Which AI pilots do we kill this quarter?

This brief is illustrative. Real briefs follow the same structure and are published only after the member who brought the decision has approved every line.

The decision, as the member framed it

A CIO at a mid-size insurer arrived with roughly forty AI pilots spread across six business units, no shared register, and a board that had started asking for “the AI number”. Legal wanted an inventory. Finance wanted a total. Nobody in the organisation had ever stopped a pilot.

The question they brought: how do we decide what to fund, gate or kill, without spending six months building a governance framework first?

What had already been ruled out

• A five-tier risk matrix. Attempted in the previous year; abandoned because nobody maintained it.
• Centralising all AI work under one team. The business units would route around it.
• Waiting for the EU AI Act timetable to settle before acting.

The three questions the table answered

1. What is the smallest inventory that is still honest?

The room converged quickly: four fields, not forty. What it does, who owns it, what decision it touches, and whether a person reviews the output. One member described maintaining anything larger as “a register that describes a different company every quarter”.

2. Who has the authority to kill a pilot?

This is where the room spent most of its time, and where the disagreement was sharpest. Three of the eight held that killing must sit with the business unit that funded it. Two argued that a unit has never once killed its own pilot and never will. The unresolved point: whether a central function can stop work it does not fund without becoming the department of no.

3. What would have to be true to move a pilot into production?

Agreement was faster here. A named accountable owner, a measured baseline from before the pilot, and a documented answer to “what happens when it is wrong”. A pilot that cannot produce all three is not ready, regardless of how good the demo is.

Where the room disagreed

On tiering. Half the table tiers by technology — generative, predictive, agentic — because it maps to the regulation. The other half tiers by consequence — who is affected when it is wrong — because it maps to the actual risk. Nobody changed their mind. Both positions are recorded here because the disagreement is more useful than a false consensus would be.

What the member decided to do

1. Publish the four-field register within three weeks, accepting it will be incomplete.
2. Tier on consequence, with two tiers, and revisit once the regulatory position is settled.
3. Give the central function the authority to gate production, not to kill pilots — which sidesteps the argument the table could not resolve.
4. Take one number to the board: how many pilots have a named owner. Not the total.

What this brief deliberately does not contain

The insurer’s name, the member’s name, the six business units, the vendors involved, and the actual pilot count. None of those change the reasoning, and all of them would identify the room.