For ten months of 2025, a Figure 02 robot loaded sheet-metal parts for welding in the body shop at BMW's plant in Spartanburg, South Carolina. Figure says the robot had 37 seconds to load parts inside an 84-second cycle, on 10-hour shifts, Monday to Friday.

Most conversations about robots start with the machine. We think the more useful place to start is the record the machine will need.

30,000+BMW X3s built with help from one Figure 02 robot during a 10-month pilot in 2025BMW Group, June 2026
65,000+hours of operation for Agility's Digit robots, with deployment commitments at nine customer facilitiesAgility Robotics SEC filing, June 2026
2026the year Tesla expects to start Optimus production, with the first units kept for trainingTesla Q2 2026 Update

1. The working deployments have narrow, tightly defined jobs

As of September 2026, a handful of humanoid robots are doing real work, though most companies label that work a pilot. Each job is narrow, repeated, and defined in detail.

At Spartanburg, BMW says the Figure 02 robot moved more than 90,000 components in about 1,250 operating hours and supported production of more than 30,000 X3s. BMW calls that a completed pilot.

In June 2026, BMW started a project with the newer Figure 03, which sorts parts into sequencing trolleys for the line. BMW says these robots add to the automation already in place, and they handle monotonous work and work that is hard on the body.

In 2024, GXO signed what Agility Robotics called the industry's first formal commercial deployment of humanoid robots. Digit robots move totes from collaborative robots onto conveyors at a SPANX facility, and GXO rents the robots as a service. Agility says its Agility Arc software directs the robots' work at GXO.

None of these robots walked into an undefined job and worked it out. Someone decided what the task was, what counted as done, and how it fit the line around it.

An AI employee needs that same clear start before it drafts a single cert.

2. A robot needs a record it can read

Our view: the companies that benefit first from humanoid robots will be readable operations. We call an operation readable when a machine can read every job, part, and handoff in its system, and a named person owns each step.

A robot sorting parts into a sequencing trolley must know which parts belong to each order and the sequence for those parts. An AI drafting a CoC needs that same system record for the heat number. If the sequence lives on a whiteboard, or in one supervisor's head, the robot has nothing to read.

This is why we say the groundwork for paperwork and the groundwork for robots are the same groundwork. A digital traveler, a receiving form with the heat number in its own field, and one record per job are what an AI employee works from today. They are also what a robot, or the software that directs it, would work from later.

3. A named owner decides what happens when a machine stops

Every machine stops sometimes. A robot drops a part, a sensor misreads, or an AI flags a heat number that does not match the receiving log. What matters is who gets the call and what they can see.

Our view: the rule we use for AI employees carries over. Each one has a named person who owns its work, a limited set of things it is allowed to touch, and a log of every step it took. When an AI flags a mismatch, the quality lead sees exactly what it read and where it stopped. A company that already runs its AI this way has the habits a robot program will need: clear owners, logged actions, and a person who decides.

What to do this week

  1. Pick one repetitive physical task on your shop floor, such as kitting or moving parts to the line, and write down where the instructions for it come from today: a system, a paper list, or a person's memory.
  2. For one job, check whether you can trace a part from its mill cert to the shipment in your records without asking anyone. Note every step where you had to ask.
  3. Name one owner for each handoff in that job (receiving, cut, weld, inspect, ship) and write the names on the traveler template.
  4. Hold any robot vendor meeting until steps 1 to 3 have clear answers. The records you build for this are the same ones an AI employee needs, so they pay off whether a robot ever arrives or not.

For the full argument, read Manufacturing's ChatGPT Moment Hasn't Happened Yet: Why AI Fits Your Paperwork and How to Prepare. To start the records a robot would need, read From Paper Traveler to Traceability: Why Digital Forms Come Before AI.

Questions people ask

Are humanoid robots working in factories yet?

A few are, mostly in pilots. BMW reports that a Figure 02 robot supported production of more than 30,000 X3s during a 10-month pilot in 2025, and Agility Robotics reports more than 65,000 hours of operation for its Digit robots. Most deployments are narrow, repeated tasks at large companies.

What should a mid-size manufacturer do about humanoid robots now?

Make the work digital, traceable, and owned. Keep one record per job, put each heat number in its own field, and give each handoff a named owner. Those records pay off with AI today, and a robot would need them later.

Terms in this piece

Robots as a service
renting robots for a monthly or hourly fee instead of buying them.
AI employee
an AI set up like a new hire, with its own login, one defined job, and a named person who reviews its work.
Readable operation
an operation where every job, part, and handoff is recorded in a system a machine can read, with a named owner for each step.
Traceability
the ability to follow any part back through every step of the job to the material and documents it came from.