Most companies now have an AI policy, but far fewer have an operating model that says who approves a new use case, who can stop one and who answers when an agent misbehaves.
An AI governance operating model is the organizational design behind the controls: the bodies, roles, decision rights, workflows and metrics that make policy real. The evidence here comes from published surveys by McKinsey, IBM, Gartner, the IAPP and SailPoint, from OMB memorandum M-25-21 and from the 2026 U.S. bank model risk guidance.
Each survey reflects its own publisher's sample, so none of it locates your company in a distribution. Regulatory mapping is left to the companion paper, and this playbook covers the people and the process.
None of this needs a new framework. It needs three decisions written down: where approval rights sit, what the council is allowed to do, and how agents are recorded and controlled. Each one can be checked by asking for a single document, and each one is missing in a company that has a policy but no operating model.
Decide whether approvals are centralized, federated or hub-and-spoke. Centralized suits the first year and a small portfolio but turns into an approval queue as volume grows. Federated is fast and tends to produce shadow AI. Hub-and-spoke fits most companies with AI in three or more units, with the hub owning standards and high-risk approvals. The artifact is a written archetype decision with the thresholds at which a spoke lead approves alone.
Charter the council or ethics board as a decision body of seven to eleven voting members. Write its rights as verbs: approve, approve with conditions, require remediation, defer, reject and suspend a live system. Any member can trigger an emergency suspension review within 48 hours, and every decision gets a record. The artifact is a three-page charter that names the suspension right and the body allowed to override it.
Treat each agent as a distinct identity with an accountable owner, a registry entry and scoped permissions. Record its model, framework, tools, credentials, autonomy level, approval gates, kill switch and log location. Grant read access by default and require separate approval for any write, payment, external message or deletion. The artifact is one registry entry per agent, with a dedicated service identity rather than a shared human account.
In the first 30 days, get a board or CEO mandate, pick the archetype, draft the charter and RACI, and open a 30-day amnesty for unregistered AI tools and agents.
Use a single intake form, risk-tier each case by impact, data sensitivity, autonomy and reversibility, and score business value separately so governance does not become a de facto product roadmap.
Low-risk cases go to spoke leads, moderate cases add security and privacy sign-off, and high-risk cases go to the council. Agents that write data or act autonomously move up a tier.
Track registry coverage, median days from intake to decision by tier, validation evidence for Tier 3 systems and incidents by severity, and define these before the first board report.
It is the organizational design that runs AI oversight day to day. It names the bodies that decide, such as the board, an AI governance council and an ethics board, and the roles that hold decision rights, from the Chief AI Officer to product owners.
It also sets the workflow from intake to retirement and the metrics that show whether governance works. It sits underneath your policy and makes it operational.
Most mid-market and enterprise companies should use hub-and-spoke. A central hub owns standards, risk tiers, tooling and high-risk approvals, while spoke leads in each business unit approve low-risk use cases. Centralized works for the first year or a small portfolio but becomes a bottleneck.
Federated models are fast but tend to produce shadow AI, duplicated tooling and inconsistent controls.
A Chief AI Officer runs the governance operating model. Typical duties are chairing or vice-chairing the AI council, owning the use-case inventory and agent registry, setting risk tiers and the intake scoring model, approving or escalating use cases, advising on AI spending and reporting metrics to the board.
IBM found 76% of surveyed organizations had a CAIO in 2026. The test is authority to stop a deployment.
Register every agent with an owner, model, tools, credentials and autonomy level, and give it a dedicated identity with least-privilege tool permissions. Require separate approval for writes, payments, external messages and deletions.
Then match human oversight to autonomy, following Gartner's four levels: review outputs for advisory agents, approve each action for agents that act, and use monitoring, rollback and circuit breakers for autonomous ones.
A working operating model can be stood up in about 90 days, in three 30-day phases: mobilize, build and operate. OMB gave federal agencies 60 days for a CAIO and 90 days for a governance board.
The playbook gives no cost figure, but budget is a known constraint: 64% of organizations surveyed call their governance budget insufficient, and talent is the most cited limit.
You get the working parts of an operating model, ready to adapt. That includes an archetype comparison, a role-by-role guide to decision rights, a 13-activity RACI, ethics board charter essentials, a nine-step intake checklist with a risk tier routing table, an agent oversight table by autonomy level, a capability map, a metrics set, a 90-day stand-up plan and a 12-item readiness scorecard.
It reaches as far as its sources, and no further. The figures come from surveys by McKinsey, IBM, Gartner, the IAPP and SailPoint, each with its own sample and fielding dates, plus OMB memorandum M-25-21 and the 2026 U.S. bank model risk guidance.
None of it benchmarks your company. Detailed regulatory mapping sits in the companion paper on integrated AI governance.
Drop your details and we'll send The AI Governance Operating Model Playbook straight to your inbox - no spam, unsubscribe anytime.
Bring your current AI policy, your use-case inventory and a list of the agents running in production. We will work through where approval and suspension rights should sit, and how to wire intake, tiering and agent permissions into the tools your teams already use. SECTION 7 - FAQ - 5 to 8 questions
Book a governance session