Why Every AI Lab Is Hiring Forward Deployed Engineers, and What It Means for Your Portfolio

Key Takeaways

• The labs are telling you where the bottleneck is. The companies with the best models are hiring engineers to sit inside customer companies. They would not pay for that if the model did the job alone.
• Demand moved fast. Indeed data shows forward deployed engineer postings up roughly 729% in a year, and one study saw the share of companies planning to hire FDEs go from 5–10% to 70% in six months.
• The labs are coming to your portfolio through your sponsors. Anthropic and OpenAI both set up deployment ventures funded by private equity firms, with portfolio companies as a natural first market.
• A model needs four things it can't bring with it: your context, a real workflow, proof it works on your cases, and someone who owns it after launch.
• Ownership is the question to ask first. Whoever sends the engineers, the portfolio company has to be able to run the system when they leave.

The most interesting hiring signal in AI right now has nothing to do with research talent.

The labs that build the best models in the world are hiring engineers whose job is to leave the lab. They go sit inside a customer's company, get access to its systems, and stay until something works in production. Indeed data shared with Business Insider shows postings for forward deployed engineers up roughly 729% year over year as of April 2026. Anthropic, OpenAI, Palantir and Stripe are all on the list of companies hiring for the role.

Think about what that admits. If a better model were enough, the labs would put every dollar into the next model and let customers figure out the rest. They are doing the opposite. They are paying senior engineers to do the slow, specific work of fitting a model into one company's data, one team's process, and one CFO's definition of "working."

MIT's NANDA initiative put a number on how badly that last mile goes when nobody does it. In their research, about 95% of generative AI pilots were falling short of real P&L impact. (The report's definition of success is narrow and the sample is small, so I treat 95% as a direction rather than a measurement. The direction matches everything I see.)

So the labs reached the same conclusion most operators reach after their second stalled pilot. The model is the easy part. Getting it to do real work inside a real company is the job.

If you run a portfolio, that conclusion should change how you plan AI across every company you own.

The Labs Just Told You Where the Bottleneck Is

A forward deployed engineer, or FDE, is a senior engineer who works inside the customer's environment. Palantir made the model famous: rather than ship software and hope, it sent engineers to live with the customer until the software did what the customer needed.

The AI labs copied that model because they hit the same wall. A frontier model can pass a bar exam. It still can't read your ERP, doesn't know that your Midwest plants code returns differently, and has no idea which of your 40 approval steps actually matter.

Demand shot up this year. TechCrunch's write-up of the Christian & Timbers FDE study reports that at the start of 2026, only 5% to 10% of companies planned to hire FDEs, mostly for small pilots. By the end of the second quarter, that number was 70%. (Christian & Timbers is an executive search firm, so it has a stake in a hot hiring market. I still think the shape is right.)

Read the hiring plan, not the press release

Model launches get the headlines. Hiring plans tell you where a company thinks its revenue is stuck. When the labs spend on people who go on site, they are telling you that deployment is what stands between a model and a paying customer.

A Model Without a Deployment Is Just a Demo

I'll say this flat. Most AI pilots in the mid-market don't fail because the model is too weak. They fail because nobody did the unglamorous work between the demo and the workflow.

That work has a pattern. Someone has to get the data out of the systems it lives in. Someone has to pick one process and wire the model into it, with a trigger and a place the output lands. Someone has to test it on the company's own cases and set a bar the business agrees to. And someone has to own it once it's live.

None of that is research. All of it is engineering, and most of it can only happen on site.

Here's the uncomfortable part: the model gets better every quarter, and that doesn't fix any of the four. A smarter model reading the wrong data with no workflow around it is still a demo.

Budget for the last mile before you pick the model

When a portfolio company asks which model to use, the honest answer is usually "it matters less than you think." Ode's chief technologist said roughly that to TechCrunch: model selection matters, but it isn't where most of the effort goes. Put the budget conversation where the effort is.

The Labs Are Coming to Your Portfolio Through Your Sponsors

This is the part PE firms should pay the closest attention to.

In May, Anthropic and OpenAI both announced enterprise services ventures. Anthropic's founding partners are Blackstone, Hellman & Friedman and Goldman Sachs, with Apollo, General Atlantic, GIC, Leonard Green and Sequoia also backing it. OpenAI's venture, The Deployment Company, was reported to be raising from TPG, Brookfield, Advent and Bain Capital. TechCrunch's read is that both ventures will likely get preferred access to their investors' portfolio companies.

