I have spent a lot of time watching AI programs get planned for plants and warehouses. Roadmaps, steering decks, pilot lists. Plenty of activity. Very little that ends up running on the floor with someone accountable for whether it worked.
What I find interesting is how often the stall is not a model problem. The work never got narrow enough to ship.
What usually fills the calendar
A program starts broad. Someone collects use cases across sites. A vendor builds a proof of concept on a clean dataset. Leadership asks for a platform view. Safety detections get bundled in because the cameras are already there. Six months later the slide says "progress," and the dock still cannot tell you why trailers sat idle for three hours.
The people who own downtime, detention, yield, and labor minutes are still working from systems of record and shift reports. The AI work lives next to that, not inside it.
That gap is where programs die. Not in the model lab. On the handoff between a pilot and a production workflow.
A different bar
The bar that has held up for me is simpler than a transformation roadmap.
One workflow. One owner. A definition of done that a site GM or VP Ops would recognize as real.
Not "we detected anomalies." Done means the exception reaches the person who can act, within a window that matters on that shift, and the team can say whether the workflow reduced dwell, wait time, or a stop that used to show up only after the fact.
If you cannot name the owner and the done condition in one sentence, you do not have a production target yet. You have a research topic.
What that looks like on the floor
The first workflow is usually something ops already fights every week.
- When did dock work actually start and finish on this trailer?
- How long was a bay empty while appointments stacked?
- Where did freight sit in staging before putaway?
- Did the changeover follow the sequence the line expects?
- Which aisle blocked before the stop showed up in the KPI?
None of those sound like a breakthrough. They sound like questions a plant manager or warehouse lead already asks with incomplete answers.
Vision is often how the first signal arrives, because plants and DCs already paid for cameras. The camera stops being a footage library you scrub after an incident and starts looking more like a sensor for start/finish, queues, and sequence. That is a wedge, not the whole program.
Once that workflow holds, the same motion applies to the next one. Identify the high-impact ops work. Put a named owner on it. Ship it into production. Expand only after it holds.
Why roadmaps and pilots still win the calendar
Roadmaps feel responsible. Pilots feel safe. Both let a program stay in motion without forcing a production decision.
I have seen teams stack twelve use cases so nobody has to pick the first. I have seen pilots that never define who will run the alert on night shift. I have seen platform buys that assume every site needs the same stack on day one.
What changes when the bar is one shipping workflow is the conversation. You stop asking which AI initiative is furthest along on a Gantt chart. You ask which exception on this site has an owner, a done definition, and a path into the shift's actual work.
That is a harder question. It is also the one that separates a deck from something running on the floor.
The useful test
If you own a plant, a DC, or a network of sites, the useful test for any AI program is still small.
Can you point to one workflow in production, name who owns it, and say what "done" meant in operational terms?
If the answer is unclear, the stall usually started there, long before the model stopped improving.
That, to me, is the practical starting point for AI in operations. Not another roadmap. One workflow that ships.