Palantir's average customer pays it roughly $4 million a year. ServiceNow, the next-largest on the list, averages about $1.3 million per customer. Workday sits around $800,000. Almost no other public SaaS company cracks half a million.
Kevin Bai assembled that league table from public filings, and his reading of the gap is short: it's the motion. Palantir's customers aren't buying seats or consultant hours. They're buying an outcome, delivered by engineers the vendor loans them. In Bai's framing it's a design partnership, the kind usually retired after product-market fit, scaled up into the enterprise instead.
Every AI company now wants some version of that motion, and postings for the Forward Deployed Engineer title grew 1,165% year over year by the end of 2025. For a founder, the question is whether to copy it. For a buyer, the question is whether a vendor offering embedded engineers is selling something real. Both come down to two questions and one number.
Question one: what are you selling, and to whom?
Bai works on Anthropic's Applied AI team. Before that he was the first FDE at Rippling, where the team grew to roughly twenty-five people in a year, and before that he was at Palantir. His test is a two-by-two, with what you sell on one axis and who buys it on the other.
Sell a technical product to a technical buyer, the way GitHub and Datadog do, and you don't need FDEs. The buyer can integrate it alone. Sell a configurable product to a non-technical buyer, like Jira or Slack, and you don't need them either, because configuration doesn't require engineering.
The role earns its cost in exactly one quadrant: a deeply technical product sold to a buyer who isn't technical. That was Palantir's corner for two decades, and for most of that time the rest of the industry treated it as a curiosity.
Bai's hypothesis is that AI is quietly pushing more companies into that corner. An agentic platform is customizable by definition. Hand a customizable platform to a business buyer and you have a technical product owned by someone who can't deploy it alone. That gap is what the forward deployed engineer exists to close.
So the first screening question is blunt. Do you have to sell something technical to a non-technical buyer? If your customers have engineering teams that can wire your product in themselves, or if setup is just configuration, an FDE team is an expensive answer to a problem you don't have.
The number: $100,000
The quadrant tells you whether the role is useful. Contract size tells you whether you can afford it.
Tomasz Tunguz, who has written the most useful short essays on the go-to-market side of the FDE motion, put the floor at roughly $100,000 in annual contract value in a July 2025 essay. Below it, embedded engineering labor costs more than the contract can support, and the motion is dilutive no matter how strategic the relationship looks. The companies that have tried to run FDE teams on lower-ACV products have, almost without exception, run out of money before reaching scale.
Above the floor, the math improves quickly. The frontier labs run the motion on seven- and eight-figure annual contracts with single enterprise customers. Anthropic's joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs targets the mid-market, which it defines as companies with roughly $100 million to $5 billion in revenue. Contracts there land between $250,000 and $2 million, with real room to expand. The vertical AI companies (Sierra, Decagon, Harvey, Glean, Hebbia and others) mostly sell at the upper end of that range or above it, where a single customer justifies dedicated FDE attention.
For a founder, this is the diagnostic. If your average contract sits below about $100,000, the FDE motion is the wrong operating model. Either the contract has to grow or the go-to-market has to change, usually toward product-led growth. Forcing an FDE motion onto a product that can't carry it is the single most common reason early-stage AI startups burn cash without reaching the position they hoped for.
Question two: is there a platform underneath?
Landing in the right quadrant and clearing the floor still isn't enough. Bai attaches a warning that's worth reading slowly:
"If you were to implement an FDE function where each FDE is building entirely from scratch, my friends, you do not have an FDE function. You have a dev shop."
A dev shop can be a decent business. Its economics look nothing like Palantir's. The FDEs have to build on a platform of shared primitives, so that what one engineer learns at one customer makes the next deployment cheaper. Without that, every new customer costs about as much to serve as the first, and the margins stay at services levels because the work is services.
Hence Bai's second screening question: do you have a platform for the FDEs to build on, or the willingness to invest in one? A "no" to the first question means you don't need the role. A "no" to the second means you won't be able to afford it for long.
Reading the same tests from the buyer's side
The two questions work just as well across the table. If a vendor proposes to embed engineers in your organization, ask which quadrant its product sits in. A configurable tool that still needs months of custom engineering at every customer is telling you something about the product.
Then ask about the platform. The quickest check is whether the vendor's FDEs ship code into the vendor's own product, or only build custom work for single accounts. The first is a company getting cheaper to deploy with every customer, including you. The second is a dev shop with a software logo.
Finally, check the contract against the floor. A vendor running embedded engineers on deals well under $100,000 a year is running a motion that, on Tunguz's math, is dilutive at that size. That's worth knowing before your operations depend on it.
Bai defines the person in deliberately plain terms: an FDE is "nothing more than a customer-facing software engineer." Whether you need one depends on everything standing behind that engineer.