Your first AI pilot should not be the flashiest project on the vendor's slide deck. That is my verdict, and it comes from watching warehouses learn it the expensive way. The job of a first pilot is not to prove that AI is powerful. It is to prove, to the floor staff and to the boss at the same time, that this thing can run in our warehouse without making the day harder. A first AI pilot that fails on trust costs more than the software. It costs the second and third pilots too, because nobody will fund them.
I have seen this play out both ways. The bad version first. A few years ago, an e-commerce fulfillment warehouse chose intelligent wave optimization as its first AI project. It sounded impressive: the system would build waves, plan paths, and schedule labor automatically. In the first week, a promotion changed the order mix, and the algorithm split several fast-moving SKUs across different waves. Pickers walked extra distance, supervisors overrode the plan by hand, and after two overrides the floor simply stopped following the system's suggestions. The project was still nominally running, but it had been shelved in practice. The algorithm was not stupid. The pilot had chosen a scenario with almost no tolerance for error: get a wave wrong once, and the whole warehouse feels it.
The good version came from a warehouse of similar size in the same period. Its first pilot did exactly one thing: read inbound delivery documents. Before the pilot, receiving staff hand-copied product names, quantities, and lot numbers from supplier paperwork into the system, easily a quarter of an hour per delivery, with frequent copied-line errors. The pilot photographed the document, recognized the fields, and showed the result to a person for confirmation before writing anything. Within two weeks, document entry time had fallen sharply, and more importantly, the receiving team wanted to use it — because when the system got something wrong, they could fix it on the spot without rework. The money saved was modest. The trust earned was not. For the first time, the whole warehouse felt that AI was there to help, not to add work.
Do not start with the flashiest first pilot
A first pilot has to pass three hard tests, and if it fails any one of them, pick a different topic. First, mistakes must be recoverable: if the system suggests something wrong, a person can catch it and correct it on the spot, without a wrong shipment or a wrong booking. Second, it must sit beside the main process, not require rebuilding the whole warehouse process first. Third, the change must be visible within two weeks: less time, fewer errors, or fewer wasted trips, stated in concrete terms by the people doing the work. A project that fails these tests is not a bad project. It is a bad first project.
By that standard, several categories should wait. Anything where the system directly decides — picking paths, automatic task assignment, automatic price changes — should wait. Any large forecasting project that needs the whole warehouse's data cleaned first should wait. Anything touching customer promises, shipping deadlines, or financial settlement should wait. The issue is not that these scenarios lack value. It is that one mistake costs too much to be the place where trust is built.
Three low-risk, high-visibility first-pilot candidates
Three kinds of scenario fit the low-risk, high-visibility standard, in the order I would consider them. The first is document and information reading: photographing delivery notes, carton labels, and return slips, recognizing the fields, and letting a person confirm before anything is written. The technology is mature, errors are intercepted before they land, and receiving and returns use it every day, so the floor feels the difference immediately.
The second is anomaly alerts: stock running low, a SKU with no movement for days, a lot approaching expiry, a location where book and physical stock have disagreed for weeks. The system only alerts. It does not act. A wrong alert costs a glance. A right alert saves a supervisor from staring at reports. The advantage of this category is that it does not take anyone's job. It takes over part of their watching.
The third is small-scale forecast checking: comparing the system's predicted outbound volume for tomorrow against the manually built staffing plan, as a reference only, never as the schedule itself. When the forecast is right, the supervisor starts checking it voluntarily. When it is wrong, nobody ships the wrong order because of it. This category lets management see the system's judgment without having to pay for its mistakes.
Which of the three to choose depends on where the warehouse hurts most. If receiving is choked by paperwork, choose the first. If supervisors spend their days chasing reports, choose the second. If staffing and staging are still done by gut feel, choose the third. And here is the counter-intuitive judgment: the most valuable first pilot is often not the one that saves the most money. It is the one the floor most wants to keep using. A pilot that saves a lot but that nobody uses will cost you a second round of persuasion when you try to scale it. A pilot the floor asks to keep is already halfway scaled.
Judging the pilot in 30, 60, and 90 days
A first AI pilot should be judged in three stages, not in a single review after three months. At 30 days, ask whether it is usable: is the system stable, does the floor actually use it, and can people correct its mistakes quickly? If any of those three fails, pause and fix it rather than pushing harder. At 60 days, ask whether it is useful: compared with before the pilot, at least one of document entry time, anomaly detection time, or error count must have genuinely improved, by an amount you can state as a number, even if the number is modest. At 90 days, ask whether it can expand: can the same scenario run on another product family, another shift, or another warehouse without major rework? A pilot that cannot expand, however good its numbers look, is a showpiece, not a foundation.
Write the success criteria down before signing anything with a vendor, and write them as criteria a person can verify: what recognition accuracy rate, what share of alerts that turn out genuinely useful, how many times a day the floor actually uses the system. Numbers shown in a vendor demo need to be retested on your documents, your lighting, and your product mix. Accuracy in a demo environment usually drops in a real warehouse, and that is not deception. Real warehouses are simply messier than demos.
So here is the question to take back to your warehouse: if you could pick only one scenario for your first AI pilot, would you pick the one that looks good in the boss's report, or the one that gets the receiving team home half an hour earlier? In my experience, warehouses that pick the second get to have a second and third pilot. Warehouses that pick the first often find that the first pilot was also the last. Earn the floor's trust first, then show the boss the numbers. Do not reverse that order.
By Lao Mi, Supply Chain Veteran (20 yrs)







