Field service
Robots in field service: they work where someone already structured the job
Can robots replace field service technicians? Anthropic's robot exposure data says robots work in structured sites. Field work isn't one. Structure the data.
Can robots replace field service technicians? Anthropic just put numbers on the question behind robots in field service. On 30 September 2026, Anthropic researchers Russell Legate-Yang and Maxim Massenkoff published a robot exposure study. They took the roughly 19,000 job tasks in the O*NET database, picked out 7,594 physical ones, and rated each against robots that exist today. The headline is big: robots can already perform 74% of physical tasks in the US. The detail underneath is the part that matters for anyone who runs technicians. Almost all of that exposure sits in places where someone already built the environment around the machine.
TL;DR: Robots in field service won’t replace technicians soon, because robots work where the environment has been structured for them and a customer’s site never is. Anthropic’s study (Legate-Yang and Massenkoff, 30 September 2026) finds robots can do 74% of US physical tasks in some setting, but only 2% of physical tasks in unstructured environments, and they are cost-competitive for 0.3% of all job tasks today. It lists installation and repair among the occupation groups with relatively low exposure. The lesson for field service runs the other way: the automation that pays now is in the information around the visit, and it needs the same thing the robot needs. A structured environment. For software, that means structured job data.
Robots work in rooms someone redesigned
The study’s rubric is the useful part, more than the headline percentages. Every physical task gets one of four ratings. E0: no robot can do it. E1: a robot can do it in a purpose-built robotic environment, like an assembly line. E2: in a structured human workplace, like a logistics warehouse. E3: in an unstructured environment, like a city road.
Weighted by working time across the whole economy, the split looks like this:
| Exposure level (Anthropic, Sep 2026) | Share of all US working time | Example the study gives | What it means for field service |
|---|---|---|---|
| Cognitive and interpersonal work (not rated for robots) | 54% | Teaching classes, designing systems | Customer calls, diagnosis, quoting. LLM territory, not robot territory |
| E0: no robot can do it | ~12% | Providing first aid, taking people into custody | Most hands-on repair in a crawlspace or a vault |
| E1: purpose-built environment | ~23% | Assembling electronics, dispensing medications | The depot bench, not the truck |
| E2: structured human facility | ~10% | Moving warehouse goods, cleaning floors | Your parts warehouse |
| E3: unstructured environment | ~1% | Driving, regulating field irrigation, inspecting property from the air | Drones and inspection rigs on a site |
Look at where the mass is. E1 and E2 together are a third of all working time. E3 is one percent. The study says it directly: most robots need highly structured environments, and automating an E1 task could require redesigning the workplace to suit the robot.
That is the whole story of automation, physical or digital. The machine works after someone has made the space legible to it. Somebody laid out the warehouse aisles, standardized the totes, put fiducials on the floor. The robot gets the credit. The layout did the work.
Your technician works in E3 by definition
A field technician doesn’t control the room. The boiler is where the builder put it in 1987. The panel is behind a water heater. The meter vault has standing water in it. Every site is a one-off, and the job is mostly handling: connectors, fittings, wires, valves.
The study is blunt about this kind of work. It says present-day robots can do little of general repair work even in highly controlled environments, and it groups installation and repair with healthcare support as relatively low exposure. It expects taxi drivers and warehouse packers to see change sooner than nurses and mechanics. Capability limits block about 70% of physical tasks, and half of physical tasks would need better manipulation before robots could take them at scale. The study’s own example of the gap is the dexterity needed to untangle wires.
My favorite example from the paper is a field task. “Dig trenches” splits in two. Cutting a straight trench in open ground is rated E3, because there’s a retrofit kit that lets a hydraulic excavator do it autonomously. Careful digging by hand around existing pipes, cables and conduit is rated E0. More of the real work looks like the second one, so the task as a whole comes out unexposed. Anyone who has worked near a utility locate knows why.
Then there’s cost. Robots are cost-competitive for 0.3% of job tasks today. Costs would have to fall about 70% for that share to reach 10%, which the authors put at around 40 years at the historical price decline of roughly 3% a year. And their cost estimates include integration, installation, maintenance and human oversight, often including human supervisory, exception-handling or repair work. The robot creates a field service job of its own.
The newest field humanoid still has a person’s hands in it
On 6 October 2026, Minerva Humanoids came out of stealth with about $10 million in pre-seed funding led by General Catalyst. Its humanoid, Roger, is aimed squarely at unstructured industrial sites: offshore rigs, high-voltage power plants, bomb disposal. According to The Next Web’s report, trained specialists operate it through a VR headset and do the fine manipulation remotely. The AI handles balance, fall recovery and navigation.
That’s a sensible product. It’s also the study’s finding in hardware form. In E3, the hands are still human. The AI carries the body to the work and keeps it upright. The dexterous part, the part that is most of a technician’s job, still comes from a person, just from further away.
The environment you can structure is the information one
Here’s the useful reframe. A field service business can’t redesign its customers’ basements. It can redesign everything around the visit.
Dispatch, scheduling, parts, work orders, asset history, completion reports, invoicing. That is an environment too, and in most field operations it’s an E3 environment for software. Job types typed free-form. Asset history in a technician’s head. Parts on the truck tracked on a whiteboard. Completion notes in a PDF that someone rekeys into billing on Thursday. An optimizer dropped on that has the same problem a robot has in a crawlspace. It can’t see the room.
