← All reports

Forward Deployed Engineers. Is It Just Consulting?

The companies selling AI to do the work are hiring engineers to make it work, and the consultancies are rushing after them. The role is worth paying for when the AI is still running after the engineer has gone. Without that, it is consulting with a better badge.

The posting

The job is advertised on OpenAI's careers site: "Forward Deployed Engineer (FDE) – SF". Base pay runs from $185,000 to $300,000, plus equity, and "Travel up to 50% is required" ¹. A New York version asks for seven years of full-stack experience and pays up to $310,000 ².

It is a strange thing for an AI company to advertise. The pitch for AI is that it does the work, yet one of the companies building the most capable models is paying senior-engineer salaries for people to sit with customers until the models do something useful. It is far from alone. According to Indeed data reported by the Financial Times, monthly FDE postings rose more than 800% between January and September 2025 ^3,4. Lightcast data, reported by Fortune, puts the median advertised FDE salary above $188,000, against about $145,000 for traditional software engineers, with postings up more than 1,000% year on year in the first eight months of 2026, when all tech postings rose 13% ⁵.

So is this the role that finally gets AI into production, or consulting with a better badge?

Where the job comes from

The role comes from Palantir. Its 2020 stock market filing said its forward deployed engineers "have travelled to bases in Afghanistan and factories in the industrial Midwest to deploy our platforms" ⁶. Inside the company they are called Deltas, and its blog gives the cleanest definition anyone has written: "You can think of a Dev's focus as 'one capability, many customers,' while a Delta's focus is 'one customer, many capabilities'" ⁷.

Nabeel Qureshi, a former Palantir FDE, writes that FDEs "were typically expected to 'go onsite' to the customer's offices and work from there 3-4 days per week" ⁸. His own posting was at Airbus: "I moved out to Toulouse for a year and worked in the factory alongside the manufacturing people four days a week," building what he calls "Asana, but for building planes" ⁹. Product engineers did the other half of the job: "PD engineers then 'productize' what the FDEs build" ⁸.

That two-part design is the heart of it. The engineer in the field solves one customer's problem; the product team "productizes" what the field engineers build.

Not a solution architect, not a consultant

In our view, the difference from a solution architect is not where they sit or what they are called, but what they are accountable for. A solution architect might design the integration. In Palantir's model, a forward deployed engineer helps build it, gets it working in production and feeds what they learn back to the product team ⁸.

Palantir has always been touchy about the comparison with consultants. Its blog insists a Delta "does significantly more engineering and technical work than a consultant." Consultants, one New York Delta wrote there, generally create "a one-time analysis, recommendation, or solution to a specific problem" ⁷.

The job ads bear out the engineering claim. Palantir's New York posting promises "increasing your pain threshold to deliver real value," and asks for a year of post-college experience ¹⁰. Anthropic's London role pays £225,000 to £255,000, expects 25 to 50% travel to customer sites and asks for "Production experience with LLMs including advanced prompt engineering, agent development, evaluation frameworks, and deployment at scale" ¹¹. Salesforce wants engineers who own components "from architecture choice to deployed, validated, and observable in production" ¹². Our read: this is engineering inside the customer's environment, not integration work at arm's length.

The case that it is consulting after all

The title, though, proves little in our view. Bloomberry analysed 1,000 FDE postings. Not one ad mentioned a quota, and 68% required travel, but it classed about three in ten as "Sales Engineer+" roles: "a solutions engineer or sales engineer role that got rebranded as 'FDE'" ¹³.

The tech author Ben Piper calls the role "solutions architecture and implementation consulting, rebranded for a hiring market" ¹⁴. Palantir's own history gives the sceptics ammunition. One former FDE recalls critics calling it "a consulting company masquerading as software" ¹⁵. Its 2020 filing showed a contribution margin of 14% in 2018 and 21% in 2019, before 48% in the first half of 2020, and warned that pilots run "at no or low cost" to customers "may not result in any future revenue" ⁶.

The investors who popularised the role have argued both sides. In June 2025 Andreessen Horowitz's Joe Schmidt called it "the hottest job in startups": "Enterprises buying AI are like your grandma getting an iPhone: they want to use it, but they need you to set it up." "As they race to own the data ingestion layer," he argued, "it's shortsighted to be optimizing for 80% gross margin" ¹⁶. Seven months later, a16z's newsletter carried Marc Andrusko's warning that copycats risk becoming "expensive services businesses with a software valuation multiple and no compounding competitive advantage": "The problem is pretending you're SaaS when you're really services with a platform" ¹⁷.

