THE FDE
BET.

Forward-deployed engineers, or FDEs, are one of the most talked-about roles in AI right now. Palantir made the job famous by sending engineers to sit inside each customer until the software did the work. Many AI startups that sell to lawyers, bankers, auditors and support teams now run their own version. Most of Harvey's and Legora's FDEs are lawyers, and Rogo's are former bankers.

Investors like the approach. Five of these companies disclosed revenue close to their latest funding round, and each raised money at 39 to 100 times that revenue. A price like that assumes revenue will keep growing much faster than costs.

FDEs draw full engineering salaries, so the approach only pays if each customer needs fewer of them over time. I spent October studying fourteen of these companies, from public evidence alone, to see whether that is happening.

[THE FOURTEEN]

All fourteen sell to expert buyers and place their own people inside customers. Everything that follows comes from public sources, and how I gathered and checked it is at the end.

[STILL HIRING]

The simplest place to look is the job board, where shrinking FDE work would show up as fewer FDE openings.

FDE openings rose instead. Nine of the twelve companies with a year of history added them. Harvey went from 5 to 27, Sierra from 1 to 22 and Decagon from 3 to 21.

FDEs are also a larger share of who these companies are hiring than of who they already employ.

At eleven of twelve the filled dot sits to the right, though at Harvey and Legora only by a point. Sardine's openings are 20% FDE against 4% of its staff, and Decagon's 19% against 8%. Cinder is the one exception, hiring FDEs below its current share.

Openings have a limit, though. A rising count can mean more customers or new countries as easily as more work per customer. Work per customer is not public, so the closest proxy is money, and above all revenue per employee.

[THE MONEY]

None of the fourteen publishes its finances, so I estimated a profit and loss for each one from public pieces. I set them beside public companies with audited accounts that also send people out with their software. The charts measure profit as EBITDA, operating profit before interest, tax and accounting charges for equipment and acquisitions.

Eleven of the fourteen bring in between $150k and $250k per employee. That puts them next to BigBear.ai and C3.ai, far from Palantir's $1.76M, and by these estimates twelve of the fourteen lose money. Only Taktile and Distyl come out at or above break even. Cognition is the outlier at about $2.1M per employee, with partners doing part of its delivery work.

Part of the answer shows in how they grow. The chart calls the measure elasticity, and I will call it hiring pace.

By reported figures, Harvey's revenue grew about 260% last year and its LinkedIn headcount about 272%. Sierra and Legora grew staff almost as fast as revenue. Palantir grew revenue 93% on 6% more people. For every 10% of new revenue, the typical startup here added six to seven percent more staff, and Palantir added under one.

So most of these companies are getting leaner, just slowly. Thirteen of the fourteen brought in more per employee than a year before, but none came close to Palantir's pace. One year is also a short window, and young companies often hire ahead of revenue, so treat the hiring pace as a direction.

[WHAT PALANTIR DID]

Palantir still sends FDEs. Its filing says "our software engineers working with existing customers often manage the deployment and operation of our platforms." Over time, each of them has had less to do, the same way a good developer platform carries more of the work on its users' behalf, which I wrote about in Building Is Free.

The chart measures contribution margin, what Palantir keeps of each dollar after deploying, running and selling its software. In 2019, by its own accounts, customers still being set up lost 43 cents on the dollar while mature customers kept 55. Its 2020 filings credit the jump that followed to "a significant decrease in the time and number of software engineers required to install and deploy our software platforms." Getting a new customer working fell more than five-fold, to an average of 14 days. In 2023 it replaced pilots that took one to three months with five-day bootcamps, workshops that build a working tool on the customer's own data. Over the same stretch the share of revenue left after deploying and selling went from 21% to 74%.

C3.ai went the other way. From 2021 to 2025 it made long pilots cheaper to buy, and its staff doubled along with its revenue. Last year both fell.

[THE BET]

So the whole bet rests on one number, how fast headcount grows as revenue grows. I ran each startup forward five years twice, once hiring at its own pace and once at Palantir's long-run pace, with everything else held the same.

At their own pace, six of the fourteen still lose money in year five, and in the model Harvey sits near minus 70%. Hired at Palantir's pace, all fourteen break even within two years. The model assumes the same revenue arrives with fewer people, which is what these companies have yet to show, so read it as the size of the prize. The lowest pace any of the fourteen showed last year was Norm Ai's 0.31, so none has yet hired at Palantir's 0.26.

[WHAT TO WATCH]

FDEs are a sound way to learn a customer's work and win the account. The bet is that each one is needed a little less every quarter, and you can watch for that from outside. Compare the FDE share of a company's openings with the FDE share of its staff, the second chart above. When the openings share drops below the staff share, the product may have started doing work the people used to do. A hiring freeze or a sales push can move it too, so read it alongside revenue per employee.

Today most of these companies are getting leaner, and none is near Palantir's pace. Harvey, the largest, is still adding people about as fast as it adds revenue. The bet pays when that changes, and the job boards will show it first.

[HOW I DID IT]

Startup margins are shown before stock pay, the way most public companies report their adjusted margins, and counting it would make every startup look worse. The public companies are compared on their own reported measures, which differ in the details.

The study has limits. The fourteen were chosen for being alike, so they describe themselves and no wider market. Startup revenue is a run rate, a year of revenue projected from the latest month or quarter, and the margins are estimates with ranges that cheaper AI models would lift.