Let's be blunt: a successful AI pilot and a scalable rollout are two different animals. I've watched this movie too many times. Warehouse #1 runs a beautiful pilot — pick accuracy up, error rates down, applause in the quarterly review. Then warehouse #2 stalls. Warehouse #3? Nobody even brings it up.
Here's the thing, and it cuts both ways. Pilots succeed on people and investment: your best building, your strongest team, the vendor's on-site engineers, all concentrated on one site. Of course it works. Scaling succeeds on something else entirely: standards, discipline, and a healthy respect for the fact that no two warehouses are alike. Lift-and-shift the pilot playbook, and you're asking for trouble.
In 2023, a 3PL in Ontario, California — e-commerce fulfillment, six buildings in the network — came to me. Building #1 had piloted computer vision for pick verification. Three months in, mis-ship rates had fallen to nearly zero, and the client was thrilled. The CEO waved his hand: roll it out everywhere. Building #2, in Dallas, went live and immediately fell apart. Same SKU, but the master data — item names, barcode rules — didn't match California's at all. The model recognized images just fine, but the moment it tried to match against the system SKU, nothing lined up. Accuracy dropped from 99% in the pilot to the low 90s. The floor crew howled. The project sat frozen for two months of rework, and the data cleanup alone cost more than the entire pilot.
That's barrier number one.
Barrier #1: The Pilot's Data Doesn't Travel to Building #2
AI eats data, not slide decks. Pilot data is usually "renovated" before the project starts: the vendor and the IT team scrub the master data, wire up the interfaces, iron out the exception flows. But building #2? Its WMS might be two versions behind. Its barcode rules were set a decade ago. SKU naming conventions vary building to building, and return reason codes were never standardized.
I've seen too many companies conclude "we solved the data problem" during the pilot. No — you solved building #1's data problem. Every building carries different historical baggage, and each one caps what the data can do. Replication means redoing the pilot's data governance work at every site — and that line item was never in the pilot budget.
The fix: make it an iron rule at pilot kickoff. Every "one-time data cleanup" done for the pilot must be captured as a repeatable script and SOP — and proven at building #2 before the pilot counts as a success.
Barrier #2: The Pilot Dream Team Doesn't Transfer
Behind every successful pilot stands a dream team: the vendor's best implementation consultant, HQ's sharpest IT lead, the most cooperative site manager in the network. Those three groups are not following you to every building.
The harder problem is the people on the floor. Pilot sites are staffed with volunteers — high energy, high buy-in. At building #2, the first reaction to "AI is coming" is "is this going to replace me?" In 2022 I watched an apparel DC where pilot pickers actively helped the AI correct its mistakes, while pickers at the second site deliberately slapped labels on crooked. It wasn't a technology failure. It was a trust failure. The project team spent a full quarter on training and incentives before adoption recovered.
Two sides to this one, and both matter: the technology is the easy part to copy. The people are the hard part. Pilot success creates an illusion that the hard part was technical. It never was.
The fix: give "people" its own budget line in the rollout plan. Every new building gets at least one pilot veteran on-site for three months — training, incentives, exception handling, all of it. Skimp on this and you're planting land mines.
Barrier #3: Scaling to Multiple Sites Rewrites the ROI
This is the one that hurts. Pilot ROI is usually distorted: vendors discount or comp the POC to win the enterprise deal and throw in on-site engineers. The pilot runs in your best building, so the gains look enormous. Then the real quotes arrive: per-building software licenses, per-site integration fees, support billed by the day.
I once walked a client through the real math of a vision-based pick-verification project. The pilot and the rollout had completely different cost structures:
| Cost item | Pilot (1 building) | Replication (buildings #2–3, per building) |
|---|---|---|
| Software licenses | Discounted or free | Full price, per building |
| Systems integration | One-time, HQ IT absorbs it | Per building; legacy systems cost extra |
| Data governance | Done informally by the project team | Redone at every building |
| Training & change management | Pilot team covers it on the side | Separate workstream, separate budget |
| Returns | Best building, biggest baseline | Average buildings, diminishing marginal gains |
See the pattern? Costs climb, returns taper. A pilot with a beautiful ROI says nothing about the rollout's ROI. Plenty of projects die at building #2 — not because the technology failed, but because the CFO looked at the replication math and wouldn't sign.
And here's the contrarian take: the more successful the pilot, the more dangerous the scale-up. Success makes everyone underestimate replication difficulty. Budgets get approved loosely; obstacles get imagined away. The veteran move is the opposite — on the day the pilot kicks off, demand the quote for building #2, and budget against your hardest building, not your best one. A pilot that survives the hardest building is a real success.
Five Questions Before You Replicate
Practically speaking, before green-lighting the rollout, answer these five questions. Can't answer one? Fix it first.
- Can building #2's data interfaces and master data standards reuse the pilot's setup directly? If not, how big is the cleanup?
- Can the pilot team spare at least one person to embed at building #2 for three months?
- Do you have the vendor's multi-site quote in hand — per building, or bundled?
- Does building #2's site manager actually want this project, or was it assigned from HQ?
- Was the rollout ROI recalculated on building #2's real numbers, or just copied from the pilot?
Nail all five and your odds jump dramatically. Miss one, and you know exactly where to start.
One last piece of advice: the next time you run an AI pilot, ask the vendor one question before you sign — "Can I see the quote for warehouse #2 right now?" Ask it, and you'll find that the ending of many "pilot success" stories was written long before the pilot began.
By Lao Mi, Supply Chain Veteran (20 yrs)