And one AI company is reported to be betting against the model. Decagon, which sells AI customer service, told the newsletter Newcomer it had crossed $100 million in annualised revenue; Newcomer reports its chief executive, Jesse Zhang, arguing that keeping engineers embedded long term means your software doesn't work ¹⁸.

Defenders and critics share one test, which the venture investor Eze Vidra put best: "If the loop stops at deployment, the company is selling labour" ¹⁹.

The problem the role exists to solve

S&P Global's survey suggests much AI stalls before production. It surveyed more than 1,000 respondents in North America and Europe. Forty-two per cent of companies had abandoned most of their AI initiatives, up from 17% the year before, and the average organisation had scrapped 46% of its proofs of concept before production. The top obstacles cited were cost, data privacy and security risks ²⁰.

OpenAI's head of forward deployed engineering, Colin Jarvis, said in September that about 80% of enterprise AI struggles are deployment problems and about 20% are the model's capability ²¹. Deployment is his business, so he would say that. But if he is right, it explains why the labs keep sending people with their models: the model usually is not the hard part.

The money arrives

In 2026 the role became a business line. On 4 May, Anthropic launched an enterprise AI services company with Blackstone, Hellman & Friedman and Goldman Sachs. Its chief financial officer, Krishna Rao, said: "Enterprise demand for Claude is significantly outpacing any single delivery model" ^22,23. Seven days later OpenAI launched the OpenAI Deployment Company with more than $4 billion of initial investment, and agreed to acquire Tomoro, an applied AI consulting firm, bringing about 150 forward deployed engineers and deployment specialists ²⁴.

AWS reportedly put $1 billion into a forward deployed engineering unit in June ²⁵, and Microsoft reportedly launched a $2.5 billion "Frontier Company" in July ^26,27. In May, Google Cloud's Thomas Kurian said Google was "investing in hiring additional Forward Deployed Engineers" ²⁸; he had already declared in April, "The era of the pilot is over" ²⁹. Salesforce said in March that it had tripled its FDE team in six months ³⁰. On 2 October Anthropic committed $100 million to train 10,000 engineers, nominated by the organisations it works with, as what it calls "Frontier Deployed Engineers" by the end of 2027 ³¹.

The consultancies' hedge

The most telling detail is on OpenAI's announcement. Among the investors in its Deployment Company are Bain & Company, Capgemini and McKinsey ²⁴.

The likeliest reading is a hedge. These firms want a seat in the deployment market whether clients buy through a consultancy or directly from the lab, while helping to finance a business that could compete for the same clients. Money shows what the labs and their backers believe; it is not proof that any of it works.

Meanwhile the consultancies are building FDE teams of their own, many of them around one vendor. Accenture launched an FDE practice with Microsoft in March and a programme with ServiceNow in May ^32,33, and in September said it would build a 1,000-person FDE workforce with Google Cloud ³⁴, about 0.13% of its roughly 799,000 staff by one analysis ³⁵. Deloitte launched Deloitte Forward Deployed Engineering in December 2025 ³⁶, says it was the first global systems integrator to deliver a Salesforce FDE engagement ³⁷, and advertises an "Anthropic Forward Deployed Engineer" role for its government clients ³⁸. EY launched FDE roles in the UK and Ireland in April ³⁹, and IBM has launched "Forward Deployed Units" ⁴⁰.

What it looks like when it works

Colin Jarvis cites Morgan Stanley as one of OpenAI's FDE engagements. By his account, as Business Insider reports it, the technical groundwork took six to eight weeks; persuading the bank's advisers to trust the assistant took about four more months of pilots and feedback ⁴¹.

What filled those months was not more model. According to OpenAI's case study, the bank built "an evaluation (eval) framework to test every AI use case before deployment." Kaitlin Elliott, Morgan Stanley's head of firmwide generative AI solutions, describes the relationship: "Based on all the questions we input and outputs we're getting, we'd sit with OpenAI and say, 'What can we change about our retrieval methods to help the accuracy we need at Morgan Stanley?'" The case study also says "Access to documents has jumped from 20% to 80%" ⁴². By 2024, Morgan Stanley said 98% of its adviser teams had adopted it.

Adoption is what made it worth having, but adoption is usage, not savings. A search for published cases tying an FDE's work to money saved found only vendor-reported results, none checked independently. As far as public sources show, the proof so far is use.

The vendor's side of the loop is better documented. By Jarvis's account, as summarised from an interview, OpenAI's FDE work with Klarna produced a framework called Swarm, which became its Agents SDK ^43,44. That is Vidra's loop working as designed: work done for one customer became product other customers can use. In our view the vendor wins twice. For a customer whose money and processes go into such work, that is worth knowing up front, and pricing.

