Amazon deployed its millionth warehouse robot this year, and the more interesting number isn't the million — it's the 10%. That's the improvement in travel efficiency Amazon reports since DeepFleet, its fleet-coordination AI, started routing that many robots around its warehouses. A 10% gain sounds modest until you multiply it across a million robots and every shift they run. That's the kind of number that gets a budget approved without a debate.
Why physical AI is outscaling software agents
Most of the AI conversation over the last two years has been about agents that read, write, and reason inside a screen. Physical AI — AI that perceives and acts in the physical world, through robots, drones, and self-adjusting equipment — is scaling faster than a lot of those software pilots, and the reason isn't that the technology is more mature. It's that the outcome is unambiguous. Travel time in a warehouse aisle, defect rate on a line, inspection time on a pipeline — these are numbers a plant manager already tracks. There's no argument about whether the agent "really" helped.
Compare that to a customer-service agent pilot, where success is a mix of resolution rate, sentiment, and a dozen judgment calls about what counts as "resolved." Physical AI skips that ambiguity, and ambiguity is what kills budget approval.
What DeepFleet actually proves
DeepFleet isn't a single robot getting smarter — it's a coordination layer making decisions about how a million robots share aisles, avoid collisions, and route around congestion in real time. That's the part worth paying attention to if you're evaluating physical AI for your own operation: the value usually isn't in any individual machine, it's in the layer that coordinates many of them. A single smart robot is a nice tool. A fleet that reroutes itself around a jam before a human notices the jam is a different category of investment.
Where this shows up outside the warehouse
- Manufacturing floors — equipment that predicts its own maintenance needs and adjusts production parameters without someone flagging an anomaly first.
- Infrastructure inspection — drones running computer vision over bridges, pipelines, and power lines, catching defects faster and more consistently than a manual crew walking the same route.
- Retail and inventory — in-store robots doing shelf audits that feed straight into supply chain systems instead of waiting for a quarterly count.
None of these need warehouse-scale ambition to be worth piloting. They just need a task that's repetitive, measurable, and currently done by a person on a fixed schedule.
What IT actually has to build for this
Physical AI isn't a facilities decision that IT reviews after the fact — it changes what your infrastructure has to do.
Robots and drones generate and act on data in real time, which means the local network and edge compute have to be reliable, not best-effort. A dropped connection for a chatbot is an annoyance; a dropped connection for a robot navigating a warehouse aisle is a safety issue. The data these systems produce is also only valuable if it lands somewhere useful — a robot doing shelf audits that dumps its findings into a silo nobody checks isn't delivering the ROI the pilot promised. And the attack surface changes: an autonomous robot or drone is a new kind of endpoint, and most security teams' threat models were built around laptops and servers, not machines that move.
Where to start if you're not Amazon
You don't need a million robots to test this. Pick one task that's repetitive, currently manual, and has a baseline you can measure before you touch it — inventory counts, routine inspections, predictable material movement. Measure the baseline carefully, because the entire case for physical AI rests on being able to point at a number and say "that's what changed." Treat it like any other capital investment with a payback period, not like an AI experiment you're hoping pans out.
The organizations getting real value out of physical AI didn't start with the hardest problem — they started with the easiest one to measure, proved the number, and only then went looking for the next task worth automating.