AI for leadership: what do you decide yourself and what can you delegate?
An executive team does not need to know how a language model works. It does need to know which decisions belong to it. In practice there are four: which challenge you take on, what an agent may do on its own, who owns it, and when you stop. The rest is execution and you can delegate that. This page is the framework, not a course.
Why does it stall in the boardroom?
The pattern is the same everywhere. Someone comes in with a proposal, technical terms fly across the table, the rest of the leadership team drops out, and the decision is postponed or handed to IT. Six months later four small trials are running that nobody can judge, and nothing has changed about how the company decides.
The cause is not a lack of knowledge. It is a lack of a framework. An executive team assessing an investment in a machine has a fixed set of questions for that. For AI that set does not exist yet, so proposals get judged on enthusiasm or on distrust.
Where it really stalls
- Of the Dutch companies that do not use AI, 74.6 percent name a lack of experience as the main reason. Not the cost, not the law, not the technology.
- At companies with 20 to 50 employees, 33 percent used AI in 2025; at 100 to 250 it was 52 percent. The difference is not in the availability of the technology, because that is the same for everyone.
Source: CBS, reporting years 2024 and 2025. The full accounting is in how many Dutch companies really use AI.
Which four decisions belong to the executive team?
You cannot delegate these four, because all four are about risk and direction and not about technology.
- Which challenge. Which recurring decision is now too slow or based on too little information? That is a leadership question, not a technical one. Leave this to IT and you get a solution for the problem IT sees.
- How much autonomy. May the system only advise, may it draft, may it act after approval, or may it act on its own within set limits? Four levels, and the difference between level two and level three is the difference between a tool and a signature.
- Who owns it. One name, not a department. Someone who is accountable for what comes out and who has the authority to switch it off.
- When you stop. Agreed up front. A project without a stopping criterion is not a trial but a purchase with a delay.
Which questions do you ask about every AI proposal?
Seven questions. Anyone who can answer all seven has a proposal; anyone who cannot has an idea.
- Which decision gets faster or better because of this, and how do we notice?
- What is the baseline measurement? How long does this decision take now and how much digging does it cost?
- Which sources does the information come from, and can we get to those sources?
- Can every figure be traced back to the source it came from?
- What may the system do on its own, and where does a person approve up front?
- Who is the owner at our company, and how much time does it cost him per week?
- At which outcome do we stop?
Question four is the most important one and is skipped most often. A figure that is not traceable gets waved away in the meeting, and rightly so. Then you have paid for something nobody dares to use.
Which decision belongs at which level?
| Decision | Where it belongs | Why |
|---|---|---|
| Which challenge we take on | executive team or leadership team | about priority and about where the money sits |
| Level of autonomy of the system | executive team or leadership team | touches liability and reputation |
| Access to sources and data | executive team together with the person responsible for privacy | a legal decision, not a technical decision |
| Choice of vendor or model | delegated, within set limits | a specialist call, provided traceability and exit are arranged |
| Setup, connections, technology | delegated | execution |
| Continue or stop after the trial | executive team or leadership team | only they can hold the outcome against the strategy |
How do you put AI on the leadership team agenda?
Not as a separate agenda item called "AI", because that turns into a presentation. Instead as a fixed part of three existing items, fifteen minutes each time.
- With the numbers. One question: which figure that we are discussing today could we have known earlier? That is your list of candidate challenges right there.
- With the running trial. Status against the baseline measurement, in two numbers. No demo, no screen.
- With the risks. Once a quarter: what are our people doing with AI outside our view? In most companies that is more than the executive team thinks, and it is a bigger risk than the trial you run yourself.
How do you recognize a weak proposal?
Five signals. If you hit two or more, postpone the decision until they are gone.
- It starts with a product or a vendor and not with a decision that is too slow today.
- There is no baseline measurement, only a promise about a percentage.
- Nobody can explain where a number comes from.
- No owner has been named, only a project group.
- There is no stopping criterion, only a next phase.
When should you not use this framework?
This framework is about AI that prepares decisions. For two other situations it does not fit.
If it is about individual tools, so employees using an assistant for writing or summarizing, that does not belong on the leadership team agenda but in your policy: what may go in and what may not, and with which data. That is an agreement of half a page, not a decision process.
If it is about a system that takes decisions about people, for example in hiring, appraisal or dismissal, this framework is too light. Heavier requirements from the European AI Act apply there, and legal advice belongs up front, not afterward.
How we do this
The Sixth Seat is a Dutch AI company that builds AI agents and Management Intelligence for growing companies. We never start with a tool and always with a decision that keeps getting stuck. The decision scan is ten questions long and gives you your three biggest decision bottlenecks within three minutes, with an estimate of what they cost per year. That is exactly the material you need for the four leadership decisions above.
Then comes the 45-day pilot: one application, a baseline measurement up front, and on day 45 a decision to embed or to stop. What an agent may and may not do on its own we record per application, as described in AI at the table, people at the helm.
Frequently asked questions
Does an executive have to be able to work with AI?
No, but it helps enormously. Anyone who has once seen for themselves what a model can and cannot do asks better questions about a proposal. An hour of trying it yourself is more useful than a one day course.
What if the leadership team is divided about AI?
That is normal and rarely a problem. Disagreement about the technology you settle with a bounded trial that has a stopping criterion. Disagreement about the direction you do not settle that way; that is a strategic conversation you have to hold first.
Who should lead AI in the leadership team?
Someone with an operational interest, not someone with a technical interest. The director who suffers from decisions that take too long is a better lead than the most technical person at the table.
How much time does this cost us?
For the executive team: fifteen minutes per leadership team meeting and two sessions of an hour during a trial. For the owner: two to three hours per week. If that is not possible, do not start.
Are we better off waiting until the technology matures?
Waiting costs nothing in technology and a lot in experience. The 74.6 percent who say they are not starting for lack of experience keep that gap in place. One bounded trial of six weeks is the cheapest way to close it.
Sources
- CBS, Use of artificial intelligence (AI) by companies is increasing, February 27, 2025. Reporting year 2024, companies with ten or more employees.
- The four leadership decisions, the seven questions and the levels of autonomy come from our own approach and from the governance agreements we record per application.