The Diagnosis That Costs a Million Dollars

There is a job title that OpenAI and Anthropic are currently bidding against each other to fill, alongside enterprises in insurance, fintech, healthcare and gaming. At the frontier labs it pays $385,000 at mid-level, $610,000 at staff, and between one and one-point-two million at principal - with equity running sixty to seventy percent of the package, which means the labs are not hiring these people so much as buying them. A hiring study published this summer projects demand growing 2,100% by year's end.

The job is called forward deployed engineer. Here is what one actually does, in the words of the company that invented the role: decide what the model should do, what data it needs to touch, and - before a single line of production code ships -how success will be measured.

Read that sentence again, slowly, and notice what is missing from it. It does not say "write the integration". It says decide what to build, against what data, checked against what definition of working. That is not a coding task with a consulting flavour.

That is a diagnosis. And the market has just put a price on it.

The title is not new. The panic is

Palantir built the role somewhere between 2005 and 2010, to solve a problem its first customers (the CIA, the NSA, army intelligence) could not solve with ordinary consultants. The phrase is borrowed from military usage: a forward-deployed unit sits close to the action rather than at a rear base. Internally Palantir called them Deltas, and until 2016 the company employed more of them than it did software engineers.

So the discipline has been sitting there, named, modelled, and demonstrably profitable, for roughly two decades. What is a year old is not the job. It is the willingness to pay for it.

The failure the job exists to prevent

The study is admirably specific about why the role exists at all: most AI pilots inside companies never reach production, because the in-house engineers, once handed the system, build what was requested rather than what solves the actual problem. Read requested for activated and problem for the ideas nobody in the room happened to be holding, and this is a company-sized version of a much older failure: a mind, or an organisation standing in for one, constructing its most coherent possible story out of whatever evidence is already in front of it, and never noticing that the missing evidence was the kind it had no way of knowing it lacked.

A forward deployed engineer's first task is to go looking for that missing evidence before the build starts. Which is a description of a pre-mortem. Not a metaphor for one. A functional match. This site named the technique in its first article: imagine the deployment has already failed, and write down why, before it happens instead of after. Gary Klein costed the exercise in 2007 at twenty to thirty minutes, in a room, for free. We have since claimed ten, on the grounds that most of what a pre-mortem surfaces is not subtle once somebody is made to look for it on purpose.

Seventeen thousand, and two thousand

Here is the number that matters, and it is not the one that made the headlines.

Christian & Timbers put roughly seventeen thousand forward deployed engineers on the US market today (a good share of them already at Palantir). Of those seventeen thousand, about two thousand are judged capable of consistently delivering what the study's author calls real return, measured in multiple tens of millions of dollars of impact. The study is blunt about the distinction: "Not 2,000 available. 2,000 total."

Sit with the ratio rather than the headline. Seventeen thousand people hold the title. Seventeen thousand have the credential, the five-plus years, the sector exposure, the CV that gets them into the room. The market trusts about one in eight of them.

Whatever separates the two thousand from the other fifteen, it cannot be the qualification, because the qualification is exactly what all seventeen thousand share. It is not access to the technique either; Klein published it in the Harvard Business Review and it fits on two pages. The only remaining variable is whether the person has, repeatedly, actually done it: sat down before the budget was committed and insisted on establishing what would count as failure.

That is a habit, not a skill. And the market has just priced it at a million dollars, which is an unusual thing to pay for something available to anyone willing to hold a meeting.

The demand curve underneath this is steeper than the 2,100% projection suggests, because projections are cheap and this one has a floor under it. At the start of 2026, five to ten percent of companies were planning to hire an FDE, mostly for small pilots. By the end of the second quarter it was seventy percent, with the largest consultancies reporting a need to multiply their FDE headcount tenfold and build standing teams of twenty to a hundred. Enterprises were hiring in the middle of summer, which the study's author describes as something he has never seen.

