Your Company Has Three AI Pilots Nobody Ever Killed

If you can't name every AI pilot your team has running right now — not the ones you approved, the ones still technically connected and billing — you're paying for at least one nobody is using. The habit worth building isn't starting another pilot. It's putting a date on your calendar to decide whether to end the last one.

Rollouts get a launch. They don't get an end date.

AI pilots start with a kickoff message, a trial license, and someone's good intentions. They almost never start with a scheduled moment to ask whether they're still earning their cost. Six 2026 rollouts that actually worked all had one thing in common: a specific person accountable for a named, visible workflow. That same missing ownership is what lets a failed pilot die quietly instead of getting cancelled — no one was ever on the hook to say whether it was worth the subscription, so no one says anything, and the charge renews itself.

One guide to measuring AI impact found that persistence at month three — is anyone still running the workflow unprompted — is the cheapest, most honest signal a team has. It's also, in most companies, a number nobody checks after month one, because checking requires a date on which somebody is required to answer.

What to do differently Monday morning

A kill decision needs three things an approval decision never required: a review date set before launch, a named owner for that review, and a plain definition of "not worth it" agreed on while everyone still likes the idea — not six months later, when admitting it stalled means admitting you picked it. AI subscription pricing is still subsidised, which means a mediocre pilot is cheap enough to coast for a year past the point it stopped helping anyone. The cost signal that would normally force the conversation doesn't arrive on its own — someone has to go looking for it.

Marcus: the scheduling agent nobody opens anymore

Marcus runs operations for a 35-person architecture firm. Four months ago he piloted an AI meeting-notes agent for client calls, paid for six seats, and got genuine enthusiasm in week one. Checking the vendor's usage dashboard this month — something he'd never done before, because nothing had prompted him to — he found two seats active and the other four untouched since week three. Nobody had complained, because nobody had been assigned to notice.

He didn't cancel on the spot. He asked the two people still using it what specifically it did for them, got a real answer from one and a shrug from the other, and cut the firm down to two seats the same week — a 67% cost cut with zero complaints, because the four unused seats were never doing anything to complain about losing.

Teresa: the review date that made the call easy

Teresa leads a 20-person claims team at a regional insurer. Before her team piloted an AI first-draft tool for claims summaries, she wrote one sentence into the kickoff note: "We review this against rework rate on week eight, and if it hasn't dropped, we stop." At week eight, rework was flat — drafts were faster, but reviewers were catching the same errors they'd always caught, just in a different place.

Because the criterion existed before anyone had a favorite, ending the pilot wasn't a fight. It was just the thing the note said would happen, and it happened.

The one thing

Most AI pilots don't get killed because killing one was never anyone's job. Name an owner, a review date, and a plain number before the pilot starts, and the decision stops being a confrontation — it's just the next item on a date you already wrote down.