$10 An Hour, Behind Plexiglass

Last week, Agility Robotics' Digit started its first real job.

Not a pilot. Not a demo loop with a press tour. An actual operational deployment at a Schaeffler automotive parts factory in South Carolina, ferrying 25-pound bins from a stamping press to a conveyor belt, two four-hour shifts a day, behind a plexiglass barrier.

The Wall Street Journal got the numbers, and The Batch wrote them up this week. Read them slowly. They're the most useful sentences anyone in physical AI has published this year.

The robot is roughly at parity with the human it replaces. The displaced worker was promoted to a supervisory role. There's a plexiglass barrier because Digit can't yet detect nearby humans. It works one minute per cycle. It needs to recharge between shifts.

This is what the leading edge of humanoid deployment actually looks like in 2026.

And I want to be clear about something before I keep going: the Schaeffler deployment is genuinely remarkable. Two years ago a humanoid in a real factory doing real work was a marketing video, not a paystub. Crossing into operational economics on any task is a milestone, and Agility deserves the credit they're getting for it.

But if you operate a small commercial building, the lesson here isn't that humanoid robots are coming for your facility. The lesson is what physical AI actually costs to deploy productively today, and where the economics are forgiving.

One Robot, One Task, One Plexiglass Barrier

What does $10–25 per hour buy you at Schaeffler?

• One robot.
• One predefined task — moving 25-pound baskets from point A to point B.
• Eight hours of uptime per day, split across two shifts with a recharge break.
• An environment that has been pre-mapped by Agility engineers.
• A plexiglass enclosure because the robot can't yet share floor space with humans safely.
• A workflow specified as structured tasks, not joint-motor commands.

That's the deal. It's a real deal. It's also a narrow one.

The bet that Schaeffler is making — and that hundreds of factories will make in the next few years — is that this narrow deal compounds. More robots. More tasks. Eventually, no plexiglass. McKinsey thinks the world goes from ~200 humanoids in factories today to 5 million by 2040.

I think that's roughly right. I also think it's irrelevant to most building owners reading this, because the buildings I work with don't have stamping presses and conveyor belts. They have a boiler that runs continuously, a sump pit that needs to never overflow, an electrical panel that should not draw more than its rated load, and a maintenance contractor who does walk-throughs once a week.

None of those are humanoid-shaped problems. All of them are physical-state-observation problems. And physical-state observation has a much cheaper deployable form than a humanoid robot — one that's already running, today, at a price that small buildings can actually pay.

The Same Skill, Without the Legs

Here's the part I want to make plain. Look at what Agility Digit does at Schaeffler. Strip away the legs. What's left?

A camera. A vision-language model. A planner that decides what to do based on what it sees. A motor that acts on the plan. Logging.

The legs exist to move the camera and the gripper between locations. That mobility is what justifies the $10–25/hour cost — Digit can walk to wherever the next basket is. In a stamping operation where the basket location moves, that's worth paying for.

In a 40,000-square-foot community center, the things that need to be observed don't move. The boiler stays put. The sump pit stays put. The HVAC air handler doesn't migrate. The electrical panel is bolted to the wall.

Mounting a camera in front of each of them is cheaper than buying a robot to walk to them. It also runs 24 hours a day instead of 8, with no recharge break, and no plexiglass.

This isn't a hypothetical. It's the deployment I've been running for two years.

That side-by-side isn't a fair comparison in the strict sense — Digit is doing physical work, not just observation. But for the building owner trying to decide what "physical AI for my facility" actually means in 2026, it's the right comparison. Because the value most buildings need from physical AI right now is observation, not basket-carrying.

Spot Confirms It

The other big news this month makes the same point from a different direction.

Boston Dynamics rolled Google DeepMind's Gemini Robotics ER 1.6 into Spot's AIVI-Learning module. As of April 8, every Spot customer enrolled in AIVI-Learning has a robot that can read complex gauges, measure sight-glass fullness from 0–100%, do 5S compliance audits, count pallets, and detect standing liquid on the floor.

Read those capabilities again. Every one of them is vision. Reading a gauge is vision. Measuring fluid level in a sight glass is vision. Counting pallets is vision. Detecting a puddle is vision.

The reason Spot is the platform is because Spot can walk to where the gauge is. In an oil refinery or a substation or a 200,000-square-foot factory, that mobility is genuinely necessary — the inspection points are scattered, hazardous, or hard to wire. A $75,000 inspection robot is reasonable economics in those environments.

In a 40,000-square-foot community center, the inspection points are not scattered. Most of them are visible from a small number of fixed positions. The mobility you'd pay for in Spot is unnecessary.

The vision capability — the part that's actually new and interesting — runs just as well from a $35 Raspberry Pi with a USB camera looking at the same gauge.

What the Schaeffler Numbers Actually Tell SMBs

Three things, I think.

One. Physical AI has crossed the cost-parity line for the first time on a narrow task. That's a real inflection. The next two years are going to be a lot of news about humanoid deployments expanding into new tasks, new factories, new geographies. That news is real and the trend line is real.

Two. Cost parity does not mean cost reduction. The Schaeffler economics are essentially break-even with a human. The justification for the deployment was operational, not financial — pulling an entry-level worker into a supervisory role and getting predictable throughput on a task that previously rotated through hires. If your facility doesn't have an analogous "bottleneck task that's hard to staff and easy to specify," the humanoid path doesn't have an obvious entry point yet.

Three. Most of what physical AI is good for in a small commercial building isn't manipulation. It's observation. And the observation case is already deployable, today, at a price that doesn't require a Schaeffler-scale capital budget. A $99/month managed-monitoring system on a small building gives you continuous data on more equipment than a Digit gives Schaeffler on one stamping press.

Behind the Plexiglass, In Front of the Camera

The image I keep coming back to from the Schaeffler story is the plexiglass barrier. It's there because the robot can't yet share space with humans safely. Agility says they'll add human detection next year. That's a year of plexiglass.

The reason this image stays with me: most of the buildings I instrument never need plexiglass at all. The sensors don't move. The cameras don't move. They observe from where they're mounted. The humans walk past them, occasionally glance up to confirm a green LED, and continue their work.

That's what physical AI looks like when it's deployed in a building that operates at human scale instead of factory scale. Not a humanoid behind glass. A camera on a wall, a sensor in a panel, an inference call running every five minutes, and a phone notification when something looks wrong.

Two years of running this kind of system in a 40,000-square-foot community center has taught me that this is what "smart building" actually means in practice. Not the version with a humanoid in the boiler room. The version with twenty cameras and a hundred sensors, all running on edge hardware, all reporting into a system that can flag anomalies before a human walks past them.

The Schaeffler deployment is a milestone for one kind of physical AI. The fixed-sensor deployment that Intersecto runs is a different kind. Both are real. Only one of them fits in your building today.