The industry's open secret is the pilot graveyard: proofs-of-concept that impressed everyone in the demo and never processed a single real transaction. Having both rescued and replaced our share of dead pilots, we see the same five killers on repeat. None of them are model problems.
Killer #1: Piloting the impressive workflow instead of the valuable one
Teams pick the flashy use case — the executive-facing chatbot — over the boring one with measurable money in it, like invoice exceptions or lead routing. The fix is choosing by arithmetic, not applause: hours consumed × frequency × error cost. The workflow your operations lead complains about weekly is almost always the right pilot.
Killer #2: No baseline, so success is unprovable
Six weeks in, the agent works — but nobody measured the before. How long did a human take? What was the human error rate? Without those numbers the renewal conversation runs on vibes, and vibes lose to budget reviews. Measure the manual process for two ordinary weeks before the pilot starts. It's unglamorous and it's the difference between “seems good” and “saved 31 hours a month”.
Killer #3: Nobody owns it on the business side
IT sponsors the pilot; the team whose workflow it touches hears about it in week four and treats it as a threat or a toy. Every surviving pilot we've seen had a named owner inside the affected team — reviewing escalations, correcting mistakes, and getting credit for the win. Adoption is a people problem that no model upgrade fixes.
Killer #4: Demo-grade engineering meets real inputs
The pilot was built on twenty hand-picked examples; production sent the angled phone photo of an invoice and the email written entirely in the subject line. If the pilot has no error handling, no escalation path and no logging, week one of real traffic destroys trust that takes months to rebuild. Pilot scope can be narrow; pilot engineering has to be production-shaped from day one. This is the core of how we run pilots — narrow but real.
Killer #5: No integration path
The pilot lives in a side tool where someone pastes data in and copies results out. The copy-paste human becomes the bottleneck and the single point of failure, and “rolling out” means starting the integration work that should have defined the pilot in the first place. If the pilot doesn't touch at least one real system — even read-only — it isn't piloting the hard part.
The shape of a pilot that survives
One valuable workflow, a two-week measured baseline, a named business owner, real inputs from day one, at least one live integration, and written success criteria with a promotion decision date. Eight weeks later you either have a production system with numbers behind it — or a cheap, fast, honest “no”, which is the second-best outcome a pilot can buy.