All posts Forward-Deployed Model

What Is a Forward-Deployed Engineer (and Why Enterprise AI Needs One)

What is a forward-deployed engineer?

A forward-deployed engineer (FDE) is a software engineer who embeds directly inside a customer’s environment, their repo, their standups, their Slack, their real data, to build and ship production software against the customer’s actual problems, rather than gathering requirements from a distance. First Round Review puts the job in one line: the FDE “embeds directly with a customer to build the ‘last mile’ of the product to work in production” (First Round Review).

For a company putting AI to work, the FDE is the person who closes the gap between an impressive demo and a system that runs in production, on your data, under your governance, inside your workflows. That gap is where AI budgets stall, and it is a delivery problem, not a model problem.

Key takeaways

  • A forward-deployed engineer embeds in your environment: repo, standups, Slack, real data. The output is production code, not slide decks or prototypes.
  • The model started at Palantir in the early 2010s (The Pragmatic Engineer). Palantir’s “AI FDE” is a Foundry agent, not an engineer in your building.
  • Expertise, not interest, is what stops European companies. In 2025, among EU enterprises that had ever considered AI and did not adopt it, 70.89% named lack of relevant expertise as a reason (Eurostat).
  • There is no market price. No AI lab publishes one. The rate cards a buyer can read are on the UK’s G-Cloud 15 framework, Microsoft’s Frontier Co-Engineering service among them (Digital Marketplace).
  • Half the team is yours, and that half decides whether the system outlives the engagement.
  • The question to ask any FDE provider is what stays with you when the engineers leave.

Why enterprise AI needs a forward-deployed engineer

The technology is rarely what stalls an AI programme. Delivery is.

Eurostat’s 2025 survey of EU enterprises puts a number on it. Among companies that had ever considered AI and decided not to use it, the most common reason was lack of relevant expertise (70.89%), followed by lack of clarity about the legal consequences (52.52%) and concerns about data protection and privacy (48.83%) (Eurostat). Only 20.68% said the technology was not useful to them. One of the top three blockers is skills; the other two are legal clarity and data protection, which are governance questions. None of them is solved by a better model.

The same survey shows how uneven adoption already is. In 2025, 19.95% of EU enterprises (10+ employees) used AI, against 55.03% of large enterprises. The Nordics lead: Denmark at 42.03%, Finland at 37.82% and Sweden at 35.04% (Eurostat). For a mid-cap in the EUR 250M to 2B range, below the Fortune 500 but with real data and real process complexity, the gap between “we use AI” and “AI runs a workflow that moves a number” is mostly an engineering bench the company does not have.

That is where a pilot stalls. The demo worked. The pilot got budget. Then it met real data, real permissions, real edge cases and real users who don’t behave like the sandbox, and nobody inside the company had the time or the skills to carry it the rest of the way. The pattern is covered in detail in why enterprise AI pilots fail.

A forward-deployed engineer exists to carry it the rest of the way. Instead of shipping a generic platform and hoping your team integrates it, the FDE sits inside your operation, sees where the process actually breaks, and builds the system to survive that reality, under your governance, on your data.

Where the FDE model came from: Palantir’s origin story

The FDE didn’t start as a hiring trend. It started as a necessity.

Palantir created the role in the early 2010s (The Pragmatic Engineer). Its intelligence-agency customers often couldn’t openly describe what they needed: the requirements were classified, tacit, or simply unknowable from a conference room. So instead of running a traditional requirements-gathering cycle, Palantir put engineers directly inside customer environments. Those engineers learned by observing, experimenting, and building in real time alongside the people doing the work (FDE Academy).

In 2011 Palantir retitled its solutions and integration engineers forward-deployed engineers and called them Deltas (Andreessen Horowitz). Its own description, quoted by The Pragmatic Engineer, contrasts a developer’s “one capability, many customers” with a Delta’s “one customer, many capabilities”, and states that “Deltas are part of Business Development”. That placement is worth remembering when a Delta sits across the table.

The result was a feedback loop that traditional enterprise software vendors structurally cannot match. The engineer who feels the pain is the same engineer who ships the fix.

The model maps closely onto AI work today. An AI system is not shrink-wrapped software. It has to be tuned to your language, your data, your policies and your exceptions. You cannot spec that from the outside. Someone has to be inside.

Palantir’s “AI FDE” is a product, not a person

The term now has a second meaning, and buyers mix them up. Palantir sells a feature called AI FDE. Its documentation describes it as an interactive agent that operates Foundry through conversational commands, translating natural-language requests into Foundry operations such as data transformations, code repository changes and ontology edits, within the user’s existing permissions (Palantir documentation).

