Underneath the model launches, the AI labs have started selling labor wrapped in software.
Watch where their hiring budget is moving. The marquee research roles are still on the page, but the title growing fastest is forward-deployed engineer — staffers shipped into customer offices to make the product actually work. One analysis of more than a thousand postings counted a 1,165 percent year-over-year jump in roles titled "forward deployed engineer" between January and October 2025; an independent read of Indeed listings logged a 729 percent jump comparing this April to last. OpenAI's own deployment page has gone from roughly two of these engineers in early 2025 to more than ten across eight cities and three continents. Anthropic ships the same function under "Applied AI." Google has signaled plans to hire hundreds. The labs call the motion different things. The shape is the same: send a person to do what the system cannot yet do alone.
We have seen this picture before.
In 2003, a young analytics company was founded on a product that could not survive on product alone. Its data platform was real, but it could not cross the chasm into the agencies it was trying to sell — too much ambient context, too many entrenched workflows, too much judgment that lived in people's heads rather than in fields. Its answer, worked out over the years that followed, was to invent the forward-deployed engineer: a salaried staffer dispatched to live inside the customer for months, write integrations against the customer's actual data, and back-fill what the product could not infer on its own. Two decades later, that motion is the dominant enterprise playbook of the AI labs. The names are different. The shape is identical.
The first read is the business-model read
The simplest explanation is that the labs are recycling the late-stage SaaS playbook. Software margins are getting thinner — inference is metered, training is capitalized, gross margin is fighting compute. Services revenue solves a near-term problem: it lets the labs report dollars per customer that look like enterprise ARR while the product surface catches up.
This is not a new move. The cloud era already taught us how it resolves. The category-defining enterprise software companies of the last three decades have grown services arms — ERP platforms leaned on integrators, CRM systems generated implementation partners, even the most product-led infrastructure businesses ran professional services for their top accounts. Services revenue is what a company books when the product cannot carry the bill on its own. It is not a tell about the customer. It is a tell about the surface.
When we read the FDE expansion as a financial maneuver, the question becomes the one nobody is asking out loud: how much of the AI-lab top line is product, and how much is a body that was sent to make the product work? The number matters.
The second read is the product read
The deeper read, and the one the labs lean on internally, is that enterprise adoption is broken. The reasoning runs: the models are good, but customers cannot drive transformation on their own. Workflows are messy, data is unlabeled, internal champions burn out. Forward-deployed engineering is framed as a bridge — a way to meet the customer where they are, ship outcomes, and unlock the next contract.
This reading is true. It is also a description of a product failure phrased as a customer failure.
A product that compiles expertise into its own surface does not need a person sent on a plane to install it. A product with Embedded Judgment shortens the distance between the buyer and the outcome — the workflow is the product, not a thing the product orchestrates. The MIT NANDA "State of AI in Business" report, published in mid-2025, found that the overwhelming majority of enterprise generative-AI pilots failed to produce measurable P&L impact within their planning horizon, even where adoption metrics looked healthy. Adoption was not the bottleneck. Translation was. The models could draft, summarize, route — but the customer's actual question was, "what should we do?", and the system could not answer in a way the operator could act on.
Forward-deployed engineering closes that gap by hand. Each FDE is, in effect, a temporary Standards Gap Owner — an employee sent in to notice the difference between observed and desired behavior and build the bridge between them. That is real, valuable work. It is also evidence that the product, on its own, does not yet do it.
The third read is the one nobody wants to underwrite
If the FDE expansion were only a margin tactic, services revenue would carry the story. If it were only an adoption bridge, the metric would be referenceable customers and renewal rates. The reason the motion has expanded across the major labs in the same window is that the third number — the productivity number that justifies the trillion-dollar capex story — has not arrived on schedule.
When the seller measures effort instead of outcome, the buyer is paying for motion, not change.
Goldman's mid-2024 research note questioning whether generative-AI spend would translate into proportionate returns was not a heterodox view at the time it landed. The macroeconomics literature — most pointedly the MIT working paper arguing that the realistic productivity uplift from AI in this decade would be measured in tenths of a percent of total factor productivity, not the multiples the spend implied — was pointing in the same direction from the other side. The labs, the customers, and the macro readers were each seeing a different face of the same gap.
This is where token usage theater enters. Token consumption is the metric the labs most love to disclose, because it grows monotonically, looks like usage, and rhymes with the unit economics of the cloud businesses that came before. A token, though, is an input. It is a measure of how much the model was asked to do. It is not a measure of what was changed in the customer's business by what the model did. A company that reports token growth without outcome attribution is a company describing its own effort — closer to a consultant who bills by the hour and calls that progress.
Forward-deployed engineering, ironically, is the function that could let the labs measure outcomes — those engineers sit close enough to the customer's P&L to see what moved. But the engineers themselves are evidence that the product cannot yet read that P&L on its own. The motion proves the gap it is sent to close.
There is now a second tell behind the first. The labs have begun dispatching a product-management twin to the engineer — variously titled Forward Deployed PM, Applied AI PM, or, on Anthropic's careers page, "Technical Deployment Lead, Applied AI." The historical analogue traces back to the same 2003 root we opened with. That young analytics company's customer-embedded function did not consist only of engineers; it split, deliberately, into a Forward Deployed Software Engineer and a Deployment Strategist — the latter a generalist whose job was to scope the work and translate between the platform and the client's operations. Recreating that second role is the louder tell. An engineer in the field is evidence that the product surface cannot ship judgment unaided. A product manager in the field is evidence that the product roadmap itself cannot be defined from inside the lab — that what gets built is being co-discovered at the customer's desk, one whale at a time. The labs are not only shipping bodies to make the product work. They are shipping startup founding teams, one per logo, and calling that scale.
In the weeks before this published, the motion stopped hiding inside job listings. OpenAI stood up a separate deployment company, capitalized in the billions and majority-owned by the lab, and staffed it in part by acquiring a consultancy outright — buying the forward-deployed headcount rather than posting for it. Anthropic announced an enterprise venture of its own in the same window, backed by private-equity capital and built on its Applied AI bench. Read these as good news for the model business and the structure stops making sense: you do not raise outside billions and acquire a services firm to deliver a product that already carries its own outcomes. You do it when the bodies are the deliverable.
What a different shape would look like
The reason we keep coming back to Embedded Judgment as the bar is that it is the property that makes the FDE motion unnecessary. A product that compiles the expertise of the engineer into its own surface — that knows where to look in the customer's data, what to do with what it finds, and how to translate the answer into a decision the operator can sign — does not need a person sent on a plane. The Expertise Floor is low enough that the buyer is also the implementer.
We do not pretend this is easy. The reason forward-deployed engineering exists, in any era, is that compiling judgment into a product is harder than dispatching labor. We build the way we do because we believe the only durable answer to that problem is to build the product from inside the operator's workflow — close to the data, close to the decision — rather than to keep wrapping the model in human bandwidth and calling that scale.
The FDE expansion is not a scandal. It is a signal. It tells us that the labs, who have the best read on what their own models can and cannot do unaided, have concluded that the products are not yet capable of carrying enterprise outcomes on their own. That conclusion is the most honest disclosure they have made this cycle.
The next interesting thing they could disclose would be the moment they stop needing the engineers.
A8C Ventures is an AI-native firm building technology for industries where information asymmetry costs people the most.
