12. What the CEO Does, and What the CEO Delegates

Two chapters of rules can leave a CEO with the wrong conclusion: that AI compliance is a legal matter, to be handed to a lawyer and forgotten. It is not. In every framework in the world the person who answers for how AI is used in a company is the person who runs it; the lawyer advises, the consultant builds, IT operates, and none of them signs. The good news is that what the law asks — the six things in chapter 10 — is almost exactly what a well-run AI program does anyway. Compliance is not a layer on top of the method; it is the method with a paper trail.

This chapter is the operating guide. It separates the three decisions only you can take from everything you can and should delegate; says who does what inside the company and what goes outside; places the compliance gates inside the life of an agent so that nothing is checked twice or too late; and ends with a checklist and the clauses to require from suppliers. By the end you will be able to run the first AI governance meeting of your company in an hour, with the right people in the room.

The three decisions you cannot delegate

Where the lines are. Which uses of AI your company will not make, whatever the return — and which uses require your personal sign-off before they go live. Every legal framework draws lines (prohibited practices, high-risk tiers); your company should draw its own one notch tighter, because your reputation is not protected by a compliance certificate. Hiring, evaluating people, granting credit, anything touching health or safety, anything a customer would be angry to discover: those are yours to decide, in writing, once.

Who answers. One person accountable for AI in the company, by name, in the org chart. Not a committee, not "the AI team," not the consultant. In an SME that person is often you for the first year, then a function head with a written mandate. Every framework — the UK's, Korea's, the Gulf's central-bank guidance, ISO 42001 — asks for this, and it is the first question an inspector or a large customer will ask.

How much risk, at what evidence. The level of oversight you require for each class of agent: which agents act alone inside thresholds, which propose and wait, which never act without a person; and what proof you want to see, and how often. This is the same decision you took in chapter 7 for the decision playbook, read from the side of responsibility instead of the side of speed. It cannot be delegated because it is the definition of how much you trust the machine on your behalf.

Everything else — the register, the assessments, the notices, the training records, the contracts, the technical controls — is execution, and execution is delegated.

The inventory: what AI is already inside your company

Before governing anything you need to know what you have, and every CEO I have worked with has underestimated it. AI is already in your company in three forms: features inside the software you already use — the ERP that forecasts, the CRM that scores leads, the email that drafts; tools people have adopted on their own — the assistants, translators and generators of shadow AI; and agents you have deliberately put in place. The UK's approach makes the point explicit: an AI register must include AI features inside existing software, not only what you bought as "AI."

The inventory is a table, one row per system, and it is the foundation of every other document in this chapter.

Watch out. The shadow-AI rows are the ones that matter most and the ones nobody volunteers. Across the surveys, roughly three in four employees use AI regularly and only a third feel adequately trained; the tools they use on their own send your data somewhere you have not approved and produce outputs nobody reviews. Do not run the inventory as a hunt. Run it as an amnesty: two weeks, no consequences, and the promise that the tools people like will be evaluated for adoption. You will learn more about your company in those two weeks than in a year of meetings.

Who does what inside

An SME does not need a compliance department. It needs four people who already exist to add a defined slice of AI to what they do, with the accountable person above them.

The pattern to notice: every role delivers documents that already exist in a well-run project. The training plan from chapter 5, the CoS and thresholds from chapters 3 and 7, the cost per work unit from chapter 4. Compliance adds columns, not work.

What goes outside

Four external roles, each with a defined boundary.

The lawyer confirms the legal basis, the notices, the contracts and — where a law requires it — the impact assessment. Used at three moments: the inventory, every new high-risk agent, every new country. Not used to design agents or to decide risk tiers: that is your call on their advice.

The AI consultant designs the agents with the gates built in (next section), writes the technical and semantic CoS, prepares the documentation an inspector would read, and trains your supervisors. Chapter 8 told you how to choose one; the test here is question 9 of the first meeting — a consultant who treats compliance as "not a problem" is not qualified.

The auditor — internal or external — verifies, once or twice a year, that the evidence file matches reality: that the logs exist, the reviews happened, the training records are current. Not needed in year one for most SMEs; needed the moment a large customer, a bank or a regulator asks.

The certification body applies when you decide to certify against ISO/IEC 42001, the AI management-system standard. More than 350 organizations were certified worldwide by early 2026, including the major model providers; for an SME the standard is worth borrowing as a structure long before it is worth certifying. Certify when a market or a customer requires it, not before — but ask every supplier and consultant whether they are certified, because it is becoming a line item in procurement questionnaires.

Compliance inside the method: the gates in an agent's life

The costly way to do compliance is to check at the end. The cheap way is to place five gates inside the loop you already run — request, negotiation, execution, acceptance — so that each check happens once, at the moment the information exists, by the person who has it.

Gate 1 is the one that saves money: an agent classified as high-risk before design costs a fraction of one reclassified after launch. Gate 4 is the one that keeps you honest: it is the same monthly review where you read the CoS, with three more questions. If your consultant's method does not show you these five gates on a page, ask why.

Example. A staffing agency with sixty employees wanted an agent to pre-screen applications. Gate 1 classified it as high-risk in the EU (employment decisions) and under the Korean and Brazilian regimes it also operated in. The design changed before a line was built: the agent would rank and summarize but never reject; every rejection stayed with a recruiter; candidates were told an AI had helped; the ranking logic was documented and tested for bias on a year of past data. Launch took six weeks longer than the "quick pilot" first proposed. When the first candidate asked how the decision had been made, the answer took ten minutes and one file. (Case anonymized.)

The evidence file: what you show when someone asks

Someone will ask — a customer's procurement team, a bank, a regulator, a candidate's lawyer, a journalist. The answer is a file, kept current by the accountable person, that anyone in the company can find in ten minutes. It contains, for the company: the inventory; the written lines (what we never do, what needs the CEO's sign-off); the training plan and records; the rule on personal tools; the contracts with AI suppliers and consultants. And, for each agent: its classification and the reasoning; the CoS and thresholds; the transparency notice used; the name of the human on the loop; the logs, or where they are; the monthly reviews; the assessments where a law required them; and the change history.

Every item in that list is produced by a gate or by a role in the tables above. The file is the by-product of the method, not an extra project. Where a country requires a specific document — a Korean self-assessment, a Brazilian impact assessment, a Californian notice — it is one more item in the same folder, produced at the same gate.

The CEO's checklist

Ten questions; if you can answer yes to all ten, you are ahead of most companies twice your size.

Clauses to require from suppliers and consultants

Your contracts are where compliance becomes enforceable. Eight clauses cover most cases; your lawyer will adapt the wording.

For consultants, add the three from chapter 8's relationship pact: cost transparency per agent, ownership of the CoS register and the documentation, and the exit clause with full handover.

What to take away

Three decisions are yours: where the lines are, who answers, how much risk at what evidence. Everything else is delegated to four internal roles and four external ones, each with a boundary, and produced at five gates inside the loop you already run — so the evidence file writes itself. Start with the inventory, because you cannot govern what you have not counted; and put the eight clauses in every contract, because the law you answer for is only as strong as the paper your suppliers signed.

Your to-do list.

1. Book one hour with your accountable person, legal, HR, IT and finance. Answer the ten questions in the room and write the "no" answers on the whiteboard: that is your compliance plan for the quarter.
2. Run the inventory as a two-week amnesty, shadow AI included.
3. Send the eight clauses to your lawyer and ask which are missing from current contracts.