I have spent a lot of time sitting with plant and warehouse teams trying to pick where AI should start. The conversation often opens on platforms, model choices, or a long list of use cases. What I find interesting is how rarely it opens on the work that already fails in the open.
The first workflow that can ship is almost never the most sophisticated one. It is the one ops already names without a workshop.
Exceptions that already have an owner
Every site I have spent time in has a short list of exceptions that eat a shift.
A trailer arrives and dock work does not start when it should. A bay sits empty while appointments stack. Freight clears the dock and then waits in staging. A line stops and nobody can say, from the systems of record alone, what blocked the aisle twenty minutes earlier.
Those are not research topics. Someone already owns the outcome. Detention. Dwell. Labor minutes. Throughput on that shift.
If you are looking for a first workflow, start there. Not with a blank canvas of AI ideas. With the exception that already has a name, a cost, and a person who gets asked about it in the daily standup.
Bottlenecks that form before the KPI turns red
What stalls programs is often that the KPI arrives too late.
By the time the stop shows up in the report, the queue has already formed. People waiting. Pallets waiting. Equipment waiting. The floor knew something was wrong long before the dashboard did.
A shipping workflow is one that catches that earlier signal and puts it in front of the person who can act while the shift still has time.
- When did work on this trailer actually start and finish?
- How long was the bay empty while the next appointment waited?
- Where did freight sit between dock and putaway?
- Did the changeover follow the sequence the line expects?
- Which aisle blocked before the stop hit the KPI?
None of those questions need a science project. They need a reliable answer in the window the floor can still use.
Handoffs that lose an hour
Shift change is another place I keep seeing the same pattern.
One crew knows why a bay ran slow, why a staging lane filled, why a sequence slipped. The next crew inherits the floor, not the story. An hour disappears into relearning what the previous shift already saw.
A first workflow can be as plain as making start/finish, queue location, and sequence visible across that handoff. Not a new dashboard for the office. Something the incoming lead can trust without a debrief that never quite happens.
That is the kind of work that survives contact with night shift. It has an owner. It has a done condition. It changes what the next crew does in the first thirty minutes.
How the signal often arrives
Vision shows up here for a practical reason, not a product reason.
Plants and warehouses already have cameras. Most of that video is still treated like a footage library. Someone goes back and watches after something has gone wrong.
With vision models, that same camera starts looking a little more like a sensor. You can pull structured operational data out of ordinary video: when an activity started and finished, how long equipment waited, where a queue formed, whether a process followed the expected sequence, what happened in the minutes before a stop.
That is often how the first signal arrives for exceptions, bottlenecks, and handoffs. The hardware is already paid for and mounted. The wedge is turning that video into operational data the floor can act on.
It is not the whole AI program. It is a practical way to get one workflow into production without waiting for a new sensor stack.
What "can ship" actually means
I use a short test.
Can you name the exception, bottleneck, or handoff in one sentence? Can you name who owns it on that site? Can you say what done looks like in operational terms: shorter dwell, fewer empty-bay minutes, a sequence that holds, a handoff that does not lose the first hour?
If those answers are clear, you have a candidate that can ship.
If they are not, you probably still have a roadmap item. Interesting. Not ready for the floor.
The first workflow that can ship is usually hiding in plain sight. It is the work the plant manager, site GM, or warehouse lead already fights every week.
Pick that one. Put an owner on it. Get a signal into the shift. Expand only after it holds.
That, to me, is the practical path for ops AI. Not a bigger list. One workflow the floor would recognize as real.