The Industry Just Discovered the Forward Deployed Engineer.

This week DeepLearning.AI's newsletter The Batch led with the buzziest new job title in Silicon Valley: the Forward Deployed Engineer. The idea is that an engineer gets embedded inside a client organization, sits with the people who actually do the work, and customizes and tunes the AI until it fits the messy specifics of that business. The model is spreading from the big AI labs to the rest of the industry, and it is being treated as a revelation.

If you build physical AI for real buildings, this is not a revelation. It is the only way the work has ever been done. You cannot ship a sump pump monitor from a data center. Somebody has to be in the basement.

Why the new job exists at all

The Forward Deployed Engineer exists because a model that scores well on a benchmark routinely falls apart on contact with a specific customer. The data is shaped differently than expected. The workflow has an undocumented exception that everyone in the building knows about and nobody wrote down. The thing the customer calls a "normal week" is not the thing the demo assumed. So the labs started sending engineers to live inside the customer and close that gap by hand.

That gap is exactly the one Capgemini measured this spring. In a survey of 1,678 executives at billion-dollar companies, 80% said they are already engaging with physical AI in some form, but only 4% said they are operating it at scale. Nearly two-thirds expect to cross from pilot to production within five years. Read that honestly and it says the same thing the Forward Deployed Engineer trend says: getting a physical AI system to actually run in one specific place is the hard part, and most organizations have not solved it.

A sump pump is a forward deployment

We have been running an edge AI system on a sump pump in a Watertown basement since 2024. Nothing about it shipped from the cloud. Someone physically mounted the sensor, watched a few weeks of real cycles, and learned what normal looked like for that pump in that pit with that water table. The baseline is not a generic model of sump pumps. It is a model of one pump, built in the place the pump lives.

That is forward deployment in the most literal sense. The intelligence was assembled on-site, against the actual equipment, by someone standing next to it. When the float sensor later drifted, the system caught it not because the model was clever in the abstract, but because it had been tuned to ground truth that only exists in that basement.

The same is true of the forty-device building we run in Northampton. Each device learned the normal behavior of the specific machine it watches: this compressor's draw at this load, this boiler's cycle frequency, this pipe joint's difference between condensation and a real leak. None of that came out of a box. All of it was deployed forward, into the building, and tuned there.

The frontier is paying enormous sums to avoid being on-site

The contrast this week was sharp. While The Batch was explaining the Forward Deployed Engineer, the World Intelligence Expo in Tianjin was showing the opposite bet. PaXini Technology released a ten-billion-scale embodied AI dataset and detailed its Super EID Factory, a 12,000 square meter facility with 150 standardized collection units built to manufacture roughly 200 million data sets a year. The thesis is that if you collect enough physical interaction data centrally, robots will generalize to the real world without anyone tuning them on-site.

That may well work for humanoids that have to operate in an open, unpredictable world. It is the right bet when the environment is too varied to instrument one site at a time. But a building is not an open world. A boiler room is the same boiler room next year. The cost of being on-site is small, and the payoff is a model that fits perfectly instead of generally. For small buildings, the data factory is solving a problem you do not have, at a price you cannot justify.

This is the small operator's structural advantage

Here is the part that should change how you sell. The entire industry just declared that the scarce, valuable skill is putting a competent human next to the customer's real equipment and tuning the system until it works. That is not a thing a small operator lacks. It is the only thing a small operator does.

A big vendor's whole problem is that forward deployment does not scale. Sending an engineer to live inside every customer is expensive, which is why most enterprise physical AI stalls at 4%. A local operator who already drives to the building, already knows the equipment, and already tunes each sensor on-site is not behind the frontier on this axis. On the axis that actually decides whether a deployment survives, the local operator is the frontier.

• Lead with on-site, not with the model. The differentiator is not whose algorithm is bigger. It is that someone tuned this system to this building, in this building. Say that out loud.
• Sell production, not pilots. "It has been running since 2024" beats any spec sheet, because 96% of the market cannot say it. Time-in-production is the proof the enterprise buyers are failing to generate.
• Treat the site visit as the product. The thing the labs are now hiring Forward Deployed Engineers to do is the thing you already do on every install. Price it as value, not as overhead.

The frontier spent a decade trying to make AI work from a distance and just rediscovered that someone has to show up. In physical AI for buildings, showing up was never optional. It is the whole job, and it is the reason the systems we deploy are still running while most of the market is still piloting.

Sources: DeepLearning.AI The Batch, "AI Forward Deployed Engineers" (2026-05-29); Capgemini Research Institute, "Physical AI: Taking human-robot collaboration to the next level" (survey of 1,678 executives, 2026-04-16; 80% engaging, 4% at scale); World Intelligence Expo 2026, Tianjin, PaXini Super EID Factory and ten-billion-scale embodied AI dataset (Xinhua / China.org.cn, May 2026); field deployments at The Intersecto Watertown sump pump site and Northampton 40-device building.