Why AI labs are racing to hire forward deployed engineers

Open LinkedIn this week and somebody is telling you to become a forward deployed engineer. For once, the hype has receipts. Job postings are up something like 800% in a year. OpenAI stood up a $4 billion joint venture with the gloriously unsubtle name “the deployment company.” Anthropic is building a $1.5 billion version of the same idea.

So before you rewrite your headline to “FDE: shipping AI in the enterprise,” it’s worth asking what the job actually is, and why the most valuable AI labs on earth suddenly need armies of people to do it.

The quiet part: the models don’t work yet

Not in the lab. In your company.

A frontier model can pass the bar exam and still have no idea which of your seven Salesforce instances holds the real customer record. MIT’s NANDA initiative put a number on the gap this year: 95% of enterprise GenAI pilots showed no measurable business impact. Read that again. Not 95% were bad models. 95% never made it from demo to dollars.

That’s not a model problem. It’s an everything-else problem. The data lives in a schema someone designed in 2009 and then left the company. The workflow has a compliance step that only exists because of a lawsuit nobody talks about. The “obvious” automation falls over the first time it meets an edge case the people doing the job could have warned you about on day one.

The model is genuinely the easy part now. Getting it to survive contact with a real business is the hard part, and that part still wants a human in the room.

Palantir already ran this experiment

None of this is new. It’s just new to the labs.

Palantir worked it out more than a decade ago. Sell a software license to a Fortune 500 and walk away, and you mostly bought yourself some shelfware and an awkward renewal call. So they did the deliberately unscalable thing: they sent their own engineers to sit inside the customer (at the bank, the agency, the hospital) and refused to leave until the system actually ran. They called them forward deployed engineers. The Valley smirked at how un-SaaS it all was. It also worked.

OpenAI and Anthropic are now copying that homework, except the shelfware in question is a $20-a-seat model that dazzles for a week and then quietly gets abandoned. A “deployment company” is just Palantir’s old insight with a far bigger checkbook: stop selling access, start selling outcomes, and put a person on-site to make sure the outcome shows up.

The job is translation, not prompting

This is where the LinkedIn version gets it wrong.

The people winning these roles aren’t AI whisperers. Forty clever prompt tricks are table stakes, and frankly that’s the skill the model is quickest to absorb on its own. The scarce ability is translation: sitting between a system that speaks tokens and a business that speaks claims, transactions, tickets, and regulations, and getting the two to understand each other.

In practice that means reading a legacy schema without flinching. Sitting through a security review and knowing which answer ends the meeting and which one ends the deal. Watching how the work really gets done, not how the org chart says it gets done, and catching the undocumented step everyone on the floor takes for granted. A good FDE spends a strange amount of time just watching people work.

It’s less “AI researcher” and more “consultant who can also ship code and won’t get steamrolled by the client’s IT team.” And it’s worth saying plainly: it can be a grind. You’re on someone else’s turf, on their timeline, fixing problems that appear on no model card anywhere. The glamour mostly lives in the title.

So what do you actually go learn?

If you want one of these roles, the counterintuitive move is to spend less time on AI.

Go learn how a bank actually clears a transaction. How a hospital actually codes a claim and gets paid. How a refinery actually schedules maintenance without anything catching fire. Pick a messy, regulated, high-stakes domain and get fluent in its plumbing:

• the systems of record, and which one is quietly lying
• the constraints that look dumb until you learn the regulation behind them
• the words insiders use, and the words that mark you as an outsider
• the reasons things are slow that have nothing to do with incompetence

The model is a commodity now. Everyone’s renting the same handful of them. Domain fluency is not. The person who can walk into a logistics company already speaking their language is worth more than the person with an immaculate eval suite who has never heard the phrase “bill of lading.”

The frontier moved. The intelligence got cheap; the work of making it land somewhere real did not.

So if you’re betting the next few years of your career on AI, don’t bet on the model. Bet on the messy, human, deeply specific work of making it useful inside a real organization. That’s the part that’s still hiring. And if you want to talk about building in that space, come find us.