The other evidence offered for the model is Palantir. The company once mocked as a consultancy had a gross margin of about two thirds in 2019, calculated from its filing ⁴⁵. In the second quarter of 2026 it reported revenue of $1.935 billion, up 93%; its gross margin, calculated from the filing, was about 85% ⁴⁶. Palantir's defenders read that climb as the field-to-product loop paying off; Andrusko warns that Palantir may be a category of one ¹⁷.

The catch for buyers

On 29 September Gartner predicted that 70% of enterprises will abandon agentic AI built by vendors' forward deployed engineers by 2028, and named the problem "FDE washing". "Some charge premium fees without the delivery depth, program management, or change management maturity to justify them," said its analyst, Mukul Saha ⁴⁷. Gartner estimates, CIO Dive reports, that FDE consulting fees can top $200,000 a quarter for a single use case, with demand about four times the supply of engineers ⁴⁸.

It is a forecast, and Gartner is not against the model: "If you truly have the right FDE model, we see it bringing good outcomes," Saha told CIO Dive ⁴⁸. Our read: an agent is not a dashboard you hand over. When the engineer leaves, someone has to own what it gets wrong. That is why Gartner tells buyers to contract for knowledge transfer, IP rights and an exit plan from day one, with an executive sponsor "accountable for business outcomes, not simply the implementation budget" ⁴⁷.

Measuring the engineers is harder than it should be. None of the companies researched publishes how it measures its FDEs. The metrics on offer (time to value, the share of field work reaching the core product, bespoke effort falling, retention) come from investors and advisers ^49,50,51,19.

Then there is lock-in. When the lab that sells the model also sends the engineer, Constellation Research says, "you can expect CxOs to view the effort through the lens of lock-in" ⁵². Activant Capital is sharper: "Be wary of paying platform multiples for what is really a services business riding someone else's inference margin" ⁵³. Everest Group's Peter Bendor-Samuel urges buyers to see FDEs "as a signal, rather than a solution" ⁵⁴.

For consultancies, in our view, the catch is awkward: teams built around one vendor sit uneasily with the independence clients pay them for. A consultancy can rebadge, partner with one lab, or sell clients an exit that actually happens, with the client owning the solution and able to run it. Our read is that the third is the one with a business in it: paid for outcomes, a fixed scope or capability transfer, not for the hours it takes to stay.

How to measure one, and how to hire one

Three numbers would separate the real thing from the rebadge.

• Time to the first use case in real use, by real staff, not merely deployed. /SHIFT's benchmark is 90 days. Morgan Stanley took closer to six months by Jarvis's account ⁴¹, so 90 days is a stretch; a slower vendor should be able to say why.
• The share that is bespoke. Pick one measure (code, engineering hours or cost) and track it. The Next Web reports that Jarvis said in September 2026 that about half of each OpenAI project is now custom work, down from about 90% ²¹. A reasonable target is under half within a year.
• Handover. The customer's own people running it alone for at least a month before the contract ends, against a saving named at the start.

To hire one, keep the full engineering bar: a former Palantir recruiter told First Round Review that FDEs faced the same interview loops as software engineers ⁵⁵. Candidates describe a "decomposition" round in Palantir's process, breaking down an open-ended problem with little or no code ⁵⁶. Shilpa Balaji, a former Palantir FDE and recruiter, adds: "If someone came in overly dogmatic or set in their ways, I found that was actually a red flag for an FDE" ⁵⁵. And look inside first: Salesforce says around 40 to 50% of FDE moves at the company are internal ⁵⁷.

The answer

The forward deployed engineer is the role that gets AI into production when the AI ends in real use, doing a measurable job the customer's own people can run once the engineer has gone. The problem it solves is real: much AI dies between pilot and production (S&P Global found the average company scrapped 46% of its proofs of concept), and OpenAI's head of forward deployed engineering says deployment, not the model, is most of the problem. But as far as public sources show, the proof so far is use, not savings. Whether the work flows back into the vendor's product tells you who is paying for whose R&D, and that should be priced. Without the first test, an FDE is consulting with a better badge, and Gartner already has a name for the expensive version.

The company whose posting opened this story has said as much. "From OpenAI's side," says its head of forward deployed engineering, "we should always be temporary" ²¹.

When a vendor offers forward deployed engineers, you may wish to ask two questions first: will it still be working when they have gone, and who pays for what goes back into the product? A good FDE is measured by how soon they are no longer needed.

Timeline

This article accompanies the /SHIFT episode of the same name. Pay bands are as posted in September and October 2026 and change often; Palantir's gross margins are calculated from its filings; Morgan Stanley's figures come from the bank's own 2024 statement, OpenAI's case study and Business Insider's account of OpenAI's head of forward deployed engineering.