Key takeaways

• Written policy alone doesn't govern AI use. Compliance Week's 2026 AI Compliance Survey found 83% of organizations are already using AI tools, yet only 25% have implemented strong governance frameworks.
• A workable framework follows four functions, borrowed from the NIST AI Risk Management Framework: Govern, Map, Measure, and Manage.
• Ownership has to span security, legal, and business unit leaders. No single team can govern AI use on its own.
• Runtime control, monitoring, and readiness assessment are distinct, named parts of the framework. Each needs its own owner and its own tooling, even though this piece treats them as components rather than how-to guides.
• Review cadence should combine a fixed schedule, such as quarterly, with reviews triggered by incidents or new agent deployments.

Introduction

Most organizations already have employees and AI agents doing real work inside the business. Fewer have a framework that governs how that work happens. This piece lays out the principles a workplace AI governance framework should rest on, who owns each piece, the policy layers that turn intent into practice, and how often the whole thing gets reviewed.

Why written policy isn't enough

Most enterprises have some kind of AI policy already. The problem is what happens after the policy gets published. Compliance Week's 2026 AI Compliance Survey found 83% of organizations are already using AI tools, yet only 25% have implemented strong governance frameworks.

Employees don't wait for approved tools. 98% of organizations have employees using unsanctioned AI tools, and 47% reach AI through personal or unmanaged accounts. A KPMG survey found that 48% of employees admit to uploading sensitive data into AI tools.

It needs ownership, enforcement mechanisms, and a way to see what's happening across an enterprise AI workforce that now includes both people and agents.

Principles that hold the framework together

A workplace AI governance framework works best when it's built on a small number of principles rather than a long list of rules. Visibility comes first: security and compliance teams need to know which AI tools and agents are in use, including the ones that were never approved. Enforcement comes second, because an unenforced policy has no effect at the point of use.

Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027. Governance issues in those projects get identified only after production incidents, not before deployment.

Gartner also expects AI agents to outnumber human users 10 to 1 in large enterprises within a few years. A workable framework has to govern agent identities and agent actions with the same rigor as human accounts. We found that 53% of AI interactions are autonomous actions rather than a person typing a prompt, and a McKinsey survey found 93% of respondents have exceeded their AI budgets. Governance has to cover actions initiated by agents directly, not only prompts typed by people.

Ownership: who runs this

Ownership of a workplace AI governance framework can't sit with one team. Security owns enforcement and monitoring. Legal and compliance own the policy language and regulatory mapping. Business unit leaders own the use cases their teams run, because they're closest to what the AI is doing day to day.

Employees carry a share of ownership too, even if it's informal. We found that 93% of prompts put in front of a human reviewer get approved. That approval rate suggests the review step isn't catching much of anything. A review process that approves almost everything isn't reviewing.

The four policy layers

The NIST AI Risk Management Framework organizes AI oversight into four functions: Govern, Map, Measure, and Manage, as described by NIST's AI Resource Center. NIST maintains the full framework as a general reference for trustworthy AI, and it maps cleanly onto the layers a workplace governance program needs.

Govern: the acceptable use policy

Govern starts with a written policy that sets expectations for how employees and agents can use AI. An AI acceptable use policy defines what's allowed, what's restricted, and what data can't leave the organization through an AI tool. This layer sets direction, but it doesn't enforce anything by itself.

Map and manage: runtime control

Map and Manage are where policy becomes enforcement. Runtime control applies rules at the moment someone or something uses an AI tool, including the actions agents take. This layer runs through our enterprise browser for user sessions and our network for traffic that doesn't originate from a managed browser, including much of what agents generate.

Measure: monitoring and visibility

Measure is the ongoing view into what's happening. Our AI Protect capability monitors over 18,200 AI extensions with real-time risk scoring. Monitoring turns a policy from a static document into something a security team can verify against daily use.

Readiness assessment

Before any of this gets built, most organizations need a clear picture of where they stand. A readiness assessment looks at current AI use, existing controls, and the biggest exposure points before governance work starts. The AI Playbook for Security Teams lays out this kind of assessment as a starting point rather than a one-time audit.

Setting a review cadence

A governance framework isn't a project with an end date. It needs a review cadence that matches how fast AI use changes inside the business.

A fixed schedule, reviewed quarterly, keeps policy language, ownership assignments, and enforcement rules aligned with how the organization is using AI. That cadence should also flex around incidents. We scanned nearly 34,000 public MCP servers and found that one in three carries a high or critical severity finding, and 92% of owners have no verifiable organizational affiliation. New agent deployments, new integrations, or a security incident should each trigger an out-of-cycle review rather than waiting for the next quarter.

The Measure and Manage functions from the NIST framework map directly onto this cadence. Measure supplies the data. Manage decides what changes based on what that data shows. An NBER paper found a 14% increase in productivity for call center workers assisted by generative AI. A review cadence built around named functions keeps that upside from turning into unmanaged exposure.

FAQs

Who owns AI governance inside the organization?

No single team owns it end to end. Security owns enforcement and monitoring, legal and compliance own policy language and regulatory mapping, and business unit leaders own the use cases their teams run day to day. Employees also carry an informal share, though the fact that most flagged prompts get approved says more about weak review than about how safe that AI use is.

Isn't a written AI policy enough on its own?

No. Compliance Week's 2026 AI Compliance Survey found 83% of organizations already use AI tools, yet only 25% have strong governance frameworks in place. A policy sets direction, but without runtime enforcement and monitoring it has no way to catch unsanctioned use as it happens.

What's the difference between runtime control and monitoring?

Runtime control applies rules at the moment someone or something uses an AI tool, governing what agents and users do. Monitoring is the ongoing visibility layer that shows what happened, which is what lets a team verify policy against real behavior over time.

How often should the framework be reviewed?

A fixed quarterly schedule works as a baseline, keeping policy language, ownership, and enforcement rules current. That schedule should flex whenever there's a security incident or a new agent deployment, rather than waiting for the next scheduled review.

Do we need a readiness assessment before building any of this?

Most organizations benefit from one. A readiness assessment looks at current AI use, existing controls, and the biggest exposure points before governance work starts, so the framework gets built against an accurate picture rather than assumptions.

Does this framework apply to AI agents or just human employees?

Both. Gartner expects AI agents to outnumber human users 10 to 1 in large enterprises within a few years, and we found that 53% of AI interactions are autonomous actions rather than a person typing a prompt. A framework built only around human users won't hold up against that shift.

What's the tradeoff between AI productivity gains and governance overhead?

An NBER paper found a 14% productivity increase for call center workers using generative AI assistance. The governance work exists to catch the exposure that shows up when nearly half of employees admit to uploading sensitive data into AI tools.