Four decisions define the agents mandate. Most boards have assigned none of them

Ask your operations lead who is permitted to switch off a live agent, and give them a day to answer. In most organisations running agents in more than one function, nobody comes back with a name. The agent stays on, not because anyone decided it should, but because the authority to end it was never assigned.

That unassigned authority is one of four decisions that make up the agents mandate. The others are the limits on what an agent may do without a person approving it, accountability for an action the agent performed correctly under a flawed instruction, and ownership of the record showing the action was reasonable at the time. None of the four is technical. All four tend to sit outside every existing job description, which is the structural reason the chief AI agents officer title appears at all.

Three counts, three definitions

The first figure comes from IBM’s 2026 study of 2 000 chief executives, which puts the share at 76 per cent against 26 per cent the previous year. The second comes from the 2026 AI and Data Leadership Executive Benchmark Survey, which records 38,5 per cent of organisations having appointed a chief AI officer. The third, and the narrowest, comes from a September 2026 census finding the enterprise-wide title at 24 of the Fortune 500, rising to 54 when combined data and AI titles are included.

Treat all three as indicative rather than authoritative: two are self-reported surveys and the third depends on what companies choose to disclose. The direction is consistent, the absolute level is not, and the spread tells you that intent is running well ahead of appointment – which is also why the title arrives attached to several different mandates.

The mandate that makes an agents officer distinct

A chief AI officer, broadly defined, owns the organisation’s use of models: tooling, literacy, data access, vendor relationships, policy. That mandate is largely about capability.

The agents mandate is narrower and harder, because it concerns software taking actions with consequences outside the company. The four decisions from the opening sit inside it and nowhere else:

• Which actions an agent may take without a person approving them. Not a technical setting but a standing decision about reversibility and exposure, reviewed as the agent’s remit grows.
• Who answers for an action the agent took correctly under a flawed instruction. This is the question most organisations discover they have not answered, usually after the first incident.
• What evidence exists that an action was reasonable at the time. The decision log is an accountability artefact before it is a debugging tool, and somebody has to own its contents.
• When an agent is withdrawn. Programmes accumulate agents that nobody switches off because no one holds the authority to do it.

An organisation running agents in two departments already faces all four questions. Whether a dedicated post answers them is a separate matter, addressed below.

Role-scope matrix

Two rows in this matrix cause most of the friction in practice. Autonomy limits sit with the agents officer but are implemented by whoever builds the integration, which makes the boundary easy to erode one exception at a time. Withdrawal authority is the row most often left unassigned, and an unassigned row means the agent stays live by default. Where the architecture owner and the agents owner disagree on a limit, the escalation path needs to be written down in advance rather than discovered during the disagreement – a question treated separately in the note on decision rights between the CTO and the CAAO.

Reporting line and budget ownership

No settled practice exists here, and anyone claiming otherwise is describing a preference. The benchmark survey records no consensus among respondents, with the most common single arrangement accounting for roughly a third of cases. An analysis of publicly disclosed structures reports slightly over half reporting to the chief executive, around three in ten to the CTO and the remainder to the COO or chief data officer – drawn from a small sample of disclosures rather than a representative survey.

What matters more than the line itself is a single constraint: the person who decides what an agent may do without human approval cannot report to the person whose numbers improve when that approval is removed. Where the agents officer reports into operations or into a revenue function, the autonomy decision needs an independent escalation route, typically to the audit or risk committee. This is the same separation applied to any control function and it is the one thing in this section worth insisting on.

On budget: the mandate works when it carries authority to stop spending, not to start it. An officer who can commission pilots but cannot withdraw them accumulates a portfolio nobody prunes.

Signals the role is premature

• Agents are confined to one department and every action is reviewed by a named person before it takes effect.
• No agent touches a customer, a payment or an external commitment.
• The organisation has not yet had a disagreement about whether an agent should have acted.
• Automation work is still being selected by business case rather than by queue, with fewer than a handful of live flows.

In this state the four decisions can be held by an existing executive with the remit written into their objectives. Creating a post ahead of the volume produces a role defined by its own case for existing.

Signals the role is overdue

• Two or more functions run agents and the autonomy limits differ between them for no deliberate reason.
• An action reached a customer and the internal discussion turned to whose decision it was.
• Nobody can produce, within a day, a list of live agents and what each is permitted to do.
• Approval queues have grown to the point where reviewers batch-approve, which converts oversight into paperwork.
• A vendor’s change altered agent behaviour and the organisation learned about it from an outcome rather than from a notice.

Three of these five are present in most organisations by the time a second department goes live. The third signal is the quickest test, and it is the one from the opening of this piece: ask for the list, give it a day, and see what comes back.

Fractional before permanent

For organisations hitting the second set of signals without a candidate internally, a part-time arrangement answers the four decisions sooner than a search does. The constraint is that fractional ownership works for decisions and standards, and does not work for incident response, which needs someone reachable within the hour. Scope, days and exit terms matter more than seniority here, as set out in the note on fractional contract scope. Where the brief needs drafting against a specific agent estate rather than in the abstract, that is work for an adviser who holds this mandate elsewhere and can write it from the decisions rather than from the title.

The decision in front of the board

Do not start with whether to appoint. Start by assigning the four decisions – autonomy limits, accountability for outcomes, the decision log, and withdrawal authority – to named people, with the escalation route for autonomy written down. That exercise takes one meeting and produces one of two results.

If every decision lands with an existing executive who has the time and the standing to hold it, you have your answer and no appointment is needed this year. If two or more decisions land nowhere, or land on someone whose incentives conflict with holding them, the gap is structural and will not be closed by a policy document. That is the point at which the role earns its cost – and the brief writes itself from the unassigned rows, which is a better starting point than any job description taken from the market.