An MIT study made headlines with a brutal number: roughly 95 percent of enterprise generative AI pilots show no measurable impact on the bottom line. The researchers didn’t blame the models. They traced most of the failures to the gap between a demo that impresses a steering committee and a system that has to survive contact with somebody’s actual ERP or claims database.
That gap has a name now, and increasingly, a job title attached to closing it. The job title is forward deployed engineer, usually shortened to FDE. In plain terms, a forward deployed engineer is a software engineer who embeds inside your team, builds against your own systems rather than a sandboxed demo environment, and stays until a specific business outcome is running. You are buying a result rather than a tool or a slide deck, and the engineer stays on the hook until that result exists.
This guide covers where the role came from, why OpenAI, Anthropic, AWS, and Salesforce are now racing to build what used to be a niche Palantir specialty, what an FDE does in a given week, and how the model looks once a 40-person company buys it instead of a defense contractor.
What “Forward Deployed Engineer” Actually Means
Most software gets sold as a product. You buy a license, install it, and adapt your process to fit it. A forward deployed engineer works the other way around: the software gets built, or heavily customized, to fit your process and your data, edge cases included, and the engineer doesn’t hand over a spec and leave.
That’s the shape of what’s sold today as an embedded AI engineering service. Instead of contracting for a deliverable such as a technical document or a set of API integrations, you’re contracting for an outcome: claims processed faster, say, or a scheduling system that stops double-booking crews. The engineer is measured against whether that outcome actually shows up in production. A milestone document getting signed off along the way is only a checkpoint on the road there.
Picture the difference in contract terms. A typical software vendor contract lists deliverables such as a working integration or a configured dashboard, and payment ties to those deliverables regardless of whether anyone ends up using the result day to day. An FDE engagement instead defines an outcome, something like “average claims-processing time drops from six days to two” or “the support backlog stays under 50 open tickets,” and the engineer’s job isn’t done until that number holds up under real conditions.
This model predates the current AI wave by a long way. Palantir built an entire company around it two decades ago, back when nobody was calling it AI deployment, or calling it anything at all beyond “the way we do delivery.”
Where the Term Came From: Palantir, 2006
Shyam Sankar joined Palantir in 2006 as employee number thirteen and became its first forward deployed engineer. The role grew out of a practical problem: intelligence customers had classified workflows too specific for any off-the-shelf analytics product, and those workflows kept changing. Nobody was thinking about branding yet.
So Palantir sent engineers to sit inside the customer’s building, write code against the customer’s actual data, and keep adjusting until analysts were using it for real work. Palantir later formalized this into a role split it calls Dev and Delta. As the company’s own engineering blog puts it, Devs and Deltas focus in opposite directions: you can think of a Dev’s focus as one capability serving many customers, while a Delta’s focus is many capabilities serving one customer. A third role, Echo, sits alongside Deltas as a deployment strategist who understands the customer’s industry and translates between engineering and the people who do the work.
Until around 2016, Palantir employed more forward deployed engineers than product engineers. Whatever the field teams learned, again and again, from being embedded across dozens of deployments, eventually got generalized into what became the Foundry platform. That pattern, in Nabeel Qureshi’s account of his eight years at Palantir, is what let the company post 80 percent gross margins in 2023, closer to a software company than to a services firm like Accenture, which runs around 32 percent. Embedding engineers is expensive. Industrializing what they learn is what makes it profitable at scale.
Why “FDE” Suddenly Means an AI Job
Large language models are trained on a huge amount of general text and code. Your company’s data is not in that training set, and neither is the way your team works inside its own systems. It never will be by default. Getting a general-purpose model to reliably do one specific job, connected to one specific ERP or claims system, with real users and no one supervising every output, is what’s often called the last-mile problem in AI deployment.
That last mile looks a lot like classic FDE work. Someone has to sit with the team that does the work today and learn the exceptions that never made it into any documentation, then keep tuning the system’s guardrails and evaluation criteria until it all holds up under real volume. The model itself is general. The deployment never is: it has to handle your invoices and your compliance rules, right down to the quirks of your legacy database schema.
That’s the structural reason the term jumped from a Palantir specialty to an industry-wide job title in about two years. Integration work didn’t disappear when general-purpose AI arrived. General capability got cheap, and the last mile became the scarce, valuable part of the job.
The Money Behind the Job Title
Job titles come and go. What makes this one worth taking seriously is how much capital the largest AI companies have put behind it since early 2025.
OpenAI formalized its forward deployed engineering function in January 2025 under Colin Jarvis. Within a year the team went from two people to a multi-city presence, with customers including Morgan Stanley, Klarna, and T-Mobile. In May 2026, OpenAI and Anthropic both launched joint ventures built around the same idea. OpenAI’s Deployment Company (DeployCo) is a majority-owned services venture, capitalized in the multi-billion-dollar range, and it acquired the consultancy Tomoro to bring about 150 experienced forward deployed engineers and deployment specialists on day one. Anthropic’s enterprise AI services venture, backed by Blackstone, Hellman & Friedman, and Goldman Sachs at around 1.5 billion dollars, is explicitly aimed at mid-sized companies rather than only the largest enterprises.
The same bet shows up outside the frontier labs. AWS committed 1 billion dollars to a forward deployed engineering program that extends to its partner network, and partners keep what they build along the way: reusable delivery assets, evaluation frameworks, tooling. Salesforce has committed to building out 1,000 Agentforce FDEs. On Indeed, listings for forward deployed engineer roles grew from 643 in April 2025 to over 5,300 in April 2026, a 729 percent jump in a single year.
