An AI agent on the warehouse floor: what six months taught me
Putting a language model into a warehouse sounds like a demo. Keeping it there for six months is a different exercise, and the lessons had almost nothing to do with the model.
Clean data is not a prerequisite. It is the project.
Before anything could decide anything, stock positions had to be true. Not approximately true — true. Every project that skips this ends with a very confident system making very confident mistakes.
Autonomy has levels, and you earn them
- Level one: it observes and reports
- Level two: it proposes, a person accepts
- Level three: it acts, a person is told
- Level four: it acts, and only exceptions surface
Every function moved up one level at a time, and only after the previous level had been boring for weeks. Boring is the signal.
What surprised me
The staff were not resistant. They were relieved — but only once it stopped asking them to check its work. A tool that needs supervising is a second job.
Trust is not won by being right. It is won by being right in the same way every time.
What I would do differently
Spend longer at level one. The temptation is to move fast because the demo already works. The demo is not the thing; six months of the same correct answer is the thing.
Working on this in your company?
Tell me what it does and I will tell you which parts can run themselves.