Plank Research · June 2026

The state of FDE as a service

Forward-deployed engineering went from a Palantir job title to a multi-billion-dollar category in under a year: OpenAI and Anthropic committed $5.5B+ to deployment ventures, FDE job postings grew 729% in the year to April 2026, and the failure rate of unassisted AI pilots remains the highest-stakes number in enterprise software. This report maps where the category stands in 2026, combining third-party market data with first-party analysis of 15 documented production engagements.

committed by OpenAI and Anthropic to forward-deployed engineering ventures in May 2026, $4B+ for OpenAI's Deployment Company and ~$1.5B for Anthropic's services JV
$5.5B+
committed by OpenAI and Anthropic to forward-deployed engineering ventures in May 2026, $4B+ for OpenAI's Deployment Company and ~$1.5B for Anthropic's services JV
startups founded by Palantir alumni, raising about $11.6B between them, the forward-deployed talent that seeded Anduril, Kalshi, and ElevenLabs
111+
startups founded by Palantir alumni, raising about $11.6B between them, the forward-deployed talent that seeded Anduril, Kalshi, and ElevenLabs
of Plank's documented engagements now ship agentic systems, with voice second at 27%
60%
of Plank's documented engagements now ship agentic systems, with voice second at 27%
of Plank's 15 documented engagements share an identical AI stack, which makes vendor neutrality a practical fact rather than a slogan
0
of Plank's 15 documented engagements share an identical AI stack, which makes vendor neutrality a practical fact rather than a slogan

Current status

Why is everyone hiring forward-deployed engineers?

Because the labs proved the economics. The OpenAI Deployment Company launched in May 2026 with a $4B+ raise led by TPG, and Anthropic stood up a roughly $1.5B services joint venture with Blackstone, Hellman & Friedman and Goldman Sachs in the same month. Both pair embedded engineers with ownership of the outcome, modeled on Palantir, whose deployment-led business just grew 85% year over year. Posted forward-deployed salaries now run $200K to $300K+.

Capital committed to deployment services

Disclosed capital behind the two frontier labs' forward-deployed engineering ventures, both announced May 2026.

OpenAI · Anthropic announcements · TechCrunch, May 2026

Why the role compounds

111+startups, ~$11.6B raised

The “Palantir Mafia”: forward-deployed engineers who left Palantir went on to found Anduril, Kalshi, and ElevenLabs, among 111+ companies. Deployment talent does not just ship pilots, it compounds into the next generation of AI companies.

getPIN.xyz Palantir Mafia tracker · Crunchbase, 2026

Who sells forward-deployed engineering in 2026

In twelve months the role spread from one company to the whole stack. Three tiers now sell it, and nearly all of the lab and consultancy capacity targets large enterprises and deploys the seller’s own stack. The startup segment is left to vendor-neutral independents.

Frontier labs

The labs spun up dedicated deployment arms and now embed their own engineers.

  • OpenAI: the OpenAI Deployment Company (DeployCo), $4B+, May 2026
  • Anthropic: a ~$1.5B enterprise AI services JV, plus an in-house Applied AI team
  • Google Cloud: a $750M partner fund and embedded Google FDEs
  • Microsoft: a Forward Deployed Engineering Practice with Accenture

Consultancies & integrators

The incumbents retrained hundreds of thousands of staff and stood up FDE practices.

  • Accenture: $5.9B GenAI bookings; a Microsoft FDE practice and a ServiceNow program
  • Deloitte: Claude across 470,000 people
  • EPAM: 10,000 Claude-certified architects, 250 forward-deployed Black Belts
  • McKinsey, BCG X, Capgemini, Cognizant, TCS, Infosys

Specialist FDE services

Independent, vendor-neutral teams that embed engineers to take AI to production.

  • Distyl AI: ex-Palantir, outcome-based, $1.8B valuation
  • Tribe AI: a 500+ forward-deployed AI engineer network
  • Mechanical Orchard: AI-native legacy modernization (GV-backed)
  • Plank: vendor-neutral FDE, AI engineering, and QA as a service

The consolidation

Who is buying the deployment layer?

Capital is one signal; M&A is the other. In two years the labs, the hyperscalers, the consultancies, and the tooling vendors all went shopping for the same three things: forward-deployed engineering teams, agent and evaluation tooling, and the talent to run them. The pattern says the value is being fought over at the deployment layer, not the model.

The flagship FDE deals

Both labs bought a forward-deployed engineering team to anchor a new services venture, explicitly copying Palantir.

OpenAI's build-out

Beyond the services venture, OpenAI bought the tooling to ship and evaluate agents: experimentation, evals, dev tools, computer-use, and retrieval.

Anthropic's build-out

Anthropic assembled the same kit a different way, from the runtime under Claude Code to SDKs, computer-use, evals, and a life-sciences vertical.

Big Tech's reverse acqui-hires

The hyperscalers used license-and-hire deals to absorb agent and model teams while sidestepping a formal merger.

Integrators rolling up delivery

The consultancies bought delivery capacity, the embedded engineers and operations staff who take AI to production at enterprise scale.

The eval & observability layer consolidates

The eval, observability, and gateway tooling the FDE toolchain runs on folded into data and security platforms.

The exception that frames the map

Palantir, which coined the forward-deployed engineer, made no deployment acquisition in this window. It built the bench in-house, and its alumni went on to found 111+ startups that have raised about $11.6B between them. Everyone else is now buying what Palantir grew.

getPIN.xyz Palantir Mafia tracker, Crunchbase, 2026

The problem

Why do AI pilots die before production?

Four independent research houses, one conclusion: most enterprise AI never ships, and the trend is worsening, not improving. S&P Global's 42% abandonment figure was 17% a year earlier, and Gartner expects more than 40% of agentic projects to be canceled by 2027. The gap is not model capability. It is integration, evaluation, and ownership of the tacit reasoning behind a company's decisions, which, as Foundation Capital argues in its context-graph thesis, was never captured as data in the first place.

The deployment gap, by study

Each bar is a separate study with its own methodology and denominator, so read them as four independent estimates of the same failure mode, not one series.

MIT NANDA 2025 · Accenture 2026 · S&P Global Oct 2025 · Gartner Jun 2025

First-party data

What 15 production engagements show

We tagged every Plank case study published at joinplank.com/showcase by workload, sector, and toolchain. Three patterns stand out: agentic systems are now the default workload, voice has become the second modality, and no two production stacks are identical, which is why vendor-neutral teams embed faster than single-stack ones.

Workload mix

Share of engagements shipping each workload (engagements can ship several, so shares sum past 100%). Agentic systems lead at 60%.

Plank engagement census, n=15 · June 2026

Sector distribution

Sector tags across the same engagements (21 tags over 15 engagements). Production AI demand is broad, not vertical-concentrated.

  • SaaS6
  • Consumer4
  • Hardware/IoT3
  • Developer tools2
  • Healthcare2
  • Fintech2
  • Manufacturing1
  • Retail1
Plank engagement census, n=15 · June 2026

Framework census

Most-used tools across engagements. Orchestration (LangGraph, Agents SDK) and realtime voice dominate; the long tail spans 35+ distinct tools.

Plank engagement census, n=15 · June 2026

The insight

Zero identical stacks in fifteen engagements

Across 15 engagements we counted 35+ distinct tools: OpenAI’s and Anthropic’s APIs, open models, three agent frameworks, two voice stacks, and classic ML alongside them. The median engagement combines four. A lab-employed FDE optimizes for one vendor’s stack; the actual market runs on mixtures. In production, vendor neutrality stops being a stance and becomes a description of the work.

The clearest outside corroboration comes from Andrew Ng: vendor-neutral FDEs are hard to find, he notes, because lab FDEs exist to integrate their employer’s product, and letting them bind your processes “significantly reduces optionality.”

The Batch, deeplearning.ai · May 29, 2026

Where this is headed

Four calls for the next 18 months

The spending backdrop makes the direction hard to miss: AI budgets keep compounding while the services share grows fastest of all.

worldwide AI software spending forecast for 2026, up about 60% year over year and on pace for $638B in 2027
$453B
worldwide AI software spending forecast for 2026, up about 60% year over year and on pace for $638B in 2027
Gartner, May 2026
dollars spent on services for every dollar on software, the labor budget AI services can now address rather than just the software seat
6:1
dollars spent on services for every dollar on software, the labor budget AI services can now address rather than just the software seat
Sequoia Capital, Services: The New Software, 2026
Accenture's GenAI new bookings in FY2025, nearly double the prior year, the cleanest demand proxy for AI implementation services
$5.9B
Accenture's GenAI new bookings in FY2025, nearly double the prior year, the cleanest demand proxy for AI implementation services
Accenture FY2025 results
Palantir's annualized revenue per employee, the operating leverage of an engineering-led model where, its CEO says, only seven of about seventy salespeople really sell
$1.5M
Palantir's annualized revenue per employee, the operating leverage of an engineering-led model where, its CEO says, only seven of about seventy salespeople really sell
Palantir Q1 2026, SEC

01

The category outgrows the labs

In twelve months the FDE title spread from Palantir to OpenAI, Anthropic, Google Cloud, Microsoft, Accenture, Deloitte, and ServiceNow, all now selling or staffing forward-deployed engineering by name. But lab and SI FDEs serve large enterprises and deploy their employer's stack, which leaves the startups and scale-ups they won't staff, the bulk of the market by company count, underserved. If prior platform shifts are any guide, independent specialists grow up around exactly that gap.