None of this proves that embedding an engineer guarantees a return for every company. It does show where the companies with the best visibility into why AI projects succeed or stall are placing their money: on deployment capacity ahead of raw model capability, which by mid-2026 is cheap and getting cheaper across every major provider.
What an FDE Actually Does, Day to Day
The job looks less like traditional software architecture and more like fieldwork with a laptop. A few things show up in almost every FDE engagement:
- Eval-first development. Before or alongside building the thing, the engineer writes a test suite specific to this task and this company’s data, so a change that quietly breaks behavior gets caught internally before a customer ever sees it.
- Context engineering. Feeding the model your company’s actual documents, data schemas, terminology, and edge cases, rather than relying on a generic prompt that happens to work in a demo.
- Integration into legacy systems. Connecting to the on-prem ERP, the ticketing tool with no modern API, or the database schema nobody has fully documented in a decade, because that’s where the workflow lives.
- Sitting with the people who do the work. Literally or over video, watching how claims processors or schedulers handle exceptions today, because those workarounds are the requirements no one wrote down.
- Iterating against a live outcome. Shipping a working version quickly, watching it run against real volume, and adjusting, rather than spending months on a spec before anything touches production.
The balance of effort differs from most software development work. Less time goes into architecting a clean, reusable system, and more into debugging why the model mishandles the invoice format this particular client’s customers happen to send.
How This Differs From a Consultant, a Solutions Engineer, or a Freelancer
The short version: a consultant is paid for advice, a solutions engineer is paid to win a deal, a freelancer is paid for hours against a fixed spec, and a forward deployed engineer is paid to reach a working outcome inside your own systems and stays until it happens.
| Role | Paid for | Where they sit | Owns the outcome |
|---|---|---|---|
| Forward deployed engineer | A working result in your systems | Embedded with your operational team | Yes, stays until it ships |
| Consultant | Advice, strategy, a recommendation | Client site periodically, workshops | Hands off a report |
| Solutions engineer | Winning the deal for a vendor’s product | Pre-sales, moves on after signing | Moves to the next prospect |
| Freelancer | Hours or a defined task | Remote, task-based | Only for the scoped task |
There’s more nuance to this comparison than fits in one section, including when each option is the cheaper or better choice for a given problem. That full breakdown lives in Embedded AI Engineer vs. Freelancer vs. Agency.
What This Means If You’re Not Palantir-Sized
A senior forward deployed engineer costs roughly 220,000 to 400,000 dollars a year fully loaded in the US market, with median advertised salaries for the role closer to 174,000 dollars. That’s before any margin a services firm adds on top if you’re buying the capability rather than hiring it directly. For most mid-sized companies in Europe or the US, hiring a full-time FDE in-house isn’t realistic, and it probably shouldn’t be the goal.
What is realistic is the same underlying model, scaled down: someone embeds with your team for a defined period, learns your systems, and ships working software against a specific outcome, without requiring a Palantir-sized budget or a multi-year enterprise contract. That’s usually sold today as an embedded AI engineer engagement, scoped to one workflow or one department rather than an entire company. AWS’s own partner program, which lets consulting firms keep the reusable delivery assets they build during an engagement, is one sign this model is being packaged for exactly this tier of the market rather than kept inside three AI labs and a handful of defense contractors.
It’s also not the right fit for every situation. If the goal is a single narrow chatbot widget or an already-templated automation, a freelancer or an agency working against a fixed spec will usually be cheaper and faster. Embedding earns its cost when the integration problem is genuinely messy: systems without clean APIs and workflows that live mostly in a few people’s heads, plus the edge cases that only surface once real volume hits. Any vendor can make a demo look impressive. Whether embedding is worth the money comes down to how messy your integration problem turns out to be once someone looks underneath the surface.
Frequently asked questions
What does a forward deployed engineer do?
A forward deployed engineer embeds inside a company's team and writes software against that company's own systems, data, and workflows until a specific outcome is running reliably. Day to day that means building evaluation tests for the exact task at hand and integrating with legacy systems that have no clean API. It also means sitting with the people who do the work to learn the exceptions no documentation captures.
What is the meaning of forward deployed engineer?
The term describes an engineer who is deployed forward, inside the customer's environment, rather than working back at a vendor's office on a generic product. Palantir coined the role in 2006 for intelligence customers whose workflows were too specific and too sensitive for off-the-shelf software, and the term has since spread across the AI industry to describe the same kind of embedded, outcome-driven engineering.
What is an FDE in AI?
In AI, FDE usually refers to an engineer who embeds with a company to get a general-purpose model working reliably on that company's specific task. That means writing evaluation suites and engineering context from the company's own data, then integrating the result into existing systems. OpenAI formalized this function in January 2025, and by mid-2026 Anthropic, AWS, and Salesforce had all built or announced similar functions.
How is a forward deployed engineer different from a consultant?
A consultant is typically paid for advice or a strategy document, and the engagement usually ends once that deliverable is handed over. A forward deployed engineer is paid to ship working code inside the client's own systems and stays engaged until a defined outcome is live, so the incentives sit closer to the client's actual result than to a report.
Can a small or mid-sized company hire a forward deployed engineer?
Yes, though rarely by hiring one full time at the roughly 220,000 to 400,000 dollar fully loaded cost a senior FDE commands in the US market. Most mid-sized firms instead buy a scaled-down version of the same model, often sold as an embedded AI engineer engagement, scoped to one workflow or department rather than an entire enterprise.