Forward Deployed Engineers: AI’s hired guns

By Iain, 22 Jul 2026: AI, Strategy, Hot takes

The last mile of AI deployment is a real person, sitting in your office, learning your business, writing your plumbing, and making things work reliably. That will remain human for the foreseeable future.

Google Trends shows a 5,000% increase in searches for “AI forward-deployed engineer” over the past five years. The curve stays flat from 2021 through mid-2024, then turns vertical. Indeed data reported by Business Insider shows postings for the role rising from 643 in April 2025 to 5,330 a year later, a 729% increase. LinkedIn’s own study found 42-fold growth between 2023 and 2025, making forward-deployed engineering the fastest-growing AI job category by a wide margin. Andreessen Horowitz called it “the hottest job in startups” in a June 2025 essay by Joe Schmidt. Median total compensation at a frontier lab is $485,000, according to a 2026 survey of 1,500 FDEs. Senior-level FDEs at those same labs can earn salaries as hog-whimperingly large as $725,000.

What an FDE does

The title comes from the military, where a forward-deployed unit is stationed close to the operational theatre rather than back at base. Palantir adopted the language in the early 2010s because the company was selling intelligence software to agencies that could not openly share their requirements.

Traditional enterprise delivery, where you gather requirements, design a solution, and ship it, was structurally incompatible with the work. Palantir’s answer was to embed engineers directly inside customer environments, building software under the same operational constraints the customer faced. The company initially called these engineers “Deltas.” By 2016, Palantir called them forward-deployed engineers and had more of them than traditional software engineers.

An FDE is a software engineer who works inside your organisation, learns your workflows, connects your systems, and builds the production plumbing needed to make a vendor’s AI product do something useful with your data. They write production code, and they stay accountable for whether the system still works after they leave. If that sounds expensive, it is. And the fact that every major AI lab has decided it needs an army of them tells you something.

Here come the consultants

In May 2026, OpenAI and Anthropic each clarified their positions within days of one another. OpenAI launched the Deployment Company on May 11, a joint venture majority-owned by OpenAI and backed by more than $4 billion from 19 investors, including TPG, Advent International, Bain Capital, and Brookfield. To staff it quickly, OpenAI acquired Tomoro, an applied AI consulting firm, adding roughly 150 forward-deployed engineers. A week earlier, Anthropic had announced its own $1.5 billion joint venture with Blackstone, Goldman Sachs, and Hellman & Friedman, now formalized as Ode with Anthropic, built on the foundation of Fractional AI and designed to embed Claude-first engineers inside mid-market companies.

The language is about “accelerating adoption” and “redesigning workflows.” Read between them, and the message is starker. The two companies building the most advanced language models on the planet have concluded that selling access to those models is not enough. Someone has to fly in, learn the client’s data, map the client’s processes, wire the client’s legacy systems, write the client’s guardrails, and babysit the result until it stops hallucinating in production. That someone costs half a million dollars a year, and the AI labs have decided the cost is worth absorbing because the alternative is that nothing ships.

Anthropic’s Garvan Doyle, Head of Forward Deployed Engineering for the Americas, put it in terms that would be refreshing if they were not also alarming. “As mid-size companies move from experimenting with AI to building it into their operations, they need partners with real implementation depth and a clear knowledge of how their businesses actually work.” Translated from corporate into English, this means your average company cannot do this on its own. The talent does not exist internally. The tooling is not mature enough to paper over the gaps. And the gap between a working demo and a working deployment remains wide enough to park $5.5 billion of investor capital inside it.

About that awkward 95% failure rate

MIT’s Project NANDA published The GenAI Divide: State of AI in Business 2025 in July that year, and the headline figure shook markets. Across more than 300 enterprise generative AI initiatives, 150 leadership interviews, and 300 public deployments, 95% showed no measurable impact on profit and loss. The remaining 5% were pulling in millions. The researchers were clear about where the problem lay. “This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach.” Buying AI from specialist vendors and building partnerships succeeded about 67% of the time, while internal builds succeeded roughly 33% of the time. The models worked fine. The deployments did not.

The ONS published fresh UK data this week that tells the same story from the other direction. AI adoption among UK businesses with 10 or more employees has tripled since late 2023, from around 12% to around 35%. That sounds like progress until you look at the depth. The average number of AI technologies per adopting business has barely moved, from 1.4 to 1.6. Only 10% of adopting businesses describe their use as extensive. Only 15% say that more than half of their employees use AI in their daily work. Only 11% report that more than half their workforce has received any AI training at all. The ONS summary is blunt: adoption is a mile wide and an inch deep. Businesses are buying the tools, but they are not deploying them in ways that change how work gets done. The gap between “we use AI” and “AI is wired into our operations” is exactly the gap the FDE exists to bridge, and the ONS numbers suggest that gap is vast.

I have been writing about this divide from different angles since BTG launched, and each angle leads to the same place. The production agent stack mapped 11 engineering concerns that a production deployment must cover and concluded that Gartner projects 40% of agent initiatives will be cancelled by 2027 because organisations underestimated the discipline required to run non-deterministic systems. Prototype purgatory examined why AI projects stall between demo and production in a loop that never ends once you swap controlled conditions for the chaos of a live business.

It’s worth understanding just how painful it is to ship production agents today, for those fortunate enough not to have had to do so. To be clear, we’re not talking about demo agents or “look what I built this weekend” prototypes. These are agents that handle sensitive data, integrate with business-critical systems, and need to satisfy compliance teams. The ones that, if you’re not losing sleep over, you’re not doing it right… consulting and integration services will capture significant revenue, helping enterprises navigate the transition. The technology is complex enough that most companies will want guidance.