Anthropic's venture is now Ode with Anthropic, built on an acquired AI boutique and staffed with about 100 engineers, according to TechCrunch's July profile. The backers will refer their portfolio companies, though Ode won't sell only to them.

Read that from the operating partner's chair. The labs decided that the fastest path into mid-market companies runs through the people who own them. If you sit on a portfolio company board, expect to hear a forward deployed pitch soon. Possibly from a firm your fund knows well.

That isn't bad news. It does mean you need a point of view before the pitch shows up.

Decide what "good" looks like before the first meeting

Write down, for each portfolio company, the one or two workflows where AI would move EBITDA. Then ask any partner, lab-backed or not, how they would get one of them to production and who would own it afterward. The answer to the second question tells you more than the first.

Engineers in the Room Beat Licenses on the Shelf

There's a reason the labs chose people over more software.

The same MIT research found that buying from specialized vendors and building partnerships succeeded about 67% of the time, while internal builds succeeded about a third as often. My read: the partnerships that work bring people who have done the deployment before. A license on its own is a shelf item.

Most portfolio companies can't hire this talent directly anyway. The FDE profile is a senior engineer who can also sit with a controller, a plant manager or a claims team and translate. Those people are rare and expensive, and a $150M manufacturer is not where they are looking for their next job.

So the practical question for a portfolio company is which outside team sends engineers, and on what terms.

Tie the budget to one workflow

Licenses are easy to buy and easy to report on. "We rolled out AI to 400 people" sounds like progress at a board meeting. Ask what one workflow is now faster, cheaper or more accurate because of it. If nobody can answer, the money went to the shelf.

The Four Things a Model Can't Bring With It

Here's the framework I use when a portfolio company asks why its AI pilot stalled. A model shows up with general knowledge and an API. It needs four things the company has to supply, and each one needs an owner.

Forward deployed engineering exists to close those four gaps on site. That's the whole job. Everything else is title inflation.

Score each portfolio company on the four

You don't need a maturity model with 30 criteria. For each company, ask whether the data is reachable, whether one workflow has been picked, whether there's an eval set, and whether someone inside will own the result. Most companies have one or two of the four. The missing ones tell you where the first engagement should start.

Ownership Matters More Than Who Sends the Engineers

This is the bit that is still unsolved, and I don't want to pretend otherwise.

The FDE model has a built-in tension. The engineers are most useful while they're inside your company. The company is only better off if it can run the system after they leave. A deployment that only works with the vendor's engineers in the room is a dependency.

Lab-backed ventures, consultancies and independent firms all have to answer that question, and the incentives point in different directions. A lab-backed team will reasonably prefer its own model. Ode describes itself as "Claude-first" while saying it will use rival products when needed. That may be the right call for a given workflow. It's still worth knowing, because the model choice tends to outlast the engagement.

I don't think there's a clean answer yet on how long an FDE team should stay, or what a fair handoff looks like for a company with a five-person IT team. I do know what a bad one looks like: a final handoff meeting, a binder of docs, and nobody inside who has ever changed the code.

Make knowledge transfer part of every sprint

Ask for it in the contract. Your people should pair with the embedded engineers, see the evals, and ship changes themselves before the engagement ends. That's how we run Forward Deployed AI Engineering at PressW: our engineers work in the client's repos and rituals, and knowledge transfer happens every sprint instead of at the end.

Treat Deployment as a Portfolio Capability

The labs looked at the gap between their models and their customers' results and decided to send engineers. That's the most honest signal in AI right now.

For a portfolio, the lesson is bigger than "hire an FDE." The work of getting AI into production repeats across companies: the data access patterns, the eval habits, the way you pick a first workflow. A fund that learns it once and reuses it across ten companies will move faster than ten companies learning it separately.

I'm not sure yet how the market sorts out. Lab-backed ventures, the big consultancies and smaller specialist firms will all compete for the same portfolio companies, and the FDE title is already getting stretched. What I'm sure of is that the companies with an owner, a workflow and an eval set will get value no matter who sends the engineers.

If you're working out where to start across your portfolio, pick one company and one workflow and get it into production first. We do that work through Forward Deployed AI Engineering, and if you want to compare notes on what you're seeing, Book a Discovery Call.

FAQ