02

Agent QA becomes the bottleneck, then the standard

57% of organizations now run agents in production, and quality is the #1 deployment barrier at 32%, ahead of latency and cost (LangChain, n=1,340). As the constraint moves from building to verifying, per-PR testing by humans plus AI becomes table stakes for production AI by 2027.

03

Requirements and workflow design become the edge

Agents, evals, and MCP are becoming table stakes. The durable edge sits upstream, in the work a model cannot do for you: reading what a business actually needs, extracting the tacit rules buried in its operations, and redesigning workflows around AI rather than bolting it on.

04

Time-to-embed becomes the buying criterion

Engineering hires average around seven weeks, and full-cycle AI hiring often runs two to four months. With pilots dying on those clocks, speed of embedding decides vendors more than headcount price. Services that put a productive engineer inside a team within weeks will price at a premium and still win on total cost.

The bear case, and what it actually argues for

The skeptics deserve airtime. A Gartner analyst predicts 70% of enterprises will eventually walk away from agentic AI built through FDE-led engagements, citing vendor cost and missing internal skills. Practitioners note that FDEs are financially incentivized to deepen a vendor's footprint, and that deployment is only about 20% of an AI system's lifetime cost. Read carefully, this is not a case against embedded engineering. It is a case against renting engineers who optimize for their employer's platform and leave nothing behind. The fix is structural: vendor-neutral teams, code in the client's repos, and patterns documented so the in-house team owns the other 80% of the lifecycle.

CIO, May 2026. Computerworld, June 2026.

Methodology & sources

First-party data is derived from the 15 case studies published at joinplank.com/showcase as of June 2026. Each engagement was tagged by workload (agentic, voice, classic ML/CV, IoT/edge), sector, and every tool named in the delivered work. Engagements may carry multiple tags, so shares can sum past 100%. n=15 is a census of published engagements, not a sample of all Plank work.

Third-party data: MIT NANDA, The GenAI Divide (v0.1, 2025) found 95% of GenAI pilots show no measurable P&L return, and externally partnered deployments succeed roughly twice as often as internal builds; IDC with Lenovo (2025) put 88% of AI proofs-of-concept as never reach production; S&P Global Market Intelligence, Voice of the Enterprise(n=1,006, 2025) found 42% of companies abandoned most AI initiatives, up from 17% in 2024; and Gartner (July 2024) put 30% of GenAI projects abandoned after PoC, with 2026 follow-up coverage indicating ~50%. The FDE postings index (643 → 5,330, April 2025 → April 2026, indexed to January 2025 = 100; +729% YoY) and $170K-$200K+ posted salaries are Indeed data as reported by Business Insider (May 2026); the values are indexed, not raw live counts. Bloomberry’s analysis of 1,000+ postings (Revealera data, Nov 2025) separately measured +1,165% YoY and a $173,816 median posted salary. Capital figures are from company announcements and contemporaneous reporting: the OpenAI Deployment Company’s $4B+ TPG-anchored raise (May 2026; valuations of $10-14B circulated in coverage and are excluded as unconfirmed) and Anthropic’s $1.5B services joint venture with Blackstone, Goldman Sachs and Hellman & Friedman (CNBC, Fortune, May 2026). Spending context: Gartner (May 2026) projects $2.59T worldwide AI spending in 2026, up 47%; IDC has GenAI software services growing at ~70% five-year CAGR; and Accenture FY2025 results showed $5.9B in GenAI new bookings. Agent-production and quality-barrier figures are from LangChain’s State of Agent Engineering (n=1,340, fielded Nov-Dec 2025). The vendor-neutrality discussion draws on Andrew Ng, The Batch (May 29, 2026); the bear case on reporting in CIO (May 2026) and Computerworld (June 2026). Hiring-timeline benchmarks: Workable ATS data (~49 days, engineering); vendor-reported full-cycle AI hiring of 90-120 days. Market-structure framing draws on Sequoia Capital, Services: The New Software (March 2026), and Foundation Capital, Context Graphs (June 2026).

Questions, corrections, or data requests: hello@joinplank.com. Reuse permitted with attribution (CC BY 4.0).

Put an embedded AI team on your roadmap

Forward-deployed engineers to deploy, AI-native engineers to build, and on-demand QA pods to validate, embedded with your team, starting the same day.

Book a callhello@joinplank.com