That is a useful tool for a Foundry customer. It is not a forward-deployed engineer. It works inside one vendor’s platform, on that platform’s objects, and it cannot sit in your standup and notice that the approval step everyone describes is not the one anyone follows. When a proposal mentions FDEs, check which of the two you are being offered.

The forward deployed engineering model

Forward deployed engineering is a delivery model. Four things distinguish it from a project staffed the usual way.

The work happens in your environment. Engineers commit to your repositories, run under your identity and access controls, and build against production data. There is no separate vendor tenant where the system works and yours where it doesn’t.

The team is small and stays close to one workflow. The unit is a workflow with an owner and a number attached, not a programme with a steering committee. Scope grows when a workflow is in production, not before.

Success is measured in production. A requirements document, a roadmap or a proof of concept is an intermediate artefact. The engagement is judged on whether the system runs and whether the number moves.

The capability transfers. This is the part the model’s popularity has blurred. An FDE engagement can end with your team running the system, or with you renting it back from the provider. What decides that is structural: whose platform the system runs on, whose model it depends on, and who owns the data, prompts, evaluations and decision records it produced.

The model is now sold by very different organisations. AI labs offer it as a route into their models, as covered in OpenAI’s DeployCo and Anthropic’s Ode. Large systems integrators sell it alongside vendor partnerships, as covered in who really does forward deployed engineering now. The role is the same in each case. Who owns the result is not.

Six buyer questions follow, each answered after the roles table: how an engagement runs, who is on the team, what you provide, what stays with you, what it costs and how to judge a partner.

What a forward-deployed engineer actually does

An FDE is not a solutions consultant, a customer success manager, or a sales engineer. The distinction is that an FDE writes and ships production code. Concretely, a forward-deployed engineer:

  • Embeds in your environment: commits to your repo, joins your standups, works in your Slack, and builds against real production data rather than synthetic test sets.
  • Diagnoses the real workflow, including the undocumented exceptions and manual workarounds that never make it into a requirements doc.
  • Builds and ships the system into production under your existing governance, identity and access controls, not a walled-off demo tenant.
  • Instruments everything, so you can see what every automated step did, what it cost and what it changed.
  • Stays through the hard part: the last edge cases and integration friction, which is where pilots quietly stop.
RolePrimary outputLives in your codebase?Owns pilot to production?
Forward-deployed engineerShipped production codeYesYes
Solutions architectAdvice, MVPs and proofs of concept on anonymised or offline cuts of dataRarelyNo
Sales engineerDemos, POCsNoNo
Solutions consultantRecommendations, configRarelyNo
SaaS platform vendorGeneric productNoNo (you integrate)
Vendor “AI FDE” featureOperations inside one platformOnly within the vendor’s platformNo
Internal ML teamModels, infraYesSometimes (capacity-limited)

The solutions architect row is OpenAI’s split of the two roles, as The Pragmatic Engineer describes it: solutions architects “rarely write code on customers’ infrastructure”, while FDEs “write code directly on customer infrastructure and use customer tooling for this”. Consultants and professional services are compared in forward-deployed engineer vs consultant vs system integrator.

How does a forward deployed engineering engagement work?

Most published sequences run from a short discovery or observation phase to a build on real data and a production or licence decision. Our recommendation is narrower: one workflow and one number agreed before the build, then hand over and widen once it runs in production.

Providers describe their own sequences, and none of them is a guarantee:

ProviderWhat it publishes about the sequence
PalantirPilots and bootcamps “generally at our own expense”, before a licence (Form 10-Q)
AWSIts method, as Forrester reports it: 45 minutes to an idea, 45 hours to a prototype, 45 days to a customer engagement
Accenture with SAPDiscovery sprints, process re-imagination, data integration, rapid AI buildout and production activation (Accenture, 8 June 2026)
Manhattan Associates“a 90-day proof of concept, and it will come with forward-deployed engineers” (earnings call, quoted by Constellation Research)
Uber, in-houseEngineers watched the work for two weeks at a time before automating it (Forrester)

Two rows need more context. AWS says customer engineers move “from observers to co-builders to autonomous operators” as projects advance (Amazon). Palantir’s bootcamp, and the paid steps that follow it, are covered in what Palantir AIP costs.

The Uber line is the one buyers skip. Its CTO, Praveen Neppalli Naga, gave the reason. As Forrester paraphrases him, you cannot automate these processes by looking at process diagrams or documentation. Phase one is watching, not a requirements workshop. A generic example, not a client case: engineers sitting with a credit-control team find that one person settles most disputed invoices by checking delivery notes, a step the process map does not show. The build starts there.

