Palantir was founded in 2003 with a hard go-to-market problem: how do you sell software to an intelligence community whose work is, by design, hidden? The founders couldn’t walk into a CIA office and ask what was broken. So Stephen Cohen built a prototype in eight weeks and took it to analysts with a question: does this match what you do? The answer, the first several times, was no. But presented with a tangible thing rather than a survey, the analysts started describing their work in concrete terms. Each conversation produced a list of edits, each list produced a new version, and the cycle repeated: build, show, listen, rebuild.
Then came the part the industry missed. As Palantir scaled, the standard advice (generalize across your first segment, then move on) failed. A counter-proliferation workflow was not a counter-terrorism workflow, and a financial intelligence workflow was not a signals intelligence workflow. The generalized product was not quite right for anyone.
Enter the forward-deployed engineer
The response is most associated with Shyam Sankar, Palantir’s thirteenth employee, who became the company’s first forward-deployed engineer and is credited with coining the title. The early deployments were unusually literal about “shoulder-to-shoulder”: three-month rotations through Kandahar, midnight helicopter rides to forward operating bases, summers embedded with Navy SEAL teams debugging databases between rocket-fire alerts. The point of all this was proximity, not romance: the FDE was inside the room where the work happened, close enough to see what broke, fix it, and put the fix in front of the operator without the latency of a headquarters routing slip.
The role had a specific structure. The FDE embedded with a single customer for as long as it took, writing whatever code was required: integrations, custom views, data-model extensions, occasionally features pushed upstream into the core product.
The metaphor that explains the economics
Sankar’s framing: the FDE builds a gravel road, a fast, customer-specific, sometimes ugly path from where the product is to where the customer needs it to be. The product team’s job, back at headquarters, is to look at the gravel roads several FDEs have built, spot the pattern, and convert it into a paved highway: a generalized capability future FDEs can use without rebuilding it.
This is what separates the FDE model from professional services, where there are no highways, only an endless string of gravel roads. Over time, the road network grows until a new customer can drive on existing highways for most of their use case and only needs gravel for the last mile. At that point the customer can be served with far less FDE labor per contract dollar, and gross margins climb toward software levels.
The ontology: where the value accrues
Palantir’s second insight was architectural. If every FDE chose their own data representation, the company would have a different schema at every site and could never ship a coherent product. The solution was the ontology, a layer above the database that let each FDE define the objects, properties, and links that mattered for that customer’s domain, all in one consistent format. The same engine could power a counter-narcotics workflow and an aerospace manufacturing workflow at the same time, and improvements compounded across customers.
Sankar has been explicit that this is where Palantir’s economic value accrues. The frontier models are commoditizing, and the chip layer is concentrated in a few hands, so the value in between goes to the layer that grounds the model in the customer’s actual semantic reality. A customer two years into a deployment won’t rip and replace, because by then the ontology encodes too much of how their business actually runs.
Every AI-native startup adopting the FDE motion should be asking, from the first engagement, what its equivalent of the ontology is. Maybe it’s an agent skill library, a domain-specific eval framework, or a workflow registry. The specific form depends on the domain, but the category is constant: the artifact that captures per-customer learning in a way that produces both vendor leverage and customer lock-in. Without that artifact, the FDE motion is, accurately, services. With it, you have a software business with a services on-ramp.
By 2024, with its AI Platform built on the ontology, Palantir’s market value had grown fifty-fold in three years, and the alumni network the role produced (Anduril, Kalshi, ElevenLabs, and a long tail) was collectively worth over $50 billion. The model the industry had filed away as a government-contracting curiosity is now how most of the AI industry goes to market.
Want the full playbook, with the templates, the interview transcripts, and the week-one checklist? It’s all in my book Forward Deployed AI Engineering: A Working Guide to the Hottest Job in Software, available on Amazon.