

A complete inventory. Every agent and connected app, with a named owner and a plain statement of what data it can reach.
Approval before exposure. One check between an agent being built and an agent being shared, light enough that people still build.
Continuous review. With the evidence produced on demand instead of assembled for each audit.

The size of it

What a breach involving AI costs

USD 6 million

average cost of a breach where AI was involved in the attack, against USD 4.99 million overall

IBM Cost of a Data Breach Report 2026

Specialist Risk Group grew from 119 staff to 450 through acquisition, each one bringing its own accounts and data into the same Microsoft 365 estate. Their infrastructure lead described the problem before they fixed it:

There is no way to view it in one place in a simple contextual way.

Greg Bowles Infrastructure Team Leader, Specialist Risk Group

What we are asking for

Thirty minutes to walk through what is actually running in our tenant and what each agent can reach. Not a purchase decision and not a procurement. If the answer turns out to be that we already know, that is thirty minutes well spent. If not, it is better to find out now. Rencore inventories the digital workplace, the apps and automations teams build, and the AI agents that reach company content.

Download the one-page case (PDF)Talk to us

Frequently asked questions

Why can AI assistants reach so much?

An agent runs at the access of whoever built it, and that access was granted to a person doing a job, not to a standing capability that answers anyone who asks. Files shared too widely years ago were always exposed; now they answer questions.

What does it cost to leave agents ungoverned?

Audit exposure first, because regulators under the EU AI Act ask what AI systems run, what they reach and who approved them. Then remediation after an incident, which typically costs more than finding it beforehand. Then delay to the AI program, because when nobody can say what an agent reaches, the safe answer is no.

What does good AI agent governance look like?

A complete inventory with a named owner per agent, one approval check between building and sharing that is light enough that people still build, and continuous review with evidence produced on demand.

How do I make the case for AI agent governance internally?

Keep it to one page: what is happening (people build agents that run at their own access), what it costs (audit exposure, remediation, a stalled AI program) and what good looks like (an inventory, approval before exposure, continuous review). Then ask for thirty minutes to look at the tenant.

How does Rencore Governance help?

Rencore inventories the digital workplace, the apps and automations teams build and the AI agents that reach company content, shows what each one can reach and enforces policies on a schedule.

Go deeper

Your AI assistants can reach everything your staff can reach