About six months ago, I was deep inside a problem I couldn't name.
I was building a system to do something specific: take the kind of forensic analysis that institutional due diligence requires and make it rigorous enough that a private equity investment committee could actually rely on it. Not impressive. Reliable. Those are different standards, and the distance between them turned out to be the entire problem.
The models were extraordinary. The gap between what they could do in isolation and what an institution could defensibly act on was larger than I had expected. Not because the AI was wrong but because the workflow, audit trail and chain of accountability - the architecture that would allow someone in a boardroom to say "we ran this process and we stand behind the output" - that didn't exist. It had to be built. And it had to live inside the institution, not outside it.
I didn't have a name for what I was trying to solve. The market does now.
Between January and September 2025, monthly job postings for forward-deployed engineers (FDEs) grew by more than 800%. Salesforce has committed to hiring a thousand of them. Google, OpenAI, Anthropic and Adobe - all building benches of these specialists, embedding them directly inside enterprise clients.
An 800% surge in a role designed to bridge the gap between what AI can do and what institutions can actually rely on is not a sign that the gap is closing. It is a sign that the gap is larger than anyone admitted when they were selling the licence.
Palantir invented the forward-deployed engineer in the early 2010s. They called them Deltas then - engineers embedded directly with government clients, building in hostile data environments, under conditions where a wrong output didn't just embarrass someone. It costs something real.
By 2016, Palantir had more Deltas than software engineers. The ratio of deployers to builders exceeded one-to-one. At the time, this looked like a peculiarity of Palantir's model. In retrospect, it was a preview.
The FDE profile tells you everything: deep technical skills, yes. But also customer empathy, radical ownership and the ability to translate between what a client believes they need and what will actually survive contact with their systems, their data, their audit requirements and their board. One industry description frames it plainly - the role is like being a hands-on CTO of a high-stakes AI startup, except the startup is inside someone else's organisation and the definition of success changes every quarter.
The standard explanation is that enterprises are finally serious about AI adoption and need specialist help. That is true, but it is the comfortable version.
The harder version: the AI economy spent three years selling capability and is now confronting operationalisation. Building a model that impresses in a demo and building a system that performs reliably at institutional scale, survives regulatory scrutiny and can be defended when something goes wrong - these are not the same problem. They are barely related problems.
The FDE exists to solve the second problem, because the products were designed to showcase the first.
This pattern is not new. In the 1990s and early 2000s, SAP and Oracle sold transformative Enterprise Resource Planning (ERP) systems that would revolutionise enterprise operations. What actually happened: they spawned a multi-billion-dollar consulting ecosystem - Accenture, Deloitte, the Big Four - because the gap between "the software works" and "the software works inside your business" turned out to be wider than anyone admitted. AI is following the same trajectory, with one critical difference: ERP outputs were deterministic and expected. AI outputs are not. The stakes are higher and the need for human judgment in the loop is permanent, not transitional.
Consider what the role actually requires in 2026. Agentic workflows. Evaluation frameworks. Observability tooling. Total compensation averaging $238,000, with senior practitioners clearing $630,000, because you need to be a strong engineer and a high-stakes relationship manager simultaneously, in an environment where the client's tolerance for ambiguity is close to zero.
That combination is rare because the need for it was not supposed to exist. The pitch was that the tools were enterprise-ready.
The job market is pricing a capability gap that no one wanted to admit was there.
Every FDE being hired today is being hired by a vendor. OpenAI's FDE. Anthropic's FDE. Salesforce's bench of a thousand. They are embedded with enterprise clients, yes - but they are deployed on behalf of the product, accountable to the product team, and building institutional knowledge that by default lives in the vendor's field organisation, not in the client's.
Could a vendor FDE help you build internal capability? Yes - if that is the explicit objective. If the engagement is structured around knowledge transfer, if you're building a vendor-agnostic architecture, or if you're using the FDE to train a permanent internal team. That is possible.
But that is not the default engagement model. The default is a service relationship: the FDE embeds, solves the immediate deployment problem, builds deep expertise in how your firm uses their product, and that knowledge accretes to the vendor's field organisation. When the engagement ends, the FDE leaves. The architecture they built is often tightly coupled to the vendor's platform. The firm is left with a working system and also a dependency.
That is a service relationship. It is not a capability.
In most industries, that distinction is uncomfortable but manageable. In private equity and private credit - where the output of any AI system will eventually be tested against a capital allocation decision, a board recommendation, a regulatory enquiry or a failed deal - it is a structural risk.
The firms that will compound advantage over the next five years are the ones asking a different question upfront: not "can we hire vendor FDEs to deploy AI?" but "are we using this deployment to build permanent, internal, vendor-agnostic capability - or are we outsourcing judgment we will need to own?"
The FDE boom is revealing something that the model was never the product. The deployment was always the product. It just took an 800% surge in job postings to give that idea a job title.
The question is not whether your firm is using AI. Most are, in some form. The question is whether you are building capability or renting it - and whether you can tell the difference.
Those are not the same question. And in this industry, they rarely have the same answer.
Six months ago, I couldn't name the problem I was solving. Now the market has named it.
And it's charging you $238,000 a year to rent it back.
Sources:
— Financial Times / Indeed: FDE job posting data, January–September 2025
— Salesforce: Forward Deployed Engineer programme announcement, April 2025
— Lorien Global: Emerging AI Jobs in Demand, 2026
— Onward Search: Top AI Jobs to Watch in 2026, February 2026
— Hashnode: Complete 2026 Guide to the Forward Deployed Engineer, February 2026