The Job That Unlocks the Machine
In 1899, electric motors supplied less than five percent of the power in American factories. The motors worked. What was missing was the person who would redraw the floor. That job did not exist yet, because the decision it would own did not exist yet.
In 1899, electric motors supplied less than five percent of the power in American factories [1].
The motors worked. Owners bought them. Then they installed the new motor roughly where the steam engine had stood, hooked it to the same overhead shaft that ran the length of the building, and left the leather belts hanging where they had always hung. Output barely moved.
The surge came in the 1920s, roughly half a century late [2]. What arrived in between was not a better motor. It was a different floor plan. Once a small motor could sit on each machine, the building no longer had to be arranged around one spinning shaft. It could be arranged around the work [3].
Somebody had to see that and redraw the plant. That person was not the electrical engineer.
I keep the story close because it describes most of the AI programs I am shown. The model is good. The integration is finished. The agent runs. And the work around it is still laid out for the steam engine.
Two job titles are being hired against this gap right now. One is the Forward-Deployed Engineer, the FDE, which is older than it looks. The other is newer and stranger: the Human Systems Architect, the HSA. Both are worth understanding, and not because they are fashionable. They are the clearest evidence I have seen of something more useful than a forecast about jobs. They show the mechanism by which a technology actually creates work.
Where the value leaks
Ask why an AI initiative underdelivered and you usually get a technical answer. Wrong model, or data that was never as clean as the slide claimed.
Boston Consulting Group tells its clients to spend their effort differently: 10% on algorithms, 20% on technology and data, and the remaining 70% on people and process [4]. A consultancy recommending more change work is not a neutral witness, so I would not lean on that number as proof of anything. Treat it as a hypothesis and then check it against how these systems actually behave.
An agent inherits the process you hand it. If what you hand it is the version on the wall chart, it will run the wall chart, at speed and at scale, with no sense of when to stop and ask.
Maria Hatherell, who holds the first Human Systems Architect role at Kyndryl, put the gap plainly to IT Brew. "A lot of the processes that are on paper are not 100% the process that happens in real life," she said. People run extra checks that nobody has written down [5].
Anyone who has worked inside a company recognizes that instantly. The process document says seven steps. The person doing the job runs eleven, and four of the extra ones are the reason nothing has caught fire yet.
Here is the part I think gets missed. Until now, nobody had to write those four down.
A new colleague absorbs them by sitting near people for six months. They arrive in corridors and in the pause before someone clicks send. Don't auto-reply to that account, call Rita first. That transmission is slow and almost completely invisible, and it has worked fine for a century.
An agent cannot receive knowledge that way. It sits near no one. It gets a document, a set of tools, and a prompt. So the extraction that used to happen by accident now has to happen on purpose, and doing it on purpose is real work that was never anybody's job.
What a Forward-Deployed Engineer does
The FDE predates this wave. The title came out of Palantir about twenty years ago, on a simple bet: to build software for a hard environment, put the engineer inside that environment instead of shipping a spec back to headquarters.
Marty Cagan describes the point of it as customer discovery at scale [6]. The FDE embeds with a customer to learn the problem space firsthand, builds against real data while sitting there, and then carries the pattern home so the platform team can generalize it. That last step is what separates an FDE from a consultant. A consultant's output is the client's system. An FDE's output is also the product.
The FDE bends the system toward the world.
What a Human Systems Architect does
Kyndryl's role works the same gap from the other direction. Diana Wolfe, who leads AI research and strategy for its consulting arm, describes the HSA as the person who "designs the collaboration layer between people and AI agents as a system is being created" [7]. The final clause carries the argument. While it is being built, not after adoption has stalled.
Her diagnosis of what goes wrong is the sharpest sentence I have read on this subject in a year. "Organizations are stacking intelligence onto operational structures that were never designed to carry it, then wondering why things buckle" [7].
The distinction she draws next is the one I would put on a wall. "Policy as code defines what agents are allowed to do. HSAs define what agents should do, and how they work alongside the people who depend on them" [7].
Allowed and should are two different questions with two different owners. Your security and risk people can answer the first one. The second belongs to whoever actually understands the work.
In practice the job produces a short list that turns out to be very hard to write: which decisions the agent takes alone, which ones come to a person, where an escalation goes, and who carries it when the answer is wrong [7].