I’ve seen what changes when that layer gets structured. At HomeWorks Energy, an energy-services field operation, the automation that saved 71+ hours a week wasn’t anything on a truck. It parsed PDFs that people had been working through by hand. Next to it sat an integration between Salesforce, ServiceTitan and NetSuite, so the three systems agreed on what a job was. Later I built a constraint-based scheduler for a utility that reached 96% workday efficiency across more than a dozen interacting constraints. The solver was the cheap part. The constraint model was the work. The same data problem decides whether autonomous dispatch helps or hurts, and it’s why off-the-shelf scheduling tools stall on the rules a template can’t express.
The study even points at this path in its closing section. The authors say they didn’t model how AI could affect physical work without robots, for example by better predicting when factory machines need maintenance. That’s the field service opportunity, and it needs no new hardware.
What a structured field environment looks like
Do for your data what a warehouse does for a robot. Concretely:
- One job taxonomy, shared by dispatch, the technician’s app and billing, so a “no heat” call means the same thing in all three.
- An asset record per site with model, install date and service history, written by the technician at close-out, not reconstructed later.
- Parts tied to trucks and to jobs, so the scheduler knows whether the tech can finish on the first visit.
- Completion captured as fields, not a PDF. If someone rekeys it, it isn’t structured yet.
None of this is AI. All of it is what an AI dispatcher, a parts-forecasting model or a future robot vendor will ask you for first. Same technician. Different environment around the work.
The operator read
If a robot does arrive at your sites, it will arrive the way it arrived in warehouses: into the parts of the operation someone already structured. The depot bench before the crawlspace. The inspection before the repair. The operators who get value from it will be the ones whose records already say what’s on site, what it is, and what was done to it last time. That’s the same work that makes dispatch and scheduling automation pay today, which is why I start there. If your dispatch board still runs on a whiteboard and good memory, that’s the place to start the conversation.
FAQ
- Can robots replace field service technicians?
- Not on any timeline a field service business plans around. Anthropic's robot exposure study (Russell Legate-Yang and Maxim Massenkoff, 30 September 2026) rated 7,594 physical job tasks, drawn from O*NET's roughly 19,000, against the robots that exist today. It found robots can perform 74% of physical tasks in the US in some setting, but mostly in purpose-built or structured facilities, and just 2% of physical tasks in unstructured environments. It names installation and repair among the occupation groups with relatively low exposure, and says present-day robots can do little of general repair work even in highly controlled environments. A technician's day happens in someone else's basement, roof or vault, which is the unstructured case by definition. The parts of field service that are close to automation are the information around the visit: dispatch, parts, records, reporting.
- What is a robot exposure index?
- It is a measure of how much of an occupation's work a currently available robot can do, and in what kind of environment. Anthropic's version rates each physical task on four levels: E0, no robot can do it; E1, a robot can do it in a purpose-built robotic environment such as an assembly line; E2, in a structured human facility such as a logistics warehouse; E3, in an unstructured environment such as a city road. Tasks are weighted by the time workers spend on them and by employment. The index measures what robots can do today, not forecasts, so it shows where change arrives first: the study expects taxi drivers and warehouse packers to see changes sooner than nurses and mechanics.
- How much of field service work can robots do today?
- Very little of the hands-on part, and almost none of it at a competitive cost. Across the whole US economy the study puts robots at 34% of all working hours in some setting, but cost-competitive with human labor for only 0.3% of job tasks. Capabilities block about 70% of physical tasks, and half of physical tasks would need better manipulation, the dexterity to untangle wires, before robots could take them at scale. Field repair is dense with exactly that kind of handling. Robots in field service today mostly mean inspection platforms, drones, and teleoperated machines for hazardous sites, with a trained person still in control.
- Robots vs software automation in field service: which pays first?
- Software, by a wide margin, because the environment it needs is one you can build yourself. A robot needs a structured physical site, and you don't own your customers' sites. A dispatch engine, a parts system or a work-order flow needs a structured information environment: consistent job codes, an asset history per site, parts tied to trucks, completion records captured as data rather than PDFs. That structure is within your control and it is where field service automation pays now. It is also the prerequisite for any robot that shows up later.
- How do you prepare a field service operation for automation?
- Do for your data what a factory does for a robot. Pick one decision, such as which technician takes which job, and list every input it depends on: skills and certifications, location, parts on the truck, the asset's service history, the customer window. For each input, name the source system, reconcile it where systems disagree, and make sure the technician captures it in a structured form at the end of the visit. That turns the job from an unstructured environment into a structured one for software. Then automate the decision.
- What did Anthropic's robot study find about cost?
- That cost, not capability, is the binding constraint for most tasks robots can already do. Robots are cost-competitive for 0.3% of job tasks today. Costs would need to fall about 70% for that share to reach 10%, which takes around 40 years at the roughly 3% annual price decline robots have followed since the 1990s. The study's cost estimates include integration, installation, maintenance, energy and human oversight, and note that they often include human supervisory, exception-handling or repair work. Packers and packagers are the clearest current case: a set of robots costing over $2 million to buy and install replaces about 14 workers, at around $45,000 a year per worker replaced versus about $49,000 for the human.
- What is Minerva Humanoids' robot, and does it change this?
- Minerva Humanoids came out of stealth on 6 October 2026 with about $10 million in pre-seed funding led by General Catalyst, for a humanoid called Roger aimed at offshore oil rigs, high-voltage power plants and bomb disposal. As The Next Web reported, trained specialists operate it through a VR headset and do the fine manipulation remotely, while the company's AI handles balance, fall recovery and navigation. It supports the study's finding rather than contradicting it: the newest humanoid built for unstructured industrial sites keeps a person's hands in the loop for the dexterous work.