AI + Robotics

In July 2025 Amazon delivered its millionth robot, to a fulfillment center in Japan. Its fleet now runs across more than 300 facilities. On the same day the company announced a new AI model for that fleet¹. Before reading on, make a prediction. What would you expect the AI to do? Many people guess something visible: robots that grasp odd-shaped parcels, or machines that walk the aisles like people.

The answer is less cinematic. The model, which Amazon calls DeepFleet, is a traffic manager. It coordinates how the robots move through a building, and Amazon says it improves the fleet’s travel time by 10 percent. The same announcement contained a second number that rarely makes the headlines. At its newest site in Shreveport, Louisiana, the advanced robotics need 30 percent more employees in reliability, maintenance and engineering roles¹.

Those two numbers carry this chapter’s argument. Much of the value of AI in physical work comes from how the whole workflow runs, not from a single impressive machine. And a body has to be maintained.

A wrong movement is not a wrong answer

Everything earlier in this module happened inside software. A model that drafts a wrong reply produces a sentence someone can reject. An agent that takes a wrong action in a system can often be reversed, which is why Agentic AI and Autonomous Actions built its controls around least privilege, approval and reversibility. A machine that moves wrongly can hurt someone, damage equipment or spoil a batch of product, and none of that can be undone with a click.

That difference sets the core idea. Robotics gives AI a body, and AI gives robots the flexibility to work in places that are not perfectly controlled. The opportunity for a leader is to find physical workflows where that combination creates measurable value at a level of safety and cost the organization can defend. The rest of the chapter is about how to judge those three conditions together.

Where AI Is Going Next named the physical world as the fourth dial of AI progress and showed it moving more slowly than the others, inside well-defined operating domains. This chapter looks inside that dial.

What AI adds to automation

Industrial robots are not new. Welding arms and pick-and-place machines have worked in car and electronics plants for decades, and the International Federation of Robotics now counts about 5 million of them in factories worldwide². Most of them follow programs. They are fast, precise and tireless as long as the world in front of them is known: the part arrives in the same place, in the same orientation, under the same light.

What AI adds is the ability to cope when the world is not known in advance. Vision models can recognize objects, defects, people and obstacles that vary. Planning can choose a path, a grasp or a sequence of steps that nobody scripted. Language can let a supervisor describe a task instead of programming it. Each of these widens the range of work a machine can take on. Much of this is still capability shown in laboratories and pilots. Documented results at scale, such as Amazon’s fleet routing, are fewer, and many of them are company-reported rather than independently measured.

It has taken a long time, and the reason has a name. In 1988 the roboticist Hans Moravec observed that it was comparatively easy to make computers perform like adults on intelligence tests or at checkers, and very hard to give them the perception and mobility of a one-year-old³. Moravec’s paradox still holds in a weaker form. Language models learned from a large share of the text ever written. There is no equivalent library of robot experience. When 21 research institutions pooled their robot data in 2023, the combined set held about a million real robot trajectories from 22 types of robot. A model trained on all of it did better than models trained on one robot’s data alone, which is encouraging, but the pool is tiny next to the text that trains a language model⁴. Physical skill is still expensive to learn because every example has to be collected in the physical world or simulated convincingly.

The loop that never ends

A language model answers once. A machine in the physical world runs a loop, many times a second, for as long as it is working.

The loop explains why physical AI is harder than digital AI in four specific ways. First, reliability is set by consequence. A system that completes 99 percent of tasks may be excellent at drafting emails and unacceptable at moving pallets near people, because the remaining 1 percent is a collision. From AI Assistants to AI Agents showed how small error rates compound over many steps; a robot runs thousands of steps an hour.

Second, the real world does not match the simulation. Teams train and test in simulation and on digital twins because it is cheap and safe, but friction, glare, dust, wear and people behave differently from their models. Every physical deployment needs validation in the real site, with the real mix of objects.

Third, some decisions cannot wait for the cloud. A machine that must stop for a person cannot depend on a network round trip. The practical split is to keep safety and real-time control on the machine, coordination on the site, and training, analytics and fleet updates in the cloud.

Fourth, a fleet is an operating problem. Many machines mean device management, model versions, telemetry, spare parts and incident handling. Changing the model can change how a machine moves, so a model update in physical AI deserves the testing that a software release would get in a safety-critical system.

AI proposes, the safety system decides

The most important architectural idea in this chapter is separation. A well-designed robot has layers. The AI layer perceives and proposes: there is a box, pick it up, take this path. Below it, a planner turns the proposal into movements. Below that sits a safety controller that decides whether the movement is permitted, and at the bottom the motion control that drives the motors.

