What Is AI Governance? An Operator's Guide

What is AI governance? It is the set of decisions about who may use which AI for what, what has to be true before an AI output takes effect, and who is accountable when it is wrong. It is a set of decisions, not a document.

Whether it helps or hurts depends on one design choice: steering wheel or brakes. A 200-person company whose customer, insurer or board has asked for its AI policy does not need a compliance program. It needs a register of what is in use, a policy that names a mechanism and an owner for each commitment, gates where real information arrives, and a rule for which gates a person holds and which a script holds.

What is AI governance?

AI governance is the system that decides how a company uses AI and who answers for the results. For a company with no risk function, it has four parts:

1. A register. One list of every AI tool and effort in the company, proposed and active.
2. A policy. A short document in which every commitment names a mechanism, an owner and a review cadence.
3. Gates. Checks at the points where new information arrives, such as before a pilot and before scaling, rather than one approval at the start.
4. An enforcement rule. For each gate, whether a person or a script holds it, and what counts as approval.

What goes in an AI governance policy?

Seven parts, each tied to a mechanism and an owner. This structure is one practitioner's, built in an MIT Professional Education course in October 2025: a shape that works, not a standard or a law anyone is obliged to follow. The parts are purpose, scope, principles (ten of them), implementation and monitoring, enforcement, review, and effective date. Implementation runs through four mechanisms: an AI ethics committee, continuous training, stakeholder engagement, and audits and periodic reviews.

The same author added four parts the course template lacked. Treat them as what experience suggests is missing, not as a research finding:

• An AI system of record, one central register of proposed and active AI efforts.
• Human gatekeeping at every lifecycle stage, not only at launch.
• A hallucination and disinformation clause.
• Ethical review staged to technology readiness.

A policy is scored on coverage and on operability. Most of what circulates as an AI use policy template covers the topics but stops short of naming mechanisms and owners. Every principle needs a mechanism, an owner and a cadence, or it is a values statement.

Does governance slow everything down?

Built as brakes, yes. Built as a steering wheel, it can speed work up. George Westerman of MIT Sloan puts the test in one line: "Is your governance more the steering wheel or is it more the brakes?" His example, from a conference talk written up in MIT Sloan Management Review in May 2026, is HCA Healthcare's AI steering committee; the write-up reports no measured results, so treat what follows as a pattern to copy, not a proven outcome. Four properties make that committee an accelerant:

1. The value question comes before the risk question.
2. The same questions are asked at three gates: before model development, before piloting at a small number of sites, and before scaling.
3. Risk questions direct the work rather than halt it.
4. Performance is measured continuously, with a kill switch for projects that miss expectations.

In this pattern, three lighter reviews are faster than one heavy approval, because a single up-front review forces a team to over-specify before it knows anything, then leaves the deployed system unexamined for years. Governance built as brakes stalls the work it was meant to protect, which is part of why AI pilots fail at mid-market companies. So ask whether a project is worth doing before asking what could go wrong, and ask the same questions again each time it earns more budget.

Who enforces the rules, a person or a script?

Both, and the policy should say which for each control. Anthropic's August 2026 playbook for building software with AI draws the line; it is a vendor describing its own tooling, so the control shapes carry over and the product settings do not.

• Advisory controls are anything written in plain language for a model to read. They raise the odds of compliance and guarantee nothing; in the playbook's words, this layer "makes violations rare".
• Deterministic controls are scripts that run outside the model's discretion. This layer, the playbook says, "makes them close to impossible".

Most policies described as enforced are advisory, and that is not nothing. The failure is presenting an advisory control as a guarantee.

The playbook's case for scripts is cost: governance cost rises with AI output, because exceptions wait for committees on a fixed cadence while output does not. A script is written once, enforced on every action, and logs each allow, ask or block decision with its timestamp.

Its sequence starts with people. Leadership lists the human approval gates that must survive, then decides which become automated checks. Each gate defines what counts as approval, such as a named person's sign-off, or the control is theater with a script behind it. Non-negotiable checks sit in settings an individual cannot switch off, and every block names its reason and the route to approval, because a gate that blocks with no remedy turns into a workaround. That is what human in the loop should mean: the useful gate is verification, not approval.

The playbook suggests two measures for the policy, neither with a benchmark figure to aim at: time spent waiting at each approval gate (leading, and it shows what governance costs) and gate violations reaching production, before and after (lagging).

In my own AI operating system, one rule says every save must name the files it touches, so one AI session cannot sweep up another's unfinished work. For weeks it was a written rule the AI was meant to follow, and it slipped. On October 22 I turned it into an automated check: about 70 lines, check and test together, live within minutes. In its first day it blocked five saves, one a false alarm fixed the same hour. The rule did not change. Who holds it did. The check exists because a plan my AI wrote claimed it already did; a test that evening showed it was only a written rule, and the plan was corrected to say so.

What about AI that acts on its own?

The same playbook carries over the oldest control in enterprise governance intact: the system that produced the work has no way to approve it, and it may act up to the production gate but cannot pass it. Once a team knows what an agentic workflow is, it doubles as a test of when not to use an agentic workflow: if no gate can sit between the system and production, the work is not ready for one.

It makes the separation structural, not instructional: output arrives as a proposal, such as a draft; a gate blocks release until a named person authorizes it; the system runs under its own identity, so the log can tell its actions from a person's; and it holds no standing production credentials. Capability and authority are separate grants.

For AI that runs continuously, the playbook's mature version tiers the response to metric drift: at 1 standard deviation, log only; at 2, the model diagnoses read-only; at 3, it may act, but only by opening a change into the review gate or triggering a pre-approved runbook. Detection stays a script. The pattern assumes a metric with a rolling baseline and continuous production to watch, which a smaller company often lacks, so treat it as where governance ends up, not where it starts.

Does AI regulation apply to my company?

Probably some of it, but not the laws that make headlines. Which rules apply is a question for your counsel, and your own policy should go through them too; this piece offers structure and research, not legal advice.

The three state AI safety laws bind model developers, not companies using their tools. California SB-53, New York's RAISE Act and Illinois SB 315 are all enacted, checked against the bills and official releases. Illinois sets its bar at more than $500 million in annual revenue plus very large training compute, far above a mid-market operator. The three still signal a shape: a published risk framework, an incident route and an audit cadence, so a policy carrying those three will map onto whatever arrives next.

That does not mean no AI law reaches you. A deployer-side layer applies. Most of the detail below comes from law-firm summaries rather than the rule text, so confirm each line with counsel before relying on it.

A 50-person California employer using an AI screening tool is already inside the first row. Colorado's implementing rules are still in draft, so its mechanics may change before they bind, and anything you have read about Colorado impact assessments is out of date: those duties were repealed and never took effect. Processing EU residents' data alone does not trigger the EU AI Act, though that reading, like the 2027 date, is worth confirming. Nothing federal makes these go away: the Senate struck a proposed 10-year moratorium on state AI enforcement 99 to 1 on 1 July 2025, and Executive Order 14365 has no preemptive legal effect on its own.

Not covered here: Utah, Illinois's employment and biometric privacy laws, New York City, and New York State's deployer-side rules. Those are open questions for counsel, not rules known to impose nothing.

Where does governance fail in practice?

In two places: how the AI is named, and what the policy promises.

Calling the AI a colleague. As reported by MIT Technology Review in June 2026, not read in the study itself, Emma Wiles of Boston University found people caught 18% fewer errors when the same work was attributed to an "AI employee" rather than a chatbot, and were 44% more likely to escalate questionable output to a manager than correct it, which cancels the time saved. In the same reported survey of 1,261 managers, about a third said their company frames AI agents as employees, and 23% list them on org charts. People feel on the hook for a colleague's errors and a tool's, but not an "AI employee's". So name AI systems as tools, and record a named human owner for every AI-produced output.

Promising accuracy. Andrej Karpathy, in a public talk from early 2025 that explains rather than measures, describes two fixes for hallucination, and a policy should name which it relies on: training the model to decline at the edge of what it knows, which does not generalize perfectly, or giving it tools such as search, which is why answers with citations deserve more trust. In his account, rare, specific questions are the danger zone, the shape of most business questions, and a model's confidence says nothing about its accuracy. A policy that promises accuracy is unwritable; one that survives review names what gets checked and by whom.

How does a 200-person company start?

Five steps, in this order:

1. Write the register first. Nothing can be governed that nobody can list.
2. Give every AI-produced output one named human owner. Call AI systems tools, not teammates.
3. Write the policy around mechanisms, not values. Each commitment names a mechanism, an owner and a review cadence.
4. Pick the two or three gates that must survive. For each, decide whether a person or a script holds it.
5. Set a review date and the two measures. Waiting time at each gate, and gate violations reaching production.

In the firm's own practice, governance is one of four domains an AI readiness assessment scores, alongside data, workflows and team, and a company with no policy fails it at baseline. The assessment shows which domain is weakest and which workflows need a person at which step, a sensible first step in the AI transformation work the firm does with mid-market operators.

At an enterprise client where I ran an AI readiness assessment, there was very little approved AI and some use on the side: one manager was running work documents through an AI tool IT had not approved. Nobody had a list of the AI tools in use. That list is the register, and it comes first: a company cannot govern tools it has not written down.

If you want to know which of the four pieces your company already has and which gate to write first, a 45-minute working session on your actual list of AI tools is the fastest way to settle it. Book one here.