Why most PMs' AI pilots stall after 60 days

The pattern is almost boring at this point. A PMO runs an AI pilot. Six weeks in, the slide deck looks great — hours saved, sentiment up, three flagship use cases. By month three, usage has quietly collapsed back to the two people who were already AI-curious before the pilot started.

It's not a tooling problem. It's not even really an adoption problem in the usual sense. AI pilots stall because the things that make a pilot succeed are the same things that prevent it scaling.

The honeymoon is doing the heavy lifting

Week one, everyone tries the new tool because it's new. Week six, the novelty's gone and people fall back to whatever requires the least thought. AI usage that depends on intrinsic curiosity will always decay — the curious are a small slice of any delivery team.

The pilots that survive month three did one thing the failing ones didn't: they tied AI usage to a ritual that already existed. Status Mondays. Sprint planning. The pre-steerco huddle. Anywhere there's already a forcing function on the calendar.

Five reasons pilots stall, ranked by how often we see them

1. No owner after the launch. Pilots have a sponsor for eight weeks and then a vacuum. Nobody is accountable for the second quarter's adoption number, so it isn't measured, so it drifts.
2. Tool sprawl. The pilot lands ChatGPT Enterprise, then Copilot lands a month later for the Office crowd, then someone in finance buys Claude. Three tools, three logins, three slightly different mental models. People give up.
3. No prompt library. Each PM starts from a blank box every time. The good prompts that emerged in week three never get shared, codified, or version-controlled. Knowledge stays personal.
4. One bad output kills trust. A hallucinated risk in a board pack burns six weeks of credibility. There's no recovery ritual, so the PM quietly stops using AI for anything that matters.
5. The "AI champion" leaves. Pilots are usually shoulder-carried by one enthusiast. When they move on or get promoted, there's nobody to answer questions, and the rest of the team folds.

What survives looks unglamorous

The PMOs we've seen sustain AI past six months share three traits, none of which are about the model:

• Tied to a deliverable. "Friday status pack uses the AI draft" is sticky. "Use AI more" is not.
• A shared prompt library. Notion page, GitHub repo, Confluence space — doesn't matter. What matters is that when a PM needs a stakeholder update prompt, they don't start from zero.
• A "burned by AI" channel. A Slack channel where people post hallucinations and mistakes openly. Sounds trivial. Has an outsized effect on trust, because it turns failure into a public good instead of a private embarrassment.

The honest fix is structural, not technical

If your pilot is at day 50 and the energy is fading, swapping models won't save it. What we'd do instead:

1. Pick one delivery ritual the team already does weekly. Make AI the default first step. Measure adherence, not opinions.
2. Name a permanent owner with a 12-month horizon — not the original sponsor, not the champion. Someone whose performance review includes this.
3. Publish a v1 prompt library inside two weeks. Ugly is fine. Empty is fatal.
4. Run a "what broke?" retro every fortnight for the first quarter. Trust is rebuilt in public.

Adoption isn't a launch event. It's a habit problem dressed up as a technology problem, and the PMOs that treat it that way are the ones whose dashboards still mean something in January.