Who Ships the Last Mile? The Forward Deployed Engineer and the Great Role Merge
In 2026 OpenAI, Microsoft and Anthropic each committed money, and in some cases thousands of engineers, to putting their own people inside customers' companies. This article explains the role they are copying, whether it is just consulting with a better title, and what the data says about software roles actually merging.
Why every AI vendor suddenly wants engineers inside your company
The starting point is an uncomfortable finding. MIT's NANDA initiative reported in 2025 that about 95% of the enterprise generative-AI pilots it studied showed no measurable profit-and-loss impact. Its authors pointed at integration and learning gaps, not at the quality of the models. That report was labelled preliminary, and the 95% figure is often quoted more broadly than the study supports, but it captured a mood: the models were good, and deployment was where value leaked.
The vendors' answer was to stop handing customers a product and walking away. Instead they would send engineers to sit inside the customer, write code against the customer's real systems, and stay until something worked. That job has a name, the Forward Deployed Engineer (FDE), and it is older than the AI boom.
What an FDE actually is
An FDE is an engineer employed by a product company who works embedded with a customer, writes production code to solve that customer's real problems, and passes what they learn back to the people building the core product. The role is usually traced to Palantir in the mid-2000s. You will see "2005" repeated in blogs, but no primary source confirms the year, so "mid-2000s" is the safe wording.
Palantir's own engineering blog gives the cleanest description. It splits engineers into Deltas, who work inside a customer's environment, and Devs, who build capabilities meant for many customers. The key idea in most accounts is that the loop closes: customer-specific work is meant to be pulled upstream into the product, so one customer's workaround becomes everyone's feature.
Is it a real FDE function, or consulting with a new title?
Tick every statement that is true of the setup you are looking at. The questions come from practitioner guides, and the verdict is only a rule of thumb.
How the role went from Palantir to the AI labs
Click an event for what happened, how sure we are, and where the information comes from.
Timeline of the forward deployed model
Dates come from the cited sources. Confidence badges say how firmly the date or claim is documented.
How an FDE differs from the people it gets confused with
The dividing line across the vendor and career sources is who writes and owns the production code, and when in the customer lifecycle the person is involved. Titles overlap heavily in practice, so use this as a guide to the typical case, not a rule.
FDE and its neighbours
Pick a role. Time-in-IDE figures come from a single source each and are illustrative.
How an engagement runs
No single authoritative playbook exists, so the loop below is a synthesis of how practitioner guides describe the arc. The important details are the ranges the sources actually give. Engagement length varies from roughly 6 weeks to a year across guides, with 90 to 120 days a commonly recommended embed. Discovery usually takes one to four weeks.
The engagement loop, and the step that feeds the product
Click a step, or watch the marker travel round. The amber step is the one that separates an FDE function from a services shop.
The 2026 scale-up
What changed in 2026 is scale. Three of the biggest AI vendors each announced a programme that treats embedded engineering as a core business, not a side team. They are not the same bet.
Three vendors, three versions of the model
| Vendor | What was announced | Shape | How firmly sourced |
|---|---|---|---|
| OpenAI | "The OpenAI Deployment Company", around 11 May 2026. Reported initial investment above $4 billion and a planned acquisition of the applied-AI firm Tomoro (about 150 engineers). | A separate venture, majority-owned and controlled by OpenAI, led by COO Brad Lightcap. | Announcement on openai.com; funding and acquisition details are from secondary coverage. |
| Microsoft | "Frontier Company", 2 July 2026. $2.5 billion and about 6,000 engineers and industry experts embedded with customers. | A unit Microsoft says "goes beyond" the forward-deployed label, leaning on partners such as Accenture, Capgemini, EY, KPMG and PwC. | Microsoft's own blog post; it has not said whether the money is new or reallocated. |
| Anthropic | "Claude Frontier Academy", 2 October 2026. $100 million to train 10,000 "Frontier Deployed Engineers" by end of 2027. | A credentialing programme: a multi-day in-person phase, then a 12-week residency leading a real project at the engineer's own organisation. First cohorts include Accenture, Bain, Capgemini, Deloitte, McKinsey and others. | Anthropic's announcement; first certified engineers expected in early 2027. |
The economics: margin for a moat, or services in disguise?
The strongest case for the model is Andreessen Horowitz's "Trading Margin for Moat" essay (Joe Schmidt, around June 2025). Its argument is that an AI company should accept lower gross margins early, because the implementation work builds workflows and switching costs that later support pricing power. As precedent it points to cloud-era software: summaries of the essay report ServiceNow's gross margin at 63.2% and Workday's at 54.1% around their IPOs, rising to about 79% and 75% by 2024. We could not open the essay itself, so those figures are as reported by secondary summaries.
The critics make the mirror-image argument. If every customer needs bespoke work, you have a services business with software attached, with the margins and operating burden that implies. One worked example circulating among critics puts an 80% software margin and implementation labour worth 30% of contract value at roughly 50%. Play with that arithmetic yourself.
What implementation labour does to blended gross margin
Illustrative arithmetic only: blended margin โ software margin โ implementation cost as a share of contract value. It is not company data and ignores how costs fall as deployments repeat, which is exactly the moat argument.
The strengths and the costs, side by side
| What it gives you | What it costs |
|---|---|
| Firsthand knowledge of how customers really work | Effort that scales with headcount, not like software |
| Faster integration in messy or locked-down environments | Customers may treat engineers as free services, eroding margin |
| Features validated by real use before core R&D builds them | Customisation sprawl and forked code if work never flows back |
| Trust that a salesperson rarely gets | Distance from the core team, and "client myopia" favouring one account over the market |
| Fast career growth for the engineer | Burnout from travel, firefighting and context switching; heavy travel is typical in postings (Anthropic's cite 25โ50%) |
How fast is demand growing? Depends who is counting
Almost every article about FDEs opens with a growth number, and the numbers disagree. They come from different data sources and different windows, and nearly all come from job boards or vendors with something to sell, so the safe summary is "several hundred to over a thousand percent".
Reported year-over-year growth in FDE postings
Bars are as reported and are not directly comparable. Source type is shown under each label.
The bigger story: what AI is doing to software roles
The FDE sits inside a larger argument, that AI is reshaping who does what in software. Three sources carry most of the weight, and they say different things.
1. Gartner: smaller teams, not fewer engineers
Gartner forecasts that by 2029, 60% of organisations will run smaller software engineering teams at scale, up from 15% in 2026 (reported 7 July 2026 by several outlets; we could not open Gartner's own release, and a Deloitte outlook cites a different Gartner figure, 80% by 2030). Its "tiny team" is usually four or five people, sometimes two or three, typically a product manager, a user or agent-experience designer and at least one AI-native engineer. Gartner frames it as restructuring, not cost-cutting, expects the number of teams to hold or grow, and cautions against cutting junior roles.
2. Indeed: developer postings are rebounding, but only in some layers
US software development postings, Indeed Hiring Lab (July 2026)
Different measures, different baselines. The 71% and 37% shares overlap and do not add up.
3. The entry-level squeeze: one phrase, several different measures
"Entry-level hiring is down 65%" is repeated everywhere. That number comes from SignalFire and measures new-graduate hiring at the largest tech companies against 2019. The same data shows engineering hiring at those companies down only about 11%. Stanford's "Canaries in the Coal Mine" study measures something else again. Put side by side, they are not three versions of one number.
Four "entry-level decline" figures and what each really measures
Are roles merging, or are tasks shifting?
This is where the confident headlines run ahead of the evidence. The claim that product manager, designer and engineer are collapsing into a single "AI builder" is well documented as discussion and as a few company stories, but we found no systematic dataset, such as job-posting counts or organisation charts, showing job titles actually merging. Credible voices argue the opposite: that AI moved tasks between roles while accountability stayed where it was, and that the product manager role persists with its low-leverage tasks shrinking.
How strong is the evidence for each claim?
Higher rungs are stronger evidence. Each claim sits on the highest rung we could find for it.
Where the FDE fits in the merge story
The FDE merge is the one supported by definition: its duties literally overlap neighbouring jobs. The others are discussion.
One more fact complicates the story. GitHub reportedly saw AI agents involved in about one in three pull requests, up from under one in ten a year earlier (via CNBC's coverage; we did not check it at source). More automated code does not obviously reduce demand for the person who owns the outcome, which is why the roles that grow in the data are senior engineers and anyone close to a customer's real problem.
What to do with this
If you are, or want to become, an FDE
- Agree handoff and success criteria before you start, so the engagement can end with the customer owning the result.
- Treat every workaround as a candidate product feature, and write down what you learn in a form the core team can use.
- Stay close to the core engineering team through design reviews and contributions, or the role isolates you.
- Prepare offline and restricted-environment tooling before you arrive on site.
- Quantify business impact for the customer and for your own leadership. Guard against burnout from travel and firefighting.
If you are building or buying an FDE function
- Ask the roadmap question first: what fraction of field learning shows up in the product, and how fast?
- Make engagements finite, and expect the customer to ask for more free work. Decide the policy in advance.
- Report engineers into product and engineering, not into a services profit-and-loss account.
- Review field patterns on a regular cadence and decide deliberately what generalises. Some guides suggest shipping at least one feature back to the core product per engagement as a leading indicator.
If you are an engineer choosing a path
- The data points to judgment, domain knowledge and owning outcomes gaining value, with entry-level routes narrowing. That favours learning to work close to real problems, which is also what an FDE does.
- It does not show that every role is merging into one. Build breadth, but do not assume a single "AI builder" title is coming.
Reading the numbers: what you may have seen versus what the sources say
| Claim in circulation | What the sources support |
|---|---|
| FDE postings are up more than 1,000% | About 729% on Indeed, up to 1,165% elsewhere. All job-board or vendor data; "roughly 7x to 12x" is safer. |
| Engagements last 30 to 90 days | Guides range from about 6 weeks to a year; 90 to 120 days is a common recommendation. |
| FDEs spend 30โ40% of time on user research, 5โ15 interviews a week | No source found. |
| There is a "three-client rule" for productising | Only a heuristic in some guides, used in different ways. Not a standard. |
| Software job listings are up 30% in 2026 | Indeed Hiring Lab reports about +15% since Feb 2025, and still 27.5% below Feb 2020. |
| Entry-level developer hiring is down 65% | SignalFire: new-grad hiring at the largest tech firms vs 2019. Other measures are smaller; see the chart above. |
| Gartner: 60% of organisations run smaller teams by 2029, up from 15% in 2026 | Reported consistently by news outlets (7 July 2026). We did not open Gartner's release; a Deloitte outlook cites 80% by 2030. |
| Palantir created the role in 2005 | Palantir as the origin is well supported; the exact year is not confirmed by a primary source. |
| Roles are merging into a single "AI builder" | Well documented as discussion and anecdote; not shown in labour data. Gartner's model keeps separate roles. |
๐ Reference links
- Primary or near-primary
- OpenAI โ launch of the OpenAI Deployment Company (not opened; details from secondary coverage)
- Microsoft โ Frontier Company announcement (2 Jul 2026)
- Anthropic โ Claude Frontier Academy (2 Oct 2026) and CNBC coverage
- Palantir โ A Day in the Life of a Palantir Forward Deployed Software Engineer
- Indeed Hiring Lab โ AI and Job Postings: From Destruction to Creation? (8 Jul 2026)
- Stanford SIEPR โ Canaries in the Coal Mine? and Canaries dashboard
- a16z โ Trading Margin for Moat (not opened; figures from summaries)
- Secondary coverage
- Channel Dive โ Gartner: AI won't cut jobs, it'll multiply them; CIO Dive; ITPro
- Deloitte โ software industry outlook (the 80% by 2030 figure)
- CNBC โ AI is automating coding. So why are software developer job postings on the rise?
- Startup Fortune โ SignalFire entry-level data; HatchWorks
- The Decoder โ Microsoft Frontier Company
- Healthcare IT News โ MIT: 95% of enterprise AI pilots fail to deliver measurable ROI (MIT's report itself not opened)
- CIO.com โ Stop sending me engineers
- Practitioner and vendor blogs (opinion; read with care)
- Paraform; Perspective AI โ hiring trends; Perspective AI โ playbook
- KDnuggets; Vibe Engines; PostHog; The Forward-deployed Business Model
- The Forward Deployed; InterviewBit; Netguru
- Kanerika; FullStack; HashLogics
- Product Trio โ tasks move between roles; Product Management Society; Amazing CTO
- Related on this site: AI Agents & the Enterprise Agent Economy ยท Building an AGI Startup: The Applied & Agentic Layer