The pilot that everyone forgot was running
Every mid-market company has one. A pilot kicked off with real momentum twelve months ago — a steering committee, a Slack channel, a vendor demo that wowed the executive team. Today nobody knows what state it’s in. The Slack channel has three messages from the last quarter, all from the vendor’s customer success rep. The vendor still invoices monthly. Nobody has the appetite to kill it because nobody has the appetite to admit it was a bad call.
That is what a stalled pilot looks like at month nine. The damage is mostly invisible at that point. The real damage was done in the first six weeks.
What a stalled pilot actually costs
The line-item cost is the smallest part. A typical mid-market AI pilot runs $80–250k all-in for the first six months — vendor fees, internal labor, infrastructure, the inevitable consulting carve-out. Companies budget for that and account for it.
The line item that doesn’t get accounted for is executive attention. A pilot that requires the CIO to spend two hours a week, the COO to spend an hour, and a director-level sponsor to spend ten hours, costs the company $4–6k a week in fully-loaded exec time. Stretch that across six months of stalled-but-not-killed status and you’ve burned $100k+ in attention before anyone notices.
The third cost is the one that hurts most: organizational appetite. Every stalled pilot eats the appetite for the next one. Engineering teams who watched a six-month effort produce nothing are slower to staff the next initiative. Executive sponsors who got burned once require triple the proof to sponsor a second time. Companies that have lived through two failed pilots will often refuse to fund a third even when it’s the right call.
The visible cost is real. The invisible cost compounds.
The week-2 signal
Here is the signal that predicts a stalled pilot, visible by week two of any engagement. It’s a question. Ask the project sponsor, the vendor lead, and the engineering lead the same question separately:
“What does success look like at the end of this pilot, in a single sentence?”
If you get three different answers, the pilot is going to stall. Not might — will.
The reason is mechanical. Pilots stall because the people responsible for them have different mental models of what they’re building. The sponsor thinks they’re buying a capability. The vendor thinks they’re proving a product. The engineering lead thinks they’re integrating a system. Each is correct in their own frame. None of them is steering toward the same finish line. Six weeks in, the engineering lead has built a brittle integration the vendor doesn’t want to support; the vendor has produced a demo the sponsor finds underwhelming; the sponsor has lost confidence and stopped showing up to the standup. Nobody is technically wrong. Everyone is on a different track.
I’ve now seen this pattern fifteen times. The single-sentence answer is the cheapest, fastest diagnostic in AI program management. It costs zero dollars and takes five minutes to run.
What a non-stalled pilot looks like at week 2
In contrast, here is what a healthy pilot looks like at week two:
- Sponsor, vendor, and engineering lead give the same one-sentence success definition. Phrasing varies; substance is identical.
- There is a single document — not a deck — describing the architecture, the data flow, and the integration points. Everyone has read it. Everyone has redlined it at least once.
- There is a named owner for each integration point. Not “the data team” — a person.
- There is a kill criterion. Something like “if we can’t get to 80% accuracy on this benchmark by week 6, we kill the pilot.” Stated explicitly. Stated in writing.
- There is a calendar. Weekly review with a fixed agenda. A live dashboard, even if it’s ugly.
That’s the boring stuff that prevents stalls. None of it is technical. All of it is operational.
Three early-warning behaviors that always precede a stall
Beyond the one-sentence test, these three behaviors are predictive — when you see two or more of them by week three, the pilot is going to stall:
- The standup is getting harder to schedule. Calendar Tetris is the leading indicator of executive disengagement. Once the weekly review starts slipping or getting cut to fifteen minutes, the pilot is being deprioritized.
- The vendor is sending updates instead of asking questions. Healthy pilots produce a lot of questions from the vendor — about data, about edge cases, about constraints. Stalled pilots produce status updates from the vendor with nobody on the customer side asking follow-ups.
- There is no production-equivalent test environment yet. If by week three you don’t have a sandbox that mirrors production data shapes (even with synthetic content), you will discover all the integration problems in week eight, not week three. By then the timeline has eaten its own buffer.
What to do in week 2 if you’re seeing the signals
You have three options and they get harder in this order:
Option 1: Realign. Pull the sponsor, vendor lead, and engineering lead into a single 90-minute working session. Force a written single-sentence success definition. Force a kill criterion. If they cannot produce one in the room, you do not have a pilot — you have a research project that’s pretending to be a pilot.
Option 2: Reframe. If realignment fails, openly reclassify the work as exploration. Different cadence, different budget envelope, different exec expectations. Some research projects deserve six months. None of them deserve to masquerade as pilots.
Option 3: Kill it. This is the option nobody likes. It is almost always the right one. Killing a pilot at week three is cheap. Killing one at month six is expensive. Letting one zombie for nine months is catastrophic.
The pattern I see most often in healthy AI programs is that the first three pilots are short. Six to eight weeks, narrow scope, kill criteria built in, ruthless about reclassifying when scope creeps. Companies that learn to kill pilots early build the appetite to fund bigger ones later. Companies that let pilots zombie burn the appetite for everything that comes after.
The framing question
If you’re running an AI pilot right now and you’re not sure whether it’s stalled, ask yourself: if this pilot were killed tomorrow, what would actually change in the business in 90 days? If the answer is “nothing,” it’s already stalled. You’re just waiting for someone to admit it.
The companies that get AI right aren’t the ones that don’t have failed pilots. They’re the ones that fail in week six instead of month nine.
If you’re looking at a pilot that’s drifting and you want a second opinion, an AI Readiness Audit gets you a written diagnosis in two weeks. Or schedule a call and we’ll do the single-sentence test on the line.