Hatherell calls these decision rights and autonomy boundaries [5]. I have written a book's worth of material about the agent side of that line, and the thing I underrated is how much design the human side needs, and how rarely anyone is assigned to do it.
So the pair sits like this. The FDE bends the system toward the world. The HSA bends the organization toward the system. Same seam, opposite sides.
Somebody just paid four billion dollars for this
In May, OpenAI launched a separate deployment company with four billion dollars of initial investment, and acquired Tomoro, a firm of roughly a hundred and fifty Forward-Deployed Engineers, to seed it [8]. About a week earlier, Anthropic announced a standalone enterprise services firm with Blackstone, Hellman & Friedman, and Goldman Sachs, with its own engineers embedded in the team [9].
Sit with who is doing this.
These are the two organizations on earth with the most reason to believe the model is enough. Their whole business case rests on capability. If anyone were positioned to say the technology takes it from here, it is them.
They put money next to the belief, and the money bought people. It did not buy more GPUs or a bigger sales force. It bought humans who go and sit inside your company and work out how the thing should be used.
Forecasts about the future of work are cheap. That was a purchase order.
The strongest objection, which is largely right
Now the case against everything above.
When IT Brew asked recruiters whether the HSA would be the next breakout AI role, JC Christian, president of the executive search firm Christian & Timbers, was blunt. "To me, [HSA] actually isn't that new. It sounds cool and I think it's a rebranding of what has already existed," he said. "It doesn't change the game that much" [5].
He is right, and I will go further on his behalf. Read the HSA description and you will not find one skill that was invented in 2026. Mapping how work really happens is what a Business Analyst does. Redesigning roles around a new system is Change Management. Clarifying who decides what is Organizational Design, which is about a century old.
Two more reasons to be skeptical. The role exists at one company in a handful of markets. And that company sells consulting by the day, which gives it an obvious commercial interest in a new discipline that only its people are trained for.
Those are three fair objections and I mean all of them.
Why the rebrand is the whole point
Go back to the factory.
The person who redrew that floor plan also brought no new skills to it. Laying out a plant was Mechanical Engineering, plus Cost Accounting, plus the patience to stand on a shop floor and watch what people did. Hugo Diemer put a name on the combination in 1900 and called it Industrial Engineering. Penn State opened the first program in the world in 1908 and made it a department the following year [10].
Every input was old. What was new is that the electric motor had created a decision.
As long as the drive shaft ran the building, where to put a machine was not a decision at all. It was geometry. Once every machine had its own motor, the question opened up, and it turned out to be worth an enormous amount of money, and it belonged to nobody. A job formed around it and eventually got a name.
That is how professions form. Rarely out of new skills. Usually out of old skills pointed at a decision that has just become scarce and valuable.
So Christian and I are describing one event from two sides. He is looking at the inputs and correctly seeing nothing new. I am looking at what those inputs are now pointed at.
And the new decision is easy to name. Where may this thing act without a person?
Nobody had to answer that before, because software did what it was told and then stopped. An agent runs a loop, takes actions in the world, and keeps going after you have stopped watching. The moment that is true of a system inside your company, someone has to decide how far it goes alone and who answers for it when it goes wrong. In most companies today nobody owns that question. It gets settled by default, by whoever configured the thing, on a Thursday afternoon.
A large, valuable, unowned decision. Jobs form around those.
Who actually gets these jobs
Here is the part I find genuinely encouraging, and I want to be careful not to oversell it.
These roles do not run on machine learning research. They run on knowing how the work really gets done here. Which exception matters and which one is theater. Which customer you never send an automated reply to, and what happened the last time somebody did. Why step four exists even though the process map makes it look redundant. Who to call when the system says yes and your instinct says no.
That is institutional knowledge, and it is exactly the asset a lot of people have been told is now depreciating.
The claims handler with twelve years of edge cases in her head is closer to being a Human Systems Architect than a new graduate holding a certificate in fine-tuning. She has to learn what agents can and cannot do, and she has to learn to write things down at a precision the job never asked of her before. That is a real gap and it does not close by itself. But it is a much shorter distance than the headlines imply, and it runs in the opposite direction to the one everyone is bracing for.
Nine lines that locate the job
If you want to test any of this without a budget, take one workflow where an agent is running or about to.
Write down three things the agent decides alone. Then three things that come to a person, and what triggers the handoff. Then the names of the individuals who are answerable when the output is wrong.