What four billion dollars buys that the room doesn't

Neither lab stopped at hiring individuals. OpenAI opened the OpenAI Deployment Company on 11 May 2026 with more than four billion dollars from nineteen investors at a fourteen-billion-dollar valuation, majority-owned and controlled by OpenAI. Ten days later, Anthropic made its own move by acquiring Fractional AI, the applied-AI services firm that would, two months on, be relaunched as the branded entity Ode with Anthropic, backed by Blackstone and Hellman & Friedman at a $1.5 billion valuation. Ten days apart in May, under two different names, at two different prices - and by July both labs were selling the same diagnosis at scale.

Read the investor list. Alongside TPG, Advent, Bain Capital, Brookfield, Goldman Sachs and SoftBank sit Bain & Company, McKinsey & Company and Capgemini - three firms that have spent the better part of a century selling exactly this kind of judgment by the hour, and that are named in the same announcement as delivery partners.

They did not turn up as competitors offering the cheaper alternative. They turned up as co-funders of the more expensive one. The consulting industry's response to a question it has billed for decades was not to underprice the room. It was to buy equity in the pipeline built to replace its own invoices - which is, on reflection, the correct move, and tells you what those firms privately think the room was ever worth.

This is the regulator's gift running through a different channel. Nobody is excluded by a rule here. The thing being made expensive is judgment, and the parties who can already afford to supply it have bought the toll booth.

There is a smaller detail that makes the point more cheaply than four billion dollars does. Some clients, the study's author reports, are now buying Palantir's software specifically in order to get access to Palantir's engineers. Not for the product. For the people who will ask what the product is supposed to be for.

The Clause holds no strong feelings about forward deployed engineers specifically (it rarely does about a job title - any job title) but it is entirely at home in a market that has just assigned a genuine scarcity price to an exercise with no natural scarcity in it at all. It has furnished boardrooms before where the actual constraint was never expertise. It was willingness.

The objection, which is a good one

Chris Taylor, who runs Ode with Anthropic, said something that cuts against all of this, and it should be met rather than skirted: "Many FDEs are well equipped to help you roll Claude Code out to your workforce. Very few are capable of building your flagship AI product feature".

That is a claim about scarce building capability, not scarce asking. If Taylor is right, the two thousand are rare for an entirely ordinary reason - they are unusually good engineers - and the premium is a normal premium for a hard skill, with no paradox in it anywhere.

He is very likely right about the second phase of the job, and the two thousand almost certainly are unusually good engineers. But notice that his sentence describes what happens after somebody has decided what the flagship feature should be. The job description at the top of this piece does not lead with building. It leads with deciding what to build, against what data, measured how, before production code ships. Everything Taylor describes is downstream of that, and cannot begin until it is settled.

So both are scarce. The difference is that one of them has a floor price set by fifteen years of engineering experience, and the other has a floor price of an afternoon - and the market, unable to buy the afternoon separately, has been obliged to purchase it bundled with the engineer.

The part the report tells against itself

One more detail, near the bottom of the coverage, worth sitting with. Jeff Christian, whose firm produced the 2,100% figure and whose business depends on it, volunteers that the demand may not last. "Maybe in two years, everything's automated, and agents are automating agents as opposed to humans automating agents". In the medium term he expects the need to shift from enterprise AI to physical AI, as companies try to fold humanoid robots into their workflows. Within five or ten years, he says, the role may simply go away.

A recruiter talking down the half-life of the role he is recruiting for is not a man overselling his market. Take him at his word and the shape of what is being purchased comes into focus. Not a durable profession. A temporary and very expensive workaround for the absence of a habit; priced as though it required five years of production engineering to acquire, because for now that is the only credential the market has agreed to accept as proof that somebody will actually do it.

Seventeen thousand people have the credential. Two thousand are trusted with the judgment. The same study cites successful deployments returning tens of millions of dollars once the diagnosis is finally done properly - which is also, read the other way, the sum already spent by the time anyone thought to ask.

Klein's price for asking first was never in that range. It still isn't. The market has simply found two thousand people it trusts to ask, and is now charging a premium for something that was, and remains, free to anyone willing to hold the meeting before the invoice.