Agentic AI for executives: four decisions only you can make

In this lesson you'll learn:

• the four agentic AI decisions that belong to leadership, and what to ask for each;
• the four questions that keep them true, and the maturity check that replaces tool counts;
• seven signals worth escalating, and a first thirty days that needs no budget approval.

Sound familiar?#

• The organisation has an agent-first directive and a portfolio of pilots, and none has reached production.
• The board pack shows one accuracy number and a count of AI tools adopted.
• The largest AI cost line arrived as a surprise, from finance.

None of these is a delivery failure. Each is a leadership decision being made by default, somewhere below the level that should be making it.

The four decisions, step by step#

Decision 1 · Which work is genuinely AI work#

Ask it of the roadmap, not of one feature, and expect most of it to come back as rules. Three questions in order (is there a genuine judgement call, is the volume high enough, is a wrong answer recoverable) and the first "no" ends it. A team under an agent-first directive will build agents for things a rule does better, and will not tell you, because you asked for agents. Ask for an AI-fit record per candidate, with the rejected alternative. Healthy answer: two or three of your top five are rules.

Decision 2 · What an agent may do without a person#

Per action, never per product, following what a mistake costs and whether it can be undone: acts alone, acts monitored, acts inside a veto window, needs a named approver every time, or is not delegated at all. Ask for the autonomy record, then the follow-up that matters: show me the cap. If someone opens a prompt file rather than a code file, you have found the gap that produces incidents.

Decision 3 · What you will accept as evidence#

The decision that most often goes by default. Accept a demo and you will be shown demos; accept a single accuracy figure and the risky slice will hide inside the average. Ask for a score per slice against a derived bar, with its lower bound, plus a shadow run against the people doing the work today.

Decision 4 · What you will fund past the first cycle#

A first cycle that saves time and costs more is normal. Funding past it should depend on a trajectory (the review load falling, the re-run count falling with it) not on a promise. Ask for two numbers, every cycle, from the team: the time saved and the money spent, on one line.

The four questions, every cycle#

The last question replaces the most misleading metric in the field, tools adopted, with a count of six controls a team either has or does not. The maturity model

Seven signals worth escalating#

A single accuracy number in a board pack · a cap that lives in a prompt · a launch date agreed before a shadow run · tool adoption reported as maturity · a postmortem that produced a name · a cost number that arrived from finance · a dashboard that stopped moving while work continues. None is a failure; each is a decision being made by default.

Your first thirty days#

1. Ask the four questions at the next review, then say nothing for ten seconds.
2. Ask to be shown one cap the team says is enforced.
3. Take a baseline, person-days per story, on whatever starts next. It cannot be recovered later.
4. Run the six-control check with the team, ticking only what can be shown in ten minutes.
5. Name who owns governance. If the answer is "we all do", nobody does.
6. Fund the gateway: a fortnight of platform work that turns every cost conversation into a diagnosis.

None of this needs a reorganisation, a new function or a vendor commitment. Rolling out agentic delivery

Why it matters#

Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, for escalating costs, unclear business value or inadequate risk controls. All three are leadership decisions left to default: what counts as value, what the agent may do, and what counts as evidence.

Key takeaways#

1. Four decisions are yours: what is AI work, what agents may do alone, what counts as evidence, what gets funded.
2. Four questions, every cycle, and a maturity count of six controls instead of tools adopted.
3. Thirty days, no budget: ask the questions, see one cap, take a baseline, name the governance owner.

FAQ#

What should executives know about agentic AI?#

That most of the portfolio is not AI work; that "does it work" is a number per kind of case, derived from what a mistake costs; that autonomy is decided per action and enforced in code, not in prompts; and that a first cycle can save time and cost more, which is survivable only if the team reports both numbers.

What questions should a board ask about AI agents?#

Which of these are rules? What may it do without a person, and who decided? What are the two numbers, the saving and the spend? What maturity level are we, and what is the next control?

How do you measure AI maturity in an organisation?#

By controls, not tools: a context file the agents read, a spec with a bar and an owner for every item, a harness that blocks merges per slice, caps in tool signatures, a trace that redacts, and production evidence with drift watched. The level is how many of the six a team can show.

Should we mandate that teams use AI agents?#

Mandate outcomes and evidence, not agents. A quota for agents produces agents for work a rule would do better; a requirement for an AI-fit record and two numbers per programme produces the right mix.

Apply it in your role#

Across the enterprise. The executive team asks the four questions of the whole portfolio each quarter, and the maturity count per team replaces the count of tools adopted.

The ten-minute workflow. Rehearse the review before it happens:

Play a sceptical chief executive reviewing our AI programme. Ask me the four questions one at a time:
which of these are rules; what may it do without a person, and who decided; what are the two numbers;
what maturity level are we, and what is the next control. Push back on any answer that lacks a
number or a named owner.

Sources and credits#