The waiter is part of the kitchen: what a Forward Deployed Engineer actually is, and why every AI vendor suddenly wants one

A Recrucial breakdown of the fastest-growing engineering role in enterprise AI: where the demand comes from, where the job was invented, what the work involves, what it pays, and how an engineer builds a career in it.

Forward Deployed Engineer - Pay by market tier

The EPAM number that deserved more attention

On 28 July 2026, EPAM Systems announced it had joined the OpenAI Partner Network as an Advanced Partner. As partnership press releases go, it looked routine, and most people skimmed it as such. The interesting part sits a few paragraphs down, and it is one of the clearest hiring signals the enterprise AI market has produced this year.

EPAM is committed to certifying more than 5,000 consultants in the first year, carrying more than 10,000 credentials between them. What those consultants are being certified as is a forward deployed engineer, a job title that sits oddly next to "AI specialist" or "prompt engineer" because it describes where a person sits rather than what they know. EPAM has roughly 56,500 delivery professionals, which makes this about one engineer in eleven being retrained into a role that was barely a mainstream job title three years ago.

The reason it matters is that EPAM is nowhere near alone. Look at what surrounds it.

OpenAI launched its Partner Network on 14 June 2026 with $150 million behind it and a stated goal of 300,000 certified consultants by the end of the year. A month earlier, in May 2026, it had already gone further and set up the OpenAI Deployment Company with more than $4 billion of initial investment, majority-owned by OpenAI and co-led by TPG alongside Advent, Bain Capital, and Brookfield. Its stated function is to embed forward deployed engineers inside customer organisations. In the same move, OpenAI agreed to acquire Tomoro, a UK applied-AI consultancy, which brought roughly 150 experienced forward deployed engineers and deployment specialists in a single transaction. Earlier in the year, it had announced Frontier Alliances, multi-year partnerships with BCG, McKinsey, Accenture, and Capgemini, with its own FDE team working alongside them.

Salesforce has publicly named a target of around 1,000 FDEs to staff Agentforce deployments. Anthropic posts forward-deployed engineering roles at $200,000 to $300,000 and lists "FDE" explicitly as a qualifying background. Databricks, Ramp, Glean, Scale AI, Sierra, and Mistral all run named FDE functions.

The aggregate data agrees with the anecdotes. Independent analyses put the growth in FDE job postings somewhere between 700% and 1,165% year on year, depending on the dataset and the window, and one analysis of 1,000 postings recorded October 2025 as the highest month on record. In mid-2026, roughly 1,200 live FDE postings were tracked across 669 companies, against a few dozen two years earlier.

So the honest reading of the EPAM announcement is that a large services company with 3.7% organic revenue growth in Q1 2026 has decided that retraining 5,000 engineers into forward-deployed engineers is its most credible growth lever. Its stock rose about 12.5% on the news, so the market agreed with it.

Which leaves the question worth asking. Why this role, and not "AI engineer" or "ML engineer", as the thing everyone is suddenly buying?

Where the demand actually comes from

The answer sits in a place the AI industry would rather not look, because access to the model stopped being a competitive advantage.

Three years ago, having a frontier model was the differentiator. Now any enterprise can buy API access to several of them in an afternoon, and what separates a company getting value out of AI from one that is not has moved away from model choice towards something much duller: whether the thing runs inside the business, on the business's real data, inside its real permission structure, used by people who have not quietly gone back to the old spreadsheet.

That is the last mile, and it has turned out to be considerably harder than the modelling. One forward-deployed engineer described the work more precisely than any analyst has managed, saying that the model is usually the cleanest part of the job and the hard part is finding the workflow nobody documented, the data source people actually trust, and the person who remembers why the process works the way it does.

This is why pilots stall, and the model is almost never the reason. What happens instead is that nobody could get the integration approved, or the data quality was worse than anyone had admitted, or the security team said no to the architecture in month three, or latency made the tool unusable at the actual moment of work, or the cost per request turned out to be indefensible at volume, or the intended users simply did not adopt it. All of those failures happen inside the customer's environment rather than the vendor's, which is the part that decides everything else.

Vendors worked this out and drew the obvious conclusion, which is that the problem cannot be solved from outside. Somebody has to sit in the room. And since no frontier lab can staff thousands of enterprises with its own people, OpenAI built a partner network, put $4 billion into a deployment company, bought a consultancy for its 150 engineers, and set itself a target of 300,000 certified consultants. It is the channel strategy that built AWS, Microsoft, and Salesforce, applied to a new problem: stop trying to deliver the last mile yourself, build an army of partners who do it, keep the platform revenue.

The forward-deployed engineer is the human unit of that strategy, which is why demand for the role has several years left in it and why it is structurally unlike the prompt-engineer bubble that came and went. Prompt engineering was a technique. This is an operating model.

Where the profession came from

The role was invented at Palantir, and the origin story is worth knowing because it explains the shape of the job.

Palantir was founded in 2003 and spent its early years building software for the US intelligence community. Between 2005 and 2008, the CIA was effectively its patron and its only customer, with In-Q-Tel, the intelligence community's venture arm, investing more than $2 million across two rounds. Palantir's engineers had no idea how those customers actually worked, and the customers could not have written a conventional requirements document even if they had wanted to. Co-founder Joe Lonsdale has been refreshingly unromantic about what followed, saying the model was adopted "by necessity." His argument is that the product playbook works when you solve a pain point shared across many customers, and they needed something closer to a special-operations services mentality. The line to remember is his summary of the economics: "services have supported product revenue, not vice versa."

Shyam Sankar, Palantir's thirteenth employee and now its CTO, is credited with creating the role somewhere around 2006. According to a first-hand account from one of the early hires, Sankar named it as a nod to the company's earliest customers, borrowing the military vocabulary of forward-deployed forces that those customers used about themselves. The earliest public use of the term that can be verified is Bloomberg Businessweek on 22 November 2011, which noted that in place of traditional salespeople, Palantir had what it called forward deployed engineers.

The explanation Palantir people actually give each other has nothing military about it. It is attributed to Alex Karp, and it is about restaurants. In a French restaurant, the waiting staff belongs to the kitchen staff, because they understand the food, the method, and the technique. They are not carrying plates between two rooms. They are part of a system in which what the room sees changes what the kitchen cooks. Palantir's commercial lead, Ted Mabrey, compresses the same thought into a sentence about the delivery mechanism having to be opinionated and to own that.

If you want one reason why an engineer sits inside a customer rather than visiting it, that is the reason, and it is not the speed of implementation. What gets built is determined by what the person in the room can see.

Two things circulate widely about all this, and both are wrong. Shyam Sankar never wrote an essay called "The Rise of the Forward Deployed Engineer"; that headline belongs to later authors and has been retro-attributed to him. And Palantir's internal nickname for the role, "Delta," has nothing to do with Delta Force. The company explained it publicly in 2019: in the early days, each business development team was named after a letter of the NATO phonetic alphabet, and Delta became forward-deployed engineering while Echo became the Deployment Strategist role.

Who picked it up, and how the consultancies run it

For roughly a decade, the model stayed largely inside Palantir. Then in 2025 the frontier labs copied it, having arrived at the same conclusion Palantir reached in 2006, and the services industry followed within months.

The consultancy version differs from the vendor version in one structural way, because the consultancy sells hours, so an FDE is a billable unit rather than a cost of customer acquisition. That changes the incentives, and the change shows up in how the role gets titled and leveled.

EPAM is the clearest current example, with a partner-network tier, a mass certification programme, and a proof point to put in front of buyers. The one it uses is a telecom, 1&1, where a customer-service transformation went live with more than 20 intelligent agents in under three months. That is a good pitch, and it fits in one sentence.

Deloitte posts a role called Lead Forward Deployed Engineer, Palantir with a band of $189,200 to $372,900 and a job description that reads as small-scale engineering leadership: lead pods of two to five engineers, architect and oversee delivery of LLM-enabled applications, govern end-to-end RAG pipeline design, review and contribute to production code, seven years of experience and up. It also pays more than most vendor FDE roles do.

PwC runs the same function under consulting grades, posting Forward Deployed Software Engineer at Senior Manager level. Accenture posts both "Palantir Forward Deployed Engineer" delivery roles and Associate Manager grades, and is at the same time an OpenAI Frontier Alliance partner. Capgemini, Thoughtworks, Zühlke, Xebia and a long tail of European boutiques are building comparable practices.

There is a trap in this that hiring managers should know about, and it is the single most common mistake we see when clients read CVs. A large share of the job postings that say "Palantir Forward Deployed Engineer" were not posted by Palantir. They come from Deloitte, Accenture, PwC, Inabia, ConSol and various boutiques staffing Foundry implementations. In the same way, "Senior," "Lead," "Principal" or "Architect" in front of Forward Deployed Engineer almost always signals the partner ecosystem, because Palantir itself uses no seniority prefixes at all. Its job postings say as much, describing a company that celebrates individual strengths, skills and interests instead of relying on traditional career ladders. There are no grades to put in front of the title.

Which means the title on a CV tells you far less than the employer does.

What a Forward Deployed Engineer actually does, and what clients ask for

Strip the vocabulary away and a real forward-deployed engineer meets three conditions at once. Miss any one of them and the job is something else.

One: they write production code. Python to a production standard, integrations, APIs, backend. OpenAI's own posting asks for someone highly proficient in Python, not familiar with it.

Two: they work with an external customer who has the right to say no. Not an internal stakeholder, but a customer with a procurement process, a security team and an opinion.

Three: they took something into daily operation. Not a demo and not a proof of concept, but a system people use every day, and they were still there when it broke.

Held against those three conditions, the profiles that get mistakenly proposed sort themselves out fast. A backend engineer with a RAG chatbot on GitHub satisfies the first. An ML engineer satisfies the first and sometimes the third, but has never worked inside somebody else's landscape. A sales engineer satisfies the second and not the first. An AI consultant with a strategy deck satisfies only the second, and in 2026 this is by a distance the most common substitution.

What clients specify, taken from live postings at OpenAI, Mistral, Databricks and Anthropic, clusters tightly:

• The full application stack rather than the model layer alone: APIs, backend, and enough frontend to ship an interface someone can use.

• RAG, agentic workflows, function calling and agent orchestration as things the candidate has designed, not things they have read about.

• Systematic evaluation. OpenAI's wording is precise enough to be worth borrowing: the ability to evaluate AI systems systematically using representative data and graders. It is the sharpest technical filter in the whole profile.

• Enterprise production reality: integrations, reliability, observability, security, privacy, governance, performance, and cost.

• Three registers of communication: engineers, security teams, executives.

• High agency and end-to-end ownership in ambiguity, which in plain language means somebody who needs neither a team lead nor a defined task.

• Travel and presence. Real postings ask for 20% to 50%, and Palantir's Deployment Strategist role asks for 25% to 75%.

Two items on that list deserve more weight than the rest, because they are where good candidates separate from plausible ones.

Evals are the tell. People who have run LLM systems in production talk about golden datasets, graders, regression suites and measuring degradation before release, while people from the demo world talk about prompts and model choice. The vocabulary difference is close to perfectly diagnostic and it takes about ninety seconds to detect.

Willingness to discuss the boring parts works the same way. Access rights and RBAC, PII handling, data residency, latency budgets, cost per request, observability, what happens when the model is unavailable, how a conversation gets handed to a human. A side project never generates any of these problems, so the words cannot turn up in someone's vocabulary by accident.

Inside the role there is a useful split that Palantir formalised and most vendors have since reinvented. Palantir's Delta, the Forward Deployed Software Engineer, owns the code and the technical architecture. Its Echo, the Deployment Strategist, owns which problem gets solved and whether anyone actually uses the result. Palantir's own framing of the difference is hard to improve on: a product engineer works on "one capability, many customers," a forward-deployed engineer on "one customer, many capabilities."

And then there is the proportion nobody puts in the job ad. Ask a working FDE how much of the job is not about AI and the honest answer runs from 60% to 80%: integrations, access, legacy systems, approvals, and teaching people to work differently.

What it pays

Compensation here is unusually bifurcated, and getting the tier wrong is the most common reason a search fails before it has properly started.

Segment	Range	Note
Global FDE median base	$173,816 to $185,000	Three independent datasets, 1,000+ postings each
OpenAI, AI Deployment Engineer, Enterprise (SF)	$197,000 to $278,000 plus equity	Posted band
Anthropic, Forward Deployed Engineer, Applied AI	$200,000 to $300,000	Posted band, US
Palantir FDSE	$135,000 to $200,000 posted; median TC around $211,000	Posted band plus crowdsourced TC
Palantir FDSE / OpenAI, market view	$170,000 to $340,000+ total	Elite programmes
Deloitte, Lead Forward Deployed Engineer	$189,200 to $372,900	Consultancy leadership grade
Glean, Founding FDE	$160,000 to $270,000	Posted band
Anthropic, Applied AI Architect Lead, Dublin	€215,000 to €260,000	Posted band, includes variable
Anthropic, Applied AI Architect, London	£190,000 to £230,000	Posted band
UK AI Architect, permanent median	£100,000	ITJobsWatch, up 25% year on year
UK contract day rate, AI Architect	£650 per day median	ITJobsWatch, up 8% year on year
Netherlands, consultancy architect	€80,000 to €105,000	Posted band
Netherlands, ZZP IT architect	€115 per hour excluding VAT, roughly €920 per day	Knab, sample above 20,000 freelancers
Germany, Solution Architect median	€69,800	StepStone, generalist market

Three things follow from those numbers.

European bands run at roughly 50% to 70% of their US equivalents, and the competition is not European. OpenAI, Anthropic, Palantir, Scale and Databricks all have offices in London, Dublin, Paris, Amsterdam and Munich, and they pay vendor-tier money in all of them. An employer budgeting from the German generalist architect median of €69,800 is planning to lose every vendor-sourced candidate it meets, and the Anthropic band in Dublin is three times that figure.

Equity rather than commission. Across 1,000 FDE postings, 70% mention equity, 8% mention commission or OTE, and none carry a sales quota. This is an engineering role that sits near revenue rather than a sales role that happens to code. Candidates care about the distinction a great deal, and pitching it the wrong way loses them in the first message.

Ravio data puts the European AI skills premium at about 12% at individual-contributor level. That is real, and it is much smaller than the gap between segments. Which tier you are hiring in matters more than whether the person has AI on their CV.

How an engineer gets into the role, and how the experience gets built

If you are the engineer rather than the person hiring one, the requirements here are unusually concrete, which makes this easier to plan than most AI career advice allows.

The four realistic entry routes

From product engineering into a customer-facing role. The most common path, and counter-intuitively the strongest signal, because most engineers move in the opposite direction or never move at all. A deliberate step towards customers is direct evidence of the personality the role needs, and vendors hire from here readily. Mistral's EMEA forward-deployed posting sets the bar at two or more years as a technical individual contributor, and Palantir has hired FDSEs with as little as a year of post-university experience. The formal bar is low. The real one is not.

From solutions or customer engineering, by adding production depth. A solutions engineer, customer engineer or solutions architect already has the harder half of the job. What is missing is evidence of having shipped and operated an LLM system rather than demonstrated one. Mistral lists prior Customer Engineer, Forward Deployed Engineer, Sales Engineer, Solutions Architect and Technical PM experience as a plus.

From integrator delivery. EPAM, Accenture, Deloitte, Thoughtworks, Xebia, Zühlke. This is the widest pool in Europe and the fastest way to accumulate several customers, several industries and real legacy exposure. It is also where title inflation is worst, so the burden of proof on outcomes sits higher.

From in-house deployment, which is the underrated route. Engineers who put AI into production inside their own large company are doing the same work: hunting undocumented processes, negotiating with internal system owners, driving adoption. They rarely call themselves forward-deployed engineers, so fewer recruiters find them and competition for them is lower. If that is your path, the highest-return thing you can do is describe your work in the market's vocabulary.

What to accumulate deliberately

Five things are worth engineering into your own CV over the next two years, and they map almost exactly onto what interviewers probe.

1. One production deployment you owned through the first bad month. Not the launch, but the fortnight afterwards, when data quality, permissions and user resistance all arrive together. It is the single most valuable story you can carry, and it cannot be fabricated.

2. An evaluation harness you built. A golden dataset, graders, a regression run before release, and a business metric you moved. Being able to explain how you detected degradation separates you from most of the market.

3. Numbers on cost and latency. What one request cost in production, what the latency budget was, and what you changed to fit inside both: caching, routing to a smaller model, context reduction. Almost nobody from the side-project world can answer this, which is precisely why it gets asked.

4. One security or compliance constraint that forced a redesign. PII handling, data residency, audit, tenant isolation, SSO. "Security said no, so we rebuilt it this way" is a strong marker.

5. One reusable artefact. A template, an internal framework, an integration pattern another team adopted. This is what separates a senior FDE from an architect-tier one, and it is what unlocks the next salary band.

Where to get that experience

In rough order of speed, an integrator with genuinely hands-on delivery gives you the most customers per year. A vendor field organisation, whether that is a Databricks Resident Solutions Architect role, the applied teams at OpenAI or Anthropic, Mistral or Parloa, gives you the deepest platform work and the best pay. An in-house AI platform team at a large enterprise accumulates the slowest, but it produces the most authentic legacy and adoption experience and it is by far the easiest to reach from a conventional engineering job.

One warning about credentials. OpenAI's target of 300,000 certified consultants means that within about a year an OpenAI certificate will differentiate nobody, in the same way AWS certification stopped differentiating anybody. Collect it if your employer pays for it, and do not mistake it for the asset. The asset is the production deployment you owned, the eval harness, the cost and latency numbers, the security redesign and the reusable artefact.

Where these people come from, and where they go next

The talent flow here is unusually well documented, largely because Palantir alumni have been unusually visible.

Estimates of the "Palantir mafia" vary with methodology, from 48 companies to 379, so any single number should travel with its source attached. The most striking observation comes from Nabeel Qureshi, who spent eight years there and points out that a given Y Combinator batch usually holds more ex-Palantir founders than ex-Google ones, despite Google having roughly fifty times the headcount.

The destinations say what the experience is worth. In defence, there is Anduril, co-founded by Palantir alumni Brian Schimpf, Trae Stephens, Matt Grimm, and Joe Chen, and Peregrine, founded by the person who ran Palantir's special operations business and valued at $6.8 billion in June 2026. In enterprise software, there is Hex, whose three founders all worked together at Palantir, along with Sourcegraph, Ironclad, and Distyl AI, founded by two people with roughly a decade there each, which raised $175 million at a $1.8 billion valuation and sells software plus forward-deployed engineers as the explicit product.

In Europe, which matters for anyone hiring here, the alumni network is concentrated in London: ElevenLabs, whose co-founder Mati Staniszewski was a Palantir deployment strategist, Conduct, founded by three ex-Palantir engineers and backed by a $60 million Series A co-led by Index Ventures and ICONIQ, plus Arondite, Ankar, Perceptic, Fern Labs and Valarian, with H in Paris now led by the former CEO of Palantir France. Index Ventures appears in four of those, which makes it the densest single node in the European network.

For a sense of pool size, Palantir's UK entity reported 749 employees at the end of 2024, down 11% year on year. The European alumni pool is measured in hundreds rather than thousands, and it is already covered by dense personal networks. Average tenure at Palantir runs around 3.2 years, though the people who leave to found companies cluster at five to seven.

The lesson for hiring managers is simpler than the list of names. This role is the best training the market currently offers in translating between technology and business under pressure, which is why people who do it well rarely stay individual contributors for long. A16z partner Marc Andrusko described the profile as engineers who are comfortable writing production code, dealing with bureaucracies, and sitting in rooms with colonels, CIOs and regulators, and called many of them unicorns for being both highly technical and highly effective with customers. If you hire one, plan for what they do next, or they will plan it without you.

Where is this heading

Four predictions, each grounded in something observable now.

The role will split in two, and the industry will keep confusing them. Pre-sales architects who help close deals, and delivery engineers who own production. Anthropic already runs both, and its titles actively mislead, because its Applied AI Architect is a pre-sales role while its actual forward-deployed role is Applied AI Engineer. Anyone hiring in this space has to decide which of the two they want before writing the job ad, since they attract different people with different motivations and different compensation structures.

An architect tier is emerging, and it will not be called Architect. Palantir, NVIDIA, ServiceNow, Atlassian and Workday have all used "Forward Deployed Architect," but it accounts for roughly 2% of forward-deployed postings and in Europe it barely exists. The industry is naming that layer Staff, Lead or Principal Forward Deployed Engineer instead, and the published career ladders for the role carry no architect rung at all. Nor is there a pay premium attached to the word: architect-titled bands sit at or barely above the FDE median, and Deloitte's Lead FDE band pays more than any architect-titled role we have found.

Certification will stop meaning anything, and outcomes will matter more. Once 300,000 people hold the certificate, the filter moves back to what somebody actually shipped.

And a meaningful share of the current enthusiasm will disappoint, for a reason worth taking seriously. Andrusko's warning is the sharpest version of it: companies that copy the model without the underlying discipline end up as expensive services businesses carrying a software valuation multiple and no compounding advantage. His phrase for the failure mode is hard to improve on: "You're Accenture for X' with a nicer front-end." Anaplan's CEO, Charlie Gottdiener, put the operational objection more bluntly, calling it a good selling model and a poor model for running software. And Manik Sharma, now at Kinaxis and formerly a Palantir executive, named the staffing failure directly, which is that putting 25-year-old engineers in front of the customer is a problem in itself.

The best single criterion for telling a real forward-deployed programme from a consultancy in disguise comes from Kevin Bai, now at Anthropic, previously at Palantir and a founding FDE at Rippling, who says the thing that makes an FDE programme different is that "they are never writing software from scratch" because they are always building on top of a platform. The discipline that follows from it is the useful part: anything bespoke to one customer should exist only for that customer, and anything generalisable should eventually be generalised.

That also happens to be the best question to put to a candidate. What did you keep bespoke, what did you push into the platform, and how did you decide?

How we think about this at Recrucial

We take a clear position on this role, because clients and candidates both get burned by the vagueness around it.

With clients, we calibrate before we search. Four questions decide whether a search is findable at all: is this pre-sales or delivery, does the person write code, how many days a week are they physically at the customer, and is a local language required. In DACH and France, customer-facing roles require the local language in roughly half to two-thirds of comparable postings, and clients forget this more often than any other constraint. We would rather have that conversation at the briefing than after the third rejected offer.

With candidates, we are specific about what we can and cannot promise. The people worth hiring in this market are rarely unhappy with their work. They move because grade and pay progression are undefined where they are. Anyone approaching them with "interesting problems" is wasting a message.

And we screen on three things rather than on titles. Production code, an external customer with the right to refuse, and a system that reached daily operation. Everything else is negotiable. Those three are not.

If you are hiring for this profile in Europe, or you are an engineer trying to move into it, we are happy to talk about what is realistically available at your level and in your market. That conversation is usually short, specific, and more useful than a job description.

Recrucial is a recruitment company working with engineering and AI roles across Europe. This article draws on public job postings from OpenAI, Anthropic, Palantir, Mistral, Databricks, Deloitte, Glean, and Scale AI; EPAM's July 2026 partnership announcement; market analyses of FDE hiring volume and compensation; ITJobsWatch, StepStone, Glassdoor, levels.fyi, Ravio and Knab compensation data; Companies House filings; and published accounts from current and former forward-deployed engineers, including Nabeel Qureshi, Barry McCardel, Ted Mabrey, Kevin Bai, Shilpa Balaji, and Vinoo Ganesh. Figures are current as of July 2026. Where sources disagree, we have given the range rather than picking a number.