Decide How It Ends

IBM put out a study of chief AI officers on November 17. Of the 1,667 senior AI executives they surveyed, only 18% say they have full control over AI investment decisions.
The people companies hired to run AI mostly don’t fully control where the AI money goes, which means they don’t fully control when it stops either. A project gets funded, it gets a sponsor, and from then on ending it is a negotiation with the sponsor.
In Quit, Annie Duke describes kill criteria: a measurable state plus a date, set before you start. I wrote about the same idea in AI Pilot Purgatory last year (set the measure of success up front, not after six months of “refinement”). I still believe that, but a threshold doesn’t do much if nobody has the authority to act on it. So I’d add one more line: the name of the person who can end the project, written down and signed by the sponsor before the money moves. I’ve been calling the whole thing a stop rule.
The usual answer is that we already have stage gates. Most companies do have a steering committee and a quarterly portfolio review. Look at how those usually run, though: the gate criteria get set after the money is committed, often by the team doing the work. There’s no date, so the review slides. And the sponsor sits on the committee that reviews their own project.
A stop rule flips that. The sponsor signs it while they’re asking for the money, which is when they’re most optimistic and have the least room to argue. Six months later, when the numbers come in soft, the only question is whether the sponsor’s own terms were met.
Bad news doesn’t travel up a company well…it gets softened on the way, explained, moved to next quarter’s roadmap. IBM found that 40% of AI chiefs say their organizations discuss the opportunities the performance data reveals and then delay acting on them. That number is about acting on opportunities rather than failures, but it’s the same habit as the review that keeps sliding. A stop rule gets the bad news to land before anyone has talked it into something gentler.
Most AI chiefs don’t treat this as a priority yet. Only 31% rank clear thresholds for stopping AI initiatives as critical to AI success. Among the ones IBM calls high-impact, 96% say they stop or cut investment in initiatives that fail to deliver, even when a senior leader sponsors them, against 46% of everyone else. That’s self-reported, and it doesn’t prove that stopping things made those teams effective. It could run the other way. But the gap fits what I’d expect: the teams getting results are the ones where someone can tell a sponsor no and make it stick.
So whose name goes on the stop rule? With only 18% of AI chiefs holding full control of the money, writing down your AI chief by default may not mean much. It has to be someone who can make a stop stick and who isn’t the sponsor. When the sponsor also owns the budget, that means going one level up: the CFO for a business unit’s project, the board for the CEO’s. And yes, that means somebody writes down, in advance, who gets to end the CEO’s favorite project.
There’s a fairer objection than stage gates, which is that stop rules kill exploration. Some AI work really is a bet, and you can’t know in advance what number should stop it. A stop rule can be a learning threshold: by March 31 we’ll know whether the model can hit the accuracy we need on our own data, and if it can’t, we stop. When the date comes you’re allowed to renew it, with a new threshold and a new date, signed again. What you can’t do is let the old one quietly roll into another quarter.
Pull up the list of AI initiatives you’re paying for right now. For each one, write down the result that ends it, the date you’ll check, and the name of the person who can end it. Then get the sponsor to sign it. For projects that are already funded, do it at the next budget review, since that’s the next time the sponsor is asking for money.
If you can’t fill in the name, nobody can stop that project right now.
Forward this to whoever owns your AI budget before the next review. If someone forwarded it to you, subscribe below to get the next one.