The autopsy of a failed AI pilot almost never blames the right thing. The model was fine. The vendor was fine. Even the change management was usually fine. Watch closely and the same missing piece appears every time: there was a project champion, a vendor contact, an IT owner, and an executive sponsor, and nobody owned the number.
Ask "who is accountable for invoice-processing hours falling 30% by March" and the room goes quiet. That silence is the failure, and it happened months before the pilot did.
Enthusiasm is not ownership
Every failed pilot had a champion, someone excited about the tool, who ran the demos, who forwarded the articles. Champions are necessary and insufficient, because a champion's success condition is the tool getting adopted, not the metric moving. Adoption without impact is the exact shape of the 95%: the software is live, people log in, a dashboard exists, and the P&L cannot tell the difference.
An owner is something else. An owner has a number, a date, and the authority to change the process around the tool. That last clause is the one that matters, because the returns never come from the software alone. They come from the workflow redesigned around it: the approval step removed, the handoff eliminated, the job description changed. A champion cannot do those things. An owner can, and knows they will be asked why the number did or did not move.
This is not an AI insight
We keep pretending AI investments are a new species. They are the oldest species: an operational improvement project wearing a new badge. Nobody would install a new machining center and skip asking who owns parts per hour. Nobody would re-route trucks without owning revenue per route-hour. The same discipline we argued for before spending on AI at all, a baseline, a dollarized target, kill criteria, applies to every operating investment ever made. AI's novelty just gives everyone a license to forget it.
The tell is in the language. Successful projects sound like operations, "we cut quote turnaround from three days to four hours, which lifted our win rate two points." Failed ones sound like technology: "we rolled out the platform across three departments." Rollouts are activity. Numbers are outcomes.
The fix costs nothing
Before the next pilot, write one sentence: [Name] owns [metric] moving from [baseline] to [target] by [date], and has the authority to change the process to get there.
If you cannot fill in the name, do not start the pilot. If the name refuses the number, listen to why, because they are usually telling you the target is fantasy or the authority is missing, and both are cheaper to learn now. And if the number does not move by the date, that was the kill criteria, and the discipline is to kill it, publicly, without shame. A culture that kills weak pilots fast is the culture whose strong ones scale.
One sentence, one name, one number. It filters out most of the 95% before they spend anything. That is the entire operating philosophy of our AI advisory practice, and it is why the engagements start with the P&L instead of the technology.