Before You Approve Another AI Pilot, Ask These Four Questions

Before You Approve Another AI Pilot, Ask These Four Questions

Ask a leadership team how many AI initiatives they have running. You will get a number, usually delivered with some pride. Then ask which of those initiatives changed a decision, a process or a role. The room goes quiet.

This is the pattern behind most stalled AI programmes. Not a shortage of ideas, not a shortage of budget, not resistance from staff. A missing filter. Projects get approved because someone credible was enthusiastic, because the tool demoed well, and because saying no to AI in 2026 makes you look like the person slowing everything down.

Enthusiasm is a poor approval criterion. It tells you someone is motivated. It tells you nothing about whether the organisation will be different six months later.

What actually gets approved today

In practice, three mechanisms make the decision instead of the leadership team.

The most visible use case wins, because it tells well in a meeting. The vendor with the best demo wins, because the demo was engineered to remove every objection. The department with spare capacity wins, because it is the only one that can start now.

None of those three mechanisms measures impact. They measure availability and narrative ease. That is how you end up with eleven pilots, a lot of activity, and not one decision made differently.

The four questions

This filter takes twenty minutes and one page. It does not replace a business case. It mostly prevents you from launching things that are not projects yet.

1. Whose work changes on Monday morning?

If the answer contains the words team, cross functional or eventually, there is no project yet. You need to be able to name people, a specific task, and the moment in the week when that task happens. An AI project that changes nobody's day will change nothing at all.

This is the Actors dimension of the MAKIA framework. It is also the question that kills the most projects in the first minute, which is precisely why it should be asked first.

2. What decision gets better, not just faster?

Speed is easy to promise and hard to bank. Producing a meeting summary in three minutes instead of twenty only creates value if something then happens with that summary.

The real question is about judgment. Does someone now decide with better information, a better exploration of options, better perspective? If the answer is no, you are funding comfort rather than performance. Comfort has value, but it does not justify a strategic project.

3. Who answers for the output when it is wrong?

A name, not a department. If nobody can be named, usage will stay cautious, marginal and invisible, or it will go the other way and become quietly reckless.

This single question settles a large part of the governance debate. A named accountability produces healthier behaviour than a ten page policy nobody rereads.

4. What would make us stop?

Define the failure criterion before you start. Three months, one indicator, one threshold. Without it, a pilot never stops. It falls asleep, keeps consuming licences and occupies mental space in the organisation.

A project where you cannot describe what would make it fail is a belief, not an experiment.

What this changes in practice

Four effects show up fairly quickly in organisations that adopt this filter.

The number of projects goes down and their depth goes up. Three well framed initiatives produce more transformation than twelve scattered pilots.

The conversation changes shape. People stop comparing tools and start describing real work.

Teams get involved earlier, because the first question forces you to go and talk to them before approval rather than after.

And stopping becomes acceptable. When a stop criterion exists from day one, pulling the plug is no longer a political failure, it is an experimental result.

To close

Most leaders assume their AI difficulty is technological. It is almost always decisional. The problem is not knowing what AI can do, it is knowing what the organisation chooses to do with it, and what it chooses not to do.

A simple filter applied consistently beats a forty page AI strategy applied nowhere.

Try this week

Take your current list of AI initiatives. Pick the one with the most momentum right now, not the weakest one. Answer the four questions in writing, one page maximum.

If you cannot name the person whose Monday morning changes, it is not a project yet. Twenty minutes is enough to find that out, and it saves six months of activity with no effect.