Nine lines. One afternoon.
If someone can produce them, your seam is designed and you can move faster than you think. If nobody in the building can, you have not found a documentation gap. You have found the job.
The honest limit
I am not claiming these two roles absorb the people whose work AI takes. I do not know that, nobody does, and the arithmetic is not obviously friendly. Two new job families do not offset a broad shift in demand, and anyone telling you the numbers work out is guessing.
The claim I am making is narrower and more useful. The way a technology creates work is not a slogan about freeing people up for higher-value tasks. It is a mechanism, it is visible right now with dates and funding attached, and it runs like this. A capability arrives. It opens a decision that used to be settled by physics or by habit. The decision turns out to be expensive. A role forms around owning it.
Watch the decisions, not the job boards. The boards are the lagging indicator.
Nobody in 1899 could have applied
Whether the title "Human Systems Architect" survives, I have no idea. It might be absorbed into a product role, or renamed by whoever wants the credit. Titles churn. The Forward-Deployed Engineer has been around twenty years and only just became fashionable.
That uncertainty does not weaken the argument. It is the argument.
In 1899 the Industrial Engineer did not exist. Not for lack of talent, and not for lack of skills, because every one of those skills was already in the building. The job did not exist because the decision it would own did not exist. The drive shaft answered it.
Then the shaft came down, the question opened, and inside a decade the answer had a name, a department, and a degree.
The shaft is coming down again. This time the question is where the machine goes alone. In your company, right now, that question is almost certainly being answered by accident.
Somebody is going to own it. It may as well be someone who understands the work.
Peter van Hees
References
- Warren D. Devine Jr., "From Shafts to Wires: Historical Perspective on Electrification," The Journal of Economic History 43(2), 1983, 347-372. Devine's horsepower series is summarized in EH.net's encyclopedia entry on the US economy in the 1920s: electricity supplied under 5% of manufacturing primary horsepower in 1899, 50% by 1919, and 75% by 1929. https://www.cambridge.org/core/journals/journal-of-economic-history/article/abs/from-shafts-to-wires-historical-perspective-on-electrification/500078D9B4764BA1109A7967437CF226 and https://eh.net/encyclopedia/the-u-s-economy-in-the-1920s/
- Tim Harford, "Why didn't electricity immediately change manufacturing?", BBC News, 21 August 2017, from the series 50 Things That Made the Modern Economy. https://www.bbc.com/news/business-40673694
- Paul A. David, "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox," American Economic Review 80(2), May 1990, 355-361. The original account of the group-drive to unit-drive transition and the lag between availability and measured productivity. https://econpapers.repec.org/RePEc:aea:aecrev:v:80:y:1990:i:2:p:355-61
- Joe McKendrick, "Why AI's 10-20-70 Principle Should Matter To CEOs And Everyone Else," Forbes, 26 January 2026, on BCG's recommended allocation. https://www.forbes.com/sites/joemckendrick/2026/01/26/why-ais-10-20-70-principle-should-matter-to-ceos-and-everyone-else/
- IT Brew, "Why Kyndryl introduced a new AI role," interviews with Maria Hatherell, JC Christian, and Gregory Summers. https://www.itbrew.com/stories/why-kyndryl-introduced-a-new-ai-role
- Marty Cagan, "Forward Deployed Engineers," Silicon Valley Product Group, 17 September 2025. https://www.svpg.com/forward-deployed-engineers/
- Diana Wolfe, "Why AI needs Human Systems Architects to scale," Kyndryl, 30 April 2026. https://www.kyndryl.com/au/en/about-us/news/2026/04/agentic-ai-human-systems-architect
- OpenAI, "OpenAI launches the Deployment Company to help businesses build around intelligence," 11 May 2026, and Larry Dignan, "OpenAI launches OpenAI Deployment Company, acquires Tomoro," Constellation Research, 11 May 2026. https://openai.com/index/openai-launches-the-deployment-company/ and https://www.constellationr.com/insights/news/openai-launches-openai-deployment-company-acquires-tomoro
- Anthropic, "Building a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs," May 2026. https://www.anthropic.com/news/enterprise-ai-services-company
- Penn State Department of Industrial and Manufacturing Engineering, "Department History." Hugo Diemer coined the term "industrial engineering" in 1900; the program was established in 1908 and the department in 1909. https://www.ime.psu.edu/department/history.aspx