The safety controller is deliberately simple and predictable. It enforces speed limits, keep-out zones and force limits, and when something fails, a sensor, the model, the network or the plan, its designed response is to stop or move to a safe state, not to improvise. The industrial robot safety standards were rewritten in 2025 for exactly this world. The third edition of ISO 10218, published in February 2025, says that only an application, not a robot, can be validated as collaborative, and adds cybersecurity requirements to the extent that they affect robot safety⁵⁶. A leader does not need to read the standards. A leader does need to ask who certified the safety layer, and whether anything in the AI layer can override it.

People stay in the loop by design, not as an afterthought. Driverless taxis show a useful pattern. When a Waymo vehicle meets something it cannot interpret, such as a closed lane, it can ask a remote human for context or a suggested path. The human never drives; the vehicle stays in control and decides how to use the advice⁷. The pattern generalizes: the machine attempts, uncertainty appears, a person advises, the machine continues. It lets a deployment start before full autonomy is proven, and every intervention becomes data for improving the system.

The regulation reads like the architecture above turned into law: a bounded movement space, a stop that always works, a record of safety decisions and a human who can correct the machine⁸.

Economics: cost per successful task

Where AI Is Going Next argued for judging AI by the cost of a completed task rather than the price of a model. For machines the same test applies, with heavier inputs. The full cost includes hardware, software, integration with existing systems, facility changes, safety assessment, energy, supervision, spare parts, calibration and repairs. The benefit includes throughput, quality, injuries avoided and work that can finally be staffed. AI does not remove physical maintenance, as Amazon’s 30 percent figure shows.

Two variables decide most business cases. The first is utilization. A robot that sits idle half the shift has roughly doubled its cost per task, and idle time comes from places no demo shows: exceptions, charging, weather, waiting for people and waiting for parts. The second is fit with the environment.

The matrix also puts humanoid robots in perspective. They attract attention because they could work in spaces built for people without rebuilding those spaces, and the International Federation of Robotics notes strong interest in using them in logistics and manufacturing⁹. Its view is that humanoids will complement existing robot types rather than replace them, and that if and when mass adoption happens remains uncertain¹⁰. A specialized machine built for one task in a controlled setting will often stay cheaper and more reliable. Judge any form, human-shaped or not, on reliability, safety, maintenance and throughput.

Story: the inspection drones that flew and fixed nothing

What follows is a composite, drawn from patterns common among wind-farm operators; it describes no single company.

A wind-energy company with hundreds of onshore turbines had a safety problem and a backlog. Sending rope-access technicians up a turbine to inspect its blades is slow and dangerous work, and storms left long queues of turbines waiting for a check. After an impressive demonstration, leadership bought a fleet of AI-enabled inspection drones. The vision model could flag leading-edge erosion, cracks and lightning damage from a single flight. The program was announced as a step toward autonomous maintenance.

A year later the post-mortem found that the drones had flown thousands of inspections and that repair times had barely moved. Four causes stood out.

The findings had nowhere to go. Defects landed as images in a shared folder. The work-order system could not accept them, so a planner retyped the important ones, and most were never acted on. Utilization was a fraction of the plan. Drones could not fly in high wind or rain, aviation rules on most sites required a trained pilot to keep the drone in sight, and crews spent much of the day driving between sites. A model update changed behavior. The supplier improved the defect model; false alarms rose for one blade type; nobody had a regression test, and field teams stopped trusting the flags. The metric was wrong. The program reported flights flown, which went up every month, while the outcome that mattered, verified defects repaired, was nobody’s target.

The rescoped program kept the drones. It narrowed the task to post-storm checks on one turbine model, wired the findings into work orders with a human reviewer confirming each flag, tested every model update against a set of known defects before it reached the fleet, and measured cost per verified defect closed. Climbs fell for that task, and the case for widening the scope was then made with evidence. The machine had not changed. The workflow around it had.

What this means for leaders

Start with the work, not the machine. Ask which physical tasks are expensive, dangerous, repetitive or hard to staff, and only then ask whether a machine can do them at acceptable safety and cost. Keep the safety layer outside the model and know who certifies it. Plan for people in the loop, in supervision, remote assistance and maintenance, and treat their interventions as data. Measure cost per successful task, with utilization in plain view. And from January 2027, treat any AI added to machinery in the EU as a product-safety question, not an IT project.

Check yourself

1. AI plus robotics mainly means humanoid robots.MythFact
2. In a well-designed robot, the AI layer has the final say over physical movement.MythFact
3. A 99 percent success rate can be unacceptable for a physical task.MythFact
4. Adding robots usually removes the need for maintenance staff.MythFact
5. Under the EU Machinery Regulation, safety components that learn need third-party conformity assessment.MythFact
6. Utilization can decide whether a robot’s business case works.MythFact

Reflection: the work that still has no body

What comes next

This chapter followed AI into machines on a floor, a field or a turbine. The larger change may be in the workflows those machines, agents and people share. The next chapter, Autonomous Workflows and AI-Native Organizations, asks what happens when organizations redesign their processes and operating models around AI rather than adding AI to the process they already have.