What the FDE boom tells us

Strip away the career-advice framing and the salary clickbait and the FDE phenomenon reveals four things:

The tooling is not ready for civilians. If deploying an AI agent into a production workflow were only a matter of configuration, API keys, and careful prompt engineering, you would not need a $485,000-a-year engineer sitting in your office for six months to make it work. The fact that you do need that engineer, and that the vendor is the one supplying them, means the product is not yet a product in the way normal software is. It is still an ability that requires a translator between technology and business, and that translator must be fluent in both languages. SaaS, at its best, eliminated the need for that translator. AI, at its current maturity, has brought them back.

The integration problem dwarfs the model problem. Model quality has improved faster than anyone predicted, and inference costs have dropped by orders of magnitude. The bottleneck has shifted downstream to the messy, bespoke, organisation-specific work of connecting a model to live data, legacy systems, compliance requirements, and human workflows. This is plumbing, not science, and it resists automation precisely because every customer’s plumbing is different. An FDE’s week, according to the Perspective AI survey of 1,500 FDEs, is split roughly 47% customer-facing, 31% shipping code, and 22% on internal coordination. Nearly half the job is figuring out what the customer actually needs, which is another way of saying that half the work cannot be done remotely, at scale, or in advance.

The vendor is subsidising your implementation. When OpenAI puts $500 million of its own money into the Deployment Company, and Anthropic commits $300 million to Ode, they are not doing it out of generosity. They are doing it because the revenue from API access alone does not build the kind of lock-in that justifies a $300 billion valuation. Menlo Ventures’ 2025 mid-year LLM market update found that Anthropic held roughly 32% of enterprise LLM market share, OpenAI roughly 25% (down from 50% in 2023), and Google roughly 20%. In a commoditising model market, the vendor that owns the deployment relationship owns the customer. The FDE is the mechanism by which model providers are buying distribution, or, if you prefer a16z’s framing, trading margin for moat.

The hardest problems aren’t technical. MIT’s NANDA report found that poor vision, a lack of talent, limited budgets, and underestimating the challenge were the main reasons for failure. Computerworld’s coverage quoted Jack Gold, principal analyst at J.Gold Associates, who described FDEs as “essentially hired guns for AI deployments,” brought in because internal teams could not produce results or fast ROI. The FDE does not arrive with a better model. They arrive with the ability to listen to a hospital administrator describe a broken intake process, connect that description to what a specific model can do, write the code that links the two, and stay long enough to confirm that the intake process actually improved. That mix of domain translation, technical execution, and sustained accountability is rare in any field, which is exactly why it costs what it costs.

The mainframe moat, rebuilt in months

Three days ago, IBM lost a quarter of its market value in a single session. The stock fell 25%, its worst day since Black Monday in 1987, erasing roughly $68 billion. Infrastructure revenue fell 7%, reversing a quarter in which that same division had jumped 15% on the strength of a record z17 mainframe launch. CEO Arvind Krishna told investors that customers had shifted capital spending toward AI-related hardware and away from IBM’s software and infrastructure stack.

The coverage treated this as an earnings miss. It was also a look at the oldest lock-in pattern in enterprise technology. IBM’s mainframe business has always worked the same way. You pay capex for the hardware and opex for the software that runs on it. You get the worst of both worlds. But you stay, because 40 years of COBOL, batch schedules, and accumulated transaction logic have made leaving more expensive than staying. I know this from direct experience in insurance, where we once had to buy parts from eBay for one machine because the product line it supported made good money, but nowhere near enough to justify new hardware or a migration.

Gartner published an analysis in April concluding that customers are “increasingly recognising the near-impossibility of a mainframe exit at an acceptable cost and risk.” The banks, the airlines, and the insurers that drove the first wave of IT in the 1960s are still there, still running the same systems, still paying IBM their pound of flesh. They gained access to technology before any other businesses. The price they pay is that their operations are so numerous and complex that they cannot afford to move them. Nobody opts into this arrangement. Nobody has for years. Any new company builds in the cloud, or buys commodity hardware and runs Linux. The IBM mainframe customer base is a captive population, not converts.

Now look at what the AI labs are building with their FDE programmes and ask whether the mechanism is really that different. OpenAI’s Deployment Company sends engineers to wire your systems to GPT-4o’s function-calling conventions, its tool-use patterns, its context-window behaviour. Anthropic’s Ode embeds Claude-first engineers who build guardrails tuned to Claude’s hallucination modes, eval suites calibrated to Claude’s failure patterns, orchestration logic fitted to Claude’s API surface. Every hour of FDE work is an hour of switching cost being manufactured. The plumbing they install is fitted to pipes that only one supplier makes, and the pipes are not easily interchangeable.

One difference is speed. IBM’s lock-in built up over decades of transaction volume. An FDE engagement creates meaningful switching costs in months. And unlike the mainframe, where you at least owned the hardware, the AI version is opex on opex. You pay for API access to a model you do not own, and you pay for the integration that binds you to it. Walk away, and the model-tuned plumbing you’re left with may be of very limited value for what you build next.

The labs know this. It is the entire commercial logic behind a16z’s “trading margin for moat” framing. Lose money on the deployment to own the relationship. The FDE is the delivery mechanism for lock-in dressed as a service. I wrote earlier this year about AI collapsing the switching costs that have protected incumