My work has put me on a lot of warehouse floors. What stands out is the judgment people build around automated systems. They know when something feels off before a dashboard catches up. They see a station backing up, or a process that works on paper about to fail in practice.
That experience has shaped how I think about adoption of artificial intelligence. Enterprise AI started in knowledge work, where a bad answer costs a rewrite. It is moving into physical operations now, where people work alongside intelligence that sequences orders and moves goods, and a bad answer costs a missed cutoff. The direction is fulfillment that runs with less and less manual intervention, and in parts of the building, none at all.
Lights-out fulfillment gets talked about as an engineering milestone. The harder part is adoption. An operator hands more decisions to a system only after the people running it have watched it make those decisions well on a bad day in November. Every increment of autonomy is bought with trust earned on the floor.
JUDGMENT IS ALREADY PART OF THE SYSTEM
Most AI adoption frameworks assume a laptop, a calendar, documents, and Slack threads. That makes sense, because the first wave happened there. Microsoft and LinkedIn’s 2024 Work Trend Index found 75% of knowledge workers already using AI at work. Anthropic’s Economic Index found Claude usage concentrated in computer and mathematical occupations, which made up 37.2% of conversations despite representing just 3.4% of U.S. workers.
AI behaves differently when its decisions have physical consequences. If an AI tool invents a detail in a draft email, someone usually catches it before it goes out. In a fulfillment center, a bad output is a missed cutoff, a misrouted tote, or a mispicked order. Once goods move, they have moved.
Here is the part the playbooks miss. Intelligence in fulfillment does not sit on a screen by itself. It works inside a system. It combines automation that stores and moves goods, software that sequences work and allocates stations, and data from every movement and every pick. Together they produce what a human operator buys—more capacity in the same footprint, continuity through peaks, lower cost-to-serve, and room to adapt.
Operator judgment weighs what the system recommends against what the equipment and the orders are doing. Accept it here, override it there, or slow something down before it creates a bigger problem downstream. Most knowledge workers still work with assistants that draft and suggest. On a fulfillment floor, the intelligence sits inside equipment that is already moving, and the call to let the plan run belongs to whoever is standing next to it.
AUTONOMY ARRIVES ONE DECISION AT A TIME
Nobody flips a building into autonomous operation. Intelligence takes on manual work step by step. That includes routing work, sequencing orders, allocating stations, and flagging an issue earlier than a person would have caught it. A vision model looks into a bin, identifies an item nobody trained it on, and works out how to handle it, with the workstation verifying the pick by barcode before the order moves on.
Every increment requires a choice of which decisions the system handles, which stay with a person, and what happens when the system gets stuck. Those choices get made by people who watched the last increment work. Adoption is the gate on autonomy, and the floor controls the gate.
The human role moves with it. The work does not disappear as execution gets automated. It shifts toward supervision, exception handling, and the judgment calls that sit above the system. The 2026 MHI and Deloitte annual industry report found supply chain leaders reassessing every aspect of their operations because of AI, with robotics and automation ranked second behind AI as the most disruptive technology in the sector.
The best deployments put human judgment where the data runs out. The system sees operational data at a scale no person can match, which is what catches problems early. The worker sees what has not reached the data yet, like a damaged pallet, a supplier who changed packaging, or a trailer running late.
GENERIC AI DOESN’T SURVIVE A WAREHOUSE FLOOR
Enterprise AI programs often start from the tool and go looking for work to apply it to. Physical operations force the opposite order. The work comes first. The data it generates comes second. And the intelligence is shaped by both. A general-purpose model can describe a warehouse. It cannot reach into a bin.
THE NEW ADOPTION PLAYBOOK
The adoption playbook should start with warehouse workers. Here are lessons I’ve learned from the warehouse floor:
1. Design for override. If the human in the loop is real, the loop must be usable. A worker who spots a problem cannot wait through three screens, because the tote is already moving.
2. Explain the recommendation. When work goes to a station nobody expected, the worker needs to see what the system was going for before deciding whether to let it run.
3. Treat edge cases as product feedback. They show up as a jammed tote, an odd-shaped item, or a supplier who changed packaging without telling anyone, and they show where the system needs to go next.
4. Measure adoption by operational outcomes rather than usage counts. Examples include capacity gained in the same footprint, continuity held through peaks, and lowered cost-to-serve.
5. Put warehouse workers in the design loop early. The people who know a workflow’s failure modes should shape the system before it ships.
Get these right and the system earns the next decision. Lights-out fulfillment arrives that way, as a growing list of decisions the floor no longer has to make.
THE PEOPLE NOT IN THE MEETING
Warehouse workers have spent years at the intersection of automation, data, and human judgment. They know technology earns trust through performance, and that judgment matters most when the world does something the system did not expect. They also play a critical role in determining how far automation can go.
Before anyone writes another AI adoption framework, talk to them.
Parth Joshi is chief product officer and president of AutoStore.