Nucleo runs four steps, set out on our services page: one workflow and one number, engineers inside your team, production rather than a pilot, then wider. Finding the first workflow is covered in AI use case discovery, the build order in the build sequence that holds, and the route to production in the four gates.

Who is on a forward deployed engineering team?

A small pod from the provider and named people on your side, and your side is the half that decides whether the system outlives the engagement.

Published pods are small. A Deloitte advertisement for a lead FDE describes pods of 2 to 5 engineers, anchored onshore and supported offshore (Deloitte, archived June 2026). CNBC reports that AWS embeds an initial pod of roughly five or six engineers at a customer, working alongside AI agents (CNBC). Sean Catlin of OSF Digital, who manages forward-deployed teams that implement Salesforce, describes them as often “no more than a half-dozen experts”, while traditional consultants “sometimes operate with larger teams” (SD Times). The comparison is set out in forward-deployed engineer vs implementation consultant.

Provider sideClient side
Engineers who write the code in your stack (Palantir’s Delta role)The process owner, who owns the number
Someone who finds and frames the problem before the buildThe engineer who will run the system, named in week one
A lead who directs the provider’s engineers and owns deliveryIT and security, for identity, access and the route to production
In some pods, offshore engineersThe data owner for each source the workflow reads
In AWS’s version, AI agents doing part of the buildThe works council, where one exists

Ask where the engineers sit: what offshore staffing changes is covered in what offshore scale changes. The client column is the one provider material leaves vague, and it is where handovers fail, as where embed fails sets out. Name the internal owner in week one, as the build sequence says, not at handover.

What does the client have to provide in an FDE engagement?

A measurable workflow with an owner, access to systems and real data, time from the people who hold the logic, signed processing terms and, in some cases, named oversight people and an informed works council.

  1. A workflow with a number and an owner. Deloitte’s first criterion for an FDE model is a business problem “with success defined through clear, measurable criteria” (Deloitte).
  2. Access under your own rules: the systems and real data through your identity and access controls, plus time from your business, engineering and security staff, the people CNBC reports AWS’s FDEs partner with (CNBC).
  3. The people who hold the logic. Forrester finds it in “spreadsheets, exception handling, and the heads of people who have been there 15 years or more” (Forrester). Budget their hours as you budget the engineers’.
  4. A processor contract before personal data moves; what it must cover is set out in AI use case discovery.
  5. Named oversight for a high-risk system. For a high-risk system listed in Annex III of the EU AI Act, Article 26(2) requires deployers to assign oversight to people with “the necessary competence, training and authority”, from 2 December 2027 (2 August 2028 for high-risk AI in products covered by Annex I). The design side is in human oversight for AI agents.
  6. In Germany, an informed works council; the duty to inform it and its co-determination rights are covered in the same discovery section. Less known is section 80(3) of the Works Constitution Act: when the council has to assess AI, “it is deemed necessary to call on the advice of an expert”, and section 40(1) says its expenses “are defrayed by the employer”. Budget for the expert.

One term belongs in the contract instead: who directs the engineers day to day. It is a central test of whether the engagement is a service or labour leasing, as set out in who carries the delivery risk. This is a list of points to raise, not legal advice; take the specifics to counsel.

What stays with you after an FDE engagement?

What stays is whatever the contract names, in places you control, able to run on a different model. Vendors now publish what they leave: AWS lists “deployed systems, knowledge graphs, runbooks, architectural documentation, and trained internal champions” (Amazon). The open questions are where those live and what they run on. AWS puts its semantic layer in the customer’s own AWS account, which makes it yours on AWS, a distinction tested in self-sufficient is not the same as portable. Forrester’s precedent is Pivotal Labs, which embedded with clients until the client could run the practice “without us” (Forrester). That happens only if the handover is designed in the first week.

The contract terms are written up elsewhere: the seven-item list in what a forward deployed engineering team should leave behind, and the five layers an engagement accumulates in ownership is the question, not model choice. What a Nucleo engagement leaves is a context engine in your estate: the owned, model-agnostic layer of your company’s knowledge and decision logic that your systems and agents read from (how we work).

How much does forward deployed engineering cost?

There is no market price. Public sources show three ways the cost is recovered: inside a licence or a proof of concept, by role rate card, and by the month. Other providers describe outcomes and name no price. Only the rate cards are prices a buyer can read.

ProviderWhat is publishedPricing basis
PalantirPilots and bootcamps “generally at our own expense” (Form 10-Q)Licences; UK pricing also lists paid discovery packages and pilot licences (Palantir AIP pricing)
ServiceNowAfter FDE work, a fast-food chain’s Now Assist entitlement grew “by 13x upon contract renewal” (CFO, via Constellation Research)Not stated; licence expansion cited as the result
Manhattan AssociatesA 90-day proof of concept that comes with FDEs (same source)Bundled into the proof of concept
AWS“shared goals and business results, not billable hours” (Amazon)Outcome language, no price
OpenAI Deployment Company, Ode with AnthropicNo price (our comparison)Not published
Microsoft, UK G-Cloud 15 listingLead machine learning engineer £3,120.00, senior developer £2,320.00 (Digital Marketplace)Role rate card, beside a promise of “shared outcomes”
Elsewhen, UK G-Cloud 15 listingLead machine learning engineer £1,400.00, senior developer £1,300.00 (Digital Marketplace)Role rate card

The G-Cloud figures are UK public-sector list prices with no unit stated, so read them as relative, not as rates you will be quoted. Microsoft’s entry is Microsoft Frontier Co-Engineering for AI Transformation, a catalogue listing that “embeds multidisciplinary Microsoft Forward Deployed Engineering (FDE) capability alongside customer teams”. It is a UK public-sector rate card, not a price list for Microsoft Frontier Company.

One monthly figure is public, but a buyer does not pay it. A.Team, which supplies white-label FDEs to AI software companies, quotes those vendors “roughly $25,000 per month for a senior builder, fully loaded” (A.Team). The vendor pays it to staff its own customers’ rollouts.

Analysts warn on price and on substance. Constellation Research says FDEs “aren’t cheap” and may be “part of a broader vendor dependency play” (Constellation Research). Gartner’s warning about rebranded consulting is covered in who really does forward deployed engineering now. Compare cost per workflow in production, licence or consumption included, not the rate.

Hiring your own means competing in a thin market. Lightcast counted more than 7,300 FDE job postings in the first nine months of 2026, up from fewer than 200 in 2023. Gartner’s Mukul Saha told Channel Dive that recruiters are competing for a couple of thousand practising FDEs (both via CIO Dive).

Nucleo has no rate card. Our terms, on the services page, are a fixed fee to build, a recurring fee to run, and part of the fee tied to the outcome where it is a discrete event we can attribute. That makes us harder to compare on a spreadsheet, which is a fair objection.

How do you evaluate a forward deployed engineering partner?

Judge the contract and the people, not the label. Five checks settle most of it.

  1. Is the number agreed before the build, with a production date? That is the first test in what you provide above. A partner who will not commit to one is selling discovery.
  2. Which platform relationship funds the pod? The questions to ask, and when a platform’s own team is the better hire, are in what to put on the scorecard instead.
  3. Who directs the engineers day to day? Their own lead, not your line managers, for the reasons in who carries the delivery risk.
  4. Is the FDE lead also a sales role? Deloitte’s lead FDE also represents the practice in “client pursuits” and contributes to “deal shaping” (Deloitte, archived June 2026). Ask how much of the lead’s week goes to your build.
  5. What stays when they leave? Put the seven items in the contract before the work starts. Constellation Research sets the bar: FDEs “should be redesigning workflows and teaching the customer. Otherwise, FDEs are glorified integrators.”

When you need a forward-deployed engineer (and when you don’t)

You likely need an FDE model if:

  • Your AI initiatives keep stalling between promising pilot and production rollout.
  • Your workflows have significant undocumented complexity, exceptions, or legacy-system integration.
  • Governance, identity and EU data-residency requirements make “just use the SaaS tool” a non-starter.
  • You need measurable outcomes within a quarter, not a multi-year platform migration.

You probably don’t if:

  • Your use case is fully served by an off-the-shelf tool with no meaningful customization.
  • You have enough internal engineering capacity already embedded in the workflow.
  • You’re still validating whether the problem is worth solving at all (run a cheap experiment first).
  • Nobody on your side can be named to receive the system. Embed fails when there is no owner to hand it to.

Whether the economics work is a separate question with its own evidence. The sourced figures, and what they do and don’t measure, are in the ROI of enterprise AI agents.

The bottom line for decision-makers

In 2024 the question was “which model?” Now it is “why is our pilot still a pilot?” The forward-deployed engineer is one answer: the role that embeds, takes on the operational detail and ships a system running on real data, under real governance.

The role alone does not settle who benefits. Models change every few months. What lasts is the workflow knowledge, the data and the decision logic built into the system, and an FDE engagement either leaves those inside your company or inside someone else’s platform. Ask which one before you sign.

Frequently asked questions

What is a forward-deployed engineer in simple terms? A forward-deployed engineer is a software engineer who works embedded inside a customer’s team and systems to build and ship production software against that customer’s real problems, rather than handing over a generic product and leaving integration to the client.

What’s the difference between a forward-deployed engineer and a sales engineer? A sales engineer builds demos and proofs of concept to help close a deal. A forward-deployed engineer ships production code that runs on real data under real governance. The FDE owns the move from pilot to production; the sales engineer typically hands off before it.

What is the difference between a forward deployed engineer and a solutions architect? A solutions architect advises and builds proofs of concept, usually on anonymised or offline cuts of data, and rarely writes code on the customer’s infrastructure. A forward-deployed engineer writes code directly on that infrastructure, with the customer’s tooling. That is OpenAI’s split of the two roles, as The Pragmatic Engineer describes it. The forward-deployed engineer also owns the move to production.

Why does enterprise AI specifically need forward-deployed engineers? Because the bottleneck is delivery, not models. In 2025, among EU enterprises that had ever considered AI and did not adopt it, Eurostat found the most common reason was lack of relevant expertise (70.89%). A forward-deployed engineer brings that expertise into the company’s own environment and builds the capability there.

Where did the forward-deployed engineer role come from? Palantir created it in the early 2010s, the date The Pragmatic Engineer and FDE Academy both give. Its intelligence customers could not fully specify requirements, so Palantir embedded engineers in customer environments to learn and build in real time. In 2011 it renamed its solutions and integration engineers forward-deployed engineers and called them Deltas.

Is Palantir’s AI FDE a forward-deployed engineer? No. Palantir’s AI FDE is a software agent inside Foundry that turns natural-language requests into Foundry operations such as writing transforms and running builds. It is a product feature named after the role, not a person embedded in your team.

How does a forward deployed engineering engagement work? Most published sequences run from a short discovery or observation phase to a build on real data and a production or licence decision. Forrester reports AWS’s method as 45 days to a customer engagement; Manhattan Associates offers a 90-day proof of concept that comes with forward-deployed engineers. Our recommendation is narrower: one workflow and one number agreed before the build, then hand over and widen once it runs in production.

Who is on a forward deployed engineering team? A small provider pod, two to six engineers in the examples providers publish, and named people on your side: the process owner, the engineer who will run the system, IT and security, the data owner and any works council. Your side decides whether the system outlives the engagement.

What does the client have to provide in an FDE engagement? A workflow with a measurable target and an owner, access to systems and real data, and time from the people who hold the undocumented logic. Also a processor contract before personal data moves, named oversight people for a high-risk system under the EU AI Act and, in Germany, an informed works council, which is deemed to need an expert when it assesses AI.

How much does forward deployed engineering cost? There is no market price. Palantir runs pilots and bootcamps generally at its own expense, and its UK pricing then lists paid discovery packages and pilot licences before annual licences. AWS talks about business results rather than billable hours, and the AI labs publish no price. The rate cards a buyer can read are on the UK’s G-Cloud 15 framework, Microsoft’s Frontier Co-Engineering service among them.

How do you evaluate a forward deployed engineering partner? Check that a number and a production date are agreed before the build, which platform relationship funds the pod, who directs the engineers day to day, how much of the lead’s time goes to selling rather than to your build, and what stays in your estate when they leave. The label settles none of these.

Do we hire an FDE or engage a partner who provides them? Both models exist. Hiring in-house builds durable capability in a thin market: Lightcast counted more than 7,300 FDE job postings in the first nine months of 2026, up from fewer than 200 in 2023, and a Gartner analyst puts the pool of practising FDEs at a couple of thousand. A partner that deploys FDEs gets engineers into the workflow without the hiring risk. The test for any partner is what stays with you when the engineers leave: the systems, the data, and the ability to run them.



Nucleo is an AI-native enterprise transformation company for European mid-caps. Our engineers work inside your team until a workflow runs in production, model-agnostic and EU-governed, and the intelligence they build stays with you. See how we work or write to us.

Sources: Eurostat, AI in enterprises, Palantir, AI FDE overview, First Round Review, FDE Academy, Andreessen Horowitz, The Pragmatic Engineer, Palantir Form 10-Q, quarter ended 30 June 2026, Amazon, CNBC, Forrester, Accenture, Constellation Research, Deloitte, Deloitte careers, archived June 2026, G-Cloud 15, Microsoft, G-Cloud 15, Elsewhen, A.Team, CIO Dive, SD Times, EU AI Act, Article 26, Works Constitution Act.

The future belongs to those who see it before it is obvious, and build it while everyone else is still out there scouting for the best AI tools.