OpenAI and Anthropic are now embedding Forward Deployment Engineers, a role descended from Palantir's military teams, into call centers, deal pipelines, and fund operations. The bet is roughly $4 billion. The role may eventually be replaced by the agents it deploys.
The first move was OpenAI spinning up a Deployment Company in early May, with the reported roughly $4 billion backing from private equity that gave the unit its name on a slide. Days later, Anthropic announced a joint venture with Blackstone and other financial institutions to do something similar at the same scale. Within a single week, the two leading AI model companies had each decided that the bottleneck was no longer in the lab. It was in the customer's call center, the portfolio company's due-diligence pipeline, and the fund's own back office. And so they reached past the usual consulting middlemen and wrote checks directly to the people they hoped could close the gap.
That person is the Forward Deployment Engineer, or FDE: an engineer whose job is not to build the model, but to make the model actually work inside one specific company's operations. The role is not new in spirit. It descends directly from Palantir's forward-deployed teams, the groups internally known as Echo and Delta, that embedded with military and intelligence customers in the late 2000s and early 2010s to install Palantir's data platforms where off-the-shelf software would not survive contact with the customer. In a long conversation on the 硅谷101 podcast, episode 240, Jove, who leads the Forward Deployment team at Cresta, and Oliver, a vice president at Invisible Technologies and former McKinsey consultant, describe the same posture. The engineer shows up at the client, learns the workflow, and only then starts building.
What is new is who is paying for the embedding, and why now. The capital behind the recent OpenAI and Anthropic moves is private equity, and the pressure to deploy these engineers is not coming from the AI vendors' sales teams. It is coming from the Limited Partners, the pension funds, sovereign wealth funds, and endowments, that sit above the private equity funds and have made AI transformation a fiduciary governance expectation. As the 硅谷101 E240 episode describes, general partners are being told by their own LPs that failure to integrate AI into portfolio companies is itself a governance failure. That posture pushes deployment capital toward vendors willing to embed engineers directly. A traditional consulting firm can advise a board, but it cannot embed a working AI agent inside a regional bank's loan-origination workflow the way an FDE team can — one embedded with the bank and backed by a PE sponsor whose own fund is on the line.
The role itself sits in an awkward spot in the org chart. Jove compares the FDE to a CTO and the company's internal engineering leader to a CEO. The FDE has the model knowledge; the customer has the business context; and the deployment only works if both show up. The work happens in the seams. The FDE pulls data from systems that were never designed to talk to each other, models the steps a human agent takes on a call, and rebuilds the workflow around an AI agent that can take the next call itself. Cresta, where Jove works, focuses on call centers because the workflows are high-volume, well-instrumented, and easy to measure. The same pattern shows up in PE diligence platforms, where AI is being used to read confidential information memoranda and draft memos, and in fund operations, where AI agents are starting to handle the rote work of allocating capital and reconciling positions.
Two failure modes show up again and again. The first is data integration done halfway. The FDE assumes a clean customer relationship management export and finds out, on day eight, that the call recordings live in a different system, in a different format, behind a different access policy. The second is forcing AI into a workflow where it does not fit. Not every step in a loan underwriting pipeline wants an agent; some steps want a deterministic rule, and a model gets in the way. The guests are candid about both, which is part of why the 硅谷101 E240 conversation reads as reporting rather than vendor promotion.
The harder question is whether the FDE role is durable at all. Jove and Oliver both name it, in different words, in the same stretch of the conversation. The FDE is, in part, building the training data and the integration patterns that the next generation of AI agents will use to do the FDE's own job. The lock-in that the FDE creates inside a customer is real, and so is the margin that recycles back to the model vendor and, from there, to the PE sponsor who funded the deployment company. But the agents that emerge from that work are, structurally, the FDE's replacement. The role is being paid for, in other words, by the same capital that is funding the technology that will eventually retire it. Whether that is a contradiction or a feature depends on how fast the agents improve, and on whether the LPs behind the deployment capital are willing to keep underwriting a job category that is, by design, working itself out of a job.
For now, the bet is that someone has to teach the agents the workflow, and the most expensive engineer in the room is the one who shows up at the customer to do it. The roughly $4 billion question is whether that someone, in three years, will still be human.