Score Your AI Agent Governance in 28 Checks
The average Fortune 500 company is expected to run more than 150,000 AI agents by 2028. Only 13% of organizations say they have the governance to manage them, according to Gartner.
That gap tends to surface the hard way: an agent shares confidential data, or takes an action no one approved. This checklist finds it first. Work through six areas, tick a box only when the control is in place for every agent in production, and score yourself out of 28.
Want the full guide? Download the CIO’s Guide to Governing AI Agents.
Key Takeaways
Treat agents as non-human identities, with their own credentials, owners and lifecycle.
Give each agent only the access its job needs, for only as long as it needs it.
Require human approval before agents send, pay or write to a system of record.
Set a written accuracy target before launch, and monitor against it after.
Scored under 14? Pause new launches and start with ownership and access.
Give Every Agent an Owner and an Identity
An agent running on its builder's API keys is invisible to your identity team. The market is catching on: in October 2026, IBM announced a governed identity record per agent, short-lived least-privilege credentials, and one place to discover and contain them.
Ownership (5 checks)
Each agent has a business owner and a technical owner on record.
Ownership transfers when someone changes roles or leaves.
Each agent has a written purpose: the one job it is approved to do.
There is a full inventory of agents in production, including ones built outside your main platform.
Retired agents are decommissioned, not left running.
Identity (4 checks)
Each agent has its own entry in your identity system, such as Okta or Microsoft Entra, with its own credentials.
No agent runs on a shared or personal credential.
Lifecycle status (draft, testing, production, retired) is tracked per agent.
Each identity entry lists the owner, the job, and exactly what the agent can access or change.
Shrink Access to the Job
The most common failure isn't a rogue agent. It's an agent that can see more than the person asking it.
Access Permissions (5 checks)
Each agent reaches only the data and systems its job requires.
Credentials are short-lived or rotated, never hardcoded in prompts or workflows.
Agents answer only from data the requesting user is already cleared to see.
Access can be revoked immediately.
The tools, MCP servers and integrations each agent can call are listed and approved.
Put People Before Consequential Actions
Agents can draft, check and recommend at scale. Sending, paying and writing to systems of record still need a human signature.
Human Review Points (4 checks)
Consequential actions require human approval.
Low-confidence outputs are routed to a human reviewer.
Reviewers know what they are approving and can reject or edit.
Escalation paths are defined for exceptions the agent can't handle.
Test Before Launch, Trace Everything After
"It seemed to work in the demo" is not an accuracy target. Neither is a log nobody can read back.
Evaluation (5 checks)
Each agent has a written accuracy target, such as 95% correct on real cases, and meets it before launch.
Test sets reflect real inputs, including edge cases and bad data.
Production accuracy is checked on a set cadence, with an alert if it drops.
Model and token costs are tracked per agent.
Changes to prompts, models or tools trigger re-testing.
Auditability (5 checks)
Every run is logged: inputs, outputs, tools called and data sources used.
Answers cite their sources so reviewers can check them.
Writes to business systems carry provenance back to the agent run.
Approval decisions are recorded with who approved and when.
Logs are retained to match your regulatory requirements.
What Your Score Tells You to Do
Add up your checks. Then find your weakest section. Start there, even if your total is high.
0–13, ungoverned. Agents are running ahead of controls. Pause new launches and start with ownership and access.
14–22, partly governed. Some controls exist, but not for every agent. Fix your weakest section before adding more.
23–28, governed. Controls are in place across all six areas. Expand to new teams and higher-risk workflows.
Hold Vendors to the Same 28
A governed platform answers each of these with a feature you can see, not a roadmap slide. Ask to see it live, using one of your own workflows, like an IT support desk or KYC review.
Ownership: Can I see every agent in production, who owns it and what stage it's in?
Identity: Do agents and users sign in through our SSO, with roles we control?
Access: Can an agent ever return data the requesting user isn't cleared to see?
Human review: Where can I require approval before an agent acts, and who can approve?
Evaluation: How do I test an agent before launch and roll it out to a small group first?
Auditability: Can I trace a single answer back to every prompt, tool call and source behind it?
Build Governance in From the First Agent
On StackAI, each checklist area maps to a control in the product: an agent development lifecycle with an owner at every stage, SSO and RBAC, permission-scoped answers across 500+ integrations, approval gates, feature-flagged rollouts, and traceability down to each tool call and prompt. Deploy in our cloud, your cloud or on-prem.
Want to see owners, permissions, approval gates and run logs live on a workflow like yours? Get a demo here.
Sources: IBM Newsroom, "IBM Introduces Identity and Control for AI Agents," October 27, 2026. Gartner, "Gartner Identifies Six Steps to Manage AI Agent Sprawl," April 28, 2026.

Esther Na
Product Marketing at StackAI