Your AI Agents Aren’t Governed. They’re Just Polite.
Your AI agent can read your data, call your tools and send emails.
Your safety net? A polite sentence in the prompt.
That is not a control. That is a wish.
I’m a 10-time Microsoft MVP. I have spent 26+ years in enterprise IT, from Tridion and FAST Search to SharePoint and Copilot. I have seen governance projects fail since before “Copilot” was a word. So hear me out: prompts do not protect you.
Here is the five-step path I took.
1. I believed a good concept was enough
I wrote the rules. The concept was perfect. Clean PDF. Clear roles.
Then nobody followed it. Adoption was zero.
Rules on paper stop nothing. They only record what you hoped would happen.
2. Someone showed me where real control sits
Microsoft published an open-source project called the Agent Governance Toolkit. MIT license. Public preview.
The idea is simple. Don’t ask the model to behave. Check every action in normal code, before it runs.
The project’s own docs cite research with a 100% attack success rate against well-known models. The attackers used adaptive methods. So “please follow the rules” is a weak fence.
A prompt injection is a simple trick. Someone hides an instruction in an email or a web page. Your agent reads it and obeys.
If the code says no, the answer is no. Not “unlikely”. No.
3. Three questions changed how I look at agents
Ask them about your agents today:
- Is this action allowed?
- Which agent did it?
- Can you prove what happened?
Most companies cannot answer even one.
This Agent Governance Toolkit gives you three building blocks. A policy engine with plain YAML rules. An identity for each agent, so five agents don’t share one API key. And an audit log that can’t be quietly edited.
A rule can say: delete, drop or truncate? Denied. Send an email? A human from the security team must approve first.
Two lines of code wrap your tool. Every call is then checked and logged.
Think of a bouncer at the door. The agent can ask for anything. The bouncer decides what gets in, and writes it down.
4. The one simple principle: fences beat reviews
You will never review every agent use case. Your IT team has no time. Every IT team I know says the same.
So build the fence once. Low-risk actions pass. High-risk actions need a human. That is risk classification, written as code instead of a slide. It is the same logic as a zone model: personal agents get a wide fence, enterprise agents get a tight one.
The toolkit adds more:
- It scans for agents nobody registered.
- It checks MCP tools for poisoning and hidden instructions.
- It maps controls to the OWASP Agentic Top 10, the NIST AI RMF and the EU AI Act.
That last point helps with evidence. It does not make you compliant on its own.
5. The honest limits (where most posts stop)
It is a developer toolkit. Python, TypeScript, .NET, Go and Rust. Copilot Studio is not on its integration list. So it does not replace admin controls or DLP policies for low-code agents.
It won’t fix oversharing in SharePoint either. That is a job for permissions and Purview.
It runs inside your app process. Microsoft recommends one container per agent for real isolation. And it is a preview, so expect breaking changes.
Use it as a layer, not a savior. That is fine. A layer you understand beats a promise you cannot check.
Your next 45 minutes
Install it with pip install "agent-governance-toolkit[full]". Write one policy. Block one dangerous action. Run agt verify. Show the result to your CISO.
Then ask one question. If your agent went wrong tonight, who would know? And how fast?
Building agents with Semantic Kernel, Agent Framework or Foundry?
PS: Ready to implement proper AI agent governance? Contact me, Ragnar Heil, for a consultation on Agent 365, SharePoint Advanced Management, Microsoft Purview (Information Protection, Data Loss Prevention Policies, DSPM for AI), EasyLife365 Collaboration, Solutions2Share, Rencore Governance, ShareGate Protect, Data&More or Agent 365 deployment strategies tailored to your organization’s needs. Find my calendar here at our HanseVision Governance Landing Page. Pssst… I am also offering an exquisite M365 Copilot & Agents MasterClass. Find out more here.
