CEO — Charter

In a company where agents do a growing share of the work, the chief executive's job gets narrower and heavier. Narrower because coordination, reporting and analysis stop needing a person. Heavier because what remains is almost entirely judgement under uncertainty — and judgement is the thing that does not delegate.

What this role owns#

Direction. Which markets, which products, what we refuse to do. The refusals matter more than the choices; a company that will do anything has no strategy.

Capital allocation. Where money and attention go. In an AI-operated company this includes a decision most organisations get wrong: how much to spend on building capability versus buying it.

The operating model itself. Whether this framework is followed, and whether it still fits.

Accountability for AI decisions. When an automated system produces a bad outcome, "the model decided" is not an answer a regulator, a customer or a court accepts. Someone is accountable, and in the end it is this role.

The stop button. Authority to halt any system, launch or engagement. Explicitly held here so it is never ambiguous in the moment it is needed.

What is NOT delegated to an agent#

This list is the charter's most useful section, because the temptation runs the other way:

• Hiring and letting people go. Not a capability question — an accountability one.
• Committing capital beyond a defined threshold.
• Anything irreversible: acquisitions, product shutdowns, public commitments.
• Deciding what we will not do. Agents optimise within a frame; they do not choose the frame.
• Judgement calls where the data is absent or contested. Most of the important ones.

Agents brief, analyse, draft and monitor. They do not decide any of the above.

KPIs#

The last three are unusual and deliberate. A chief executive in an AI-operated company is principally a decision throughput function, and throughput is measurable.

AI agents in this function#

Briefing agent — assembles the daily position: what changed, what needs a decision, what is overdue. Read-only.

Research agent — market, competitor and regulatory scanning, with sources. Read-only.

Decision-drafting agent — for a recurring decision type, prepares options with trade-offs and a recommendation. A human decides.

Commitment tracker — watches what was promised, to whom, by when, and surfaces slippage before it becomes a surprise.

None of these has write access to anything outside its own workspace. The pattern throughout this framework: agents prepare, humans commit.

SOPs#

• Weekly operating review — the six KPIs above, exceptions only, one hour, decisions recorded.
• Decision protocol — anything material gets: the question, the options, the recommendation, the decision, and what was rejected and why. The rejected options are what stop the same debate recurring in six months.
• Escalation — what reaches this role, from whom, in what form, and what explicitly does not.
• Monthly capital review — spend against plan, with the AI cost line separated out.
• Quarterly model review — is the operating model still the right one?

Templates#

Project Charter · decision record · board or investor update · capital request · incident escalation summary.

Workflows#

In: exception reports from every department · escalations · the weekly briefing · external signals from the research agent.

Out: decisions with recorded reasoning · capital allocations · direction changes · the stop signal.

Handoffs: COO for execution · CTO for technical strategy · Finance for capital · Legal and Compliance for anything with an obligation attached.

The critical loop: decision → recorded → visible to whoever executes it. A decision that lives only in the head of the person who made it is not a decision, it is a preference nobody else can act on.

FAQ#

Can an AI run a company?#

It can run large parts of the operations. It cannot hold accountability, and accountability is most of what this role is. Regulators, customers and courts require a person answerable for automated decisions — and that requirement is tightening, not loosening.

What changes most in an AI-operated company?#

Decision speed and the cost of analysis. Work that took a team a week now takes an agent an hour, which means the bottleneck moves decisively to judgement — and to whoever is available to exercise it.

How do you stop agents drifting from strategy?#

By writing the strategy down as constraints an agent can be given, and by reviewing outputs on a schedule. Agents optimise for what they are told to optimise for; drift is almost always an instruction problem rather than a model problem.

What is the most common mistake at this level?#

Automating decisions that should stay human because automating them is possible. The test is not capability, it is reversibility and accountability — if a mistake is hard to undo or someone must answer for it, keep a person in the loop.

What else is coming for CEO