Overview
A practical AI agent governance guide for security leaders who need to control what agents do, not just whether they are approved.
AI agents now read customer data, call business tools and change records, often without a person in the loop. Most enterprises have responded with policies, approved tool lists and risk committees. That work decides whether an agent is approved. It rarely controls what the agent does next, whose authority it uses, or how quickly it can be switched off. This guide shows where that gap sits, six places it appears in enterprise AI programs, and how to close it with controls identity teams already know how to run. It ends with a five minute test that tells you exactly where to start.

Who This Guide Is For
The guide is written for CISOs, heads of identity and AI governance leads who need AI agents to be as accountable as the people they work alongside.
It draws on what Sennovate’s identity and security operations teams see across enterprise environments, and on current research from Gartner, SailPoint, IBM, CyberArk, IANS and the OWASP GenAI Security Project. Since 2007, Sennovate has secured more than 10 million identities and delivered over 1,000 integrations.
Why AI Agent Governance Breaks Down
AI governance was built to approve tools. Agents need something more: control over what they do after approval.
Agents Act, Not Just Answer
Early enterprise AI generated text. Today’s agents query databases, send emails, open tickets and change systems through tool connections such as the Model Context Protocol (MCP). Gartner predicts that by 2028, 33% of enterprise software applications will include agentic AI, up from less than 1% in 2024, and at least 15% of day to day work decisions will be made autonomously.

Agents Already Do Things Nobody Approved
SailPoint’s research found that 80% of organizations have seen AI agents take unintended actions, including accessing unauthorized systems (39%) and sharing restricted information (33%). Yet only 44% have any governance policy for AI agents.

Nobody Owns the Problem
At companies with more than $10 billion in revenue, AI governance is split across chief AI officers, CIOs and CTOs, and 15% say it has no single owner at all. Only 7% place it with the CISO, according to the IANS and Artico Search 2026 benchmark.
Access Controls Lag Behind
CyberArk counts 82 machine identities for every human, and agents are the fastest growing group. IBM’s 2025 research found that 13% of organizations reported breaches of AI models or applications, and 97% of those lacked proper AI access controls.

The bottom line:
Approval is a moment. Governance is ongoing. Close the gap by giving every agent its own identity, an owner, scoped authority and a tested off switch.
The Six AI Agent Blind Spots
Each blind spot below is a place where an approved agent can quietly become an ungoverned one. Read them with your most powerful agent in mind.
From Access to Authority
Access answers one question: can this agent reach the system? Authority answers four more. Human employees already work this way: an analyst can see every supplier but approve purchases only below a threshold.
- On whose behalf? If the agent acts for a user, it should never hold more access than that user.
- Which actions? Define actions, not systems. “Read tickets and draft replies” is a scope; “service desk access” is not.
- Up to what limit? Value thresholds, record counts, approved recipients.
- With what oversight? Which actions need human approval first, and which are logged for review after.
The Agent Identity Lifecycle
Govern agents through the lifecycle identity teams already run for people.

| Stage | For people | For AI agents |
|---|---|---|
| Joiner | HR record triggers account creation | Agent registered as its own identity, with a named human owner, purpose and expiry date |
| Access | Role based access, request and approval | Task scoped permissions, separate credentials per tool connection, approval for high impact actions |
| Mover | Role change triggers access update | Any new tool, MCP server or data source triggers an access review |
| Review | Periodic certification by manager | Periodic certification by owner, using actual usage data and a delegation map |
| Monitor | Behavioural analytics on user activity | Logging of every tool call and data access, with alerts on actions outside the normal pattern |
| Leaver | Departure triggers revocation | Pilot end date, owner departure or project close triggers revocation, verified across all systems |
The Five Minute Revocation Test
Pick your most powerful AI agent. Give your team one minute per question, and score one point for every answer found in time.
- Find it. Is it in your identity inventory?
- Find the owner. Who is accountable, by name?
- Find its credentials. Own identity, shared account, personal token or API key?
- Find its reach. Which systems, data, tools and MCP connections?
- Switch it off. Can you remove every path with one action and prove it?
Your Score
- 5, Governed: repeat the test on your next ten agents.
- 3 to 4, Partly governed: fix the questions that took longest, usually ownership and reach.
- 0 to 2, Approved, not governed: start with Days 1 to 30 of the plan below.
A 90 Day Plan
- Days 1 to 30, See. Inventory agents, service accounts, API keys and machine trust relationships. Assign a named owner to each. Flag anything without an owner or unused in 90 days.
- Days 31 to 60, Scope. Move agents off personal and shared credentials onto their own identities. Narrow permissions to the task. Add human approval for actions that write, delete, send or pay.
- Days 61 to 90, Prove. Add agents to the next access certification. Turn on action level logging. Run the five minute revocation test on your top ten agents and fix what fails.
What You’ll Gain
Agents You Can Account For
Every agent gets an identity, an owner and a purpose, so nothing acts in your environment without someone answering for it.
Authority That Matches the Task
Scoped permissions and approval for high impact actions keep agents useful without giving them the keys to everything.
A Tested Off Switch
The five minute test turns revocation from a hope into a practiced, provable step.
Governance That Uses What You Have
The same lifecycle controls that protect people extend to agents, so there is no new program to build from scratch.
The result is AI agent governance that holds up after approval, at the moment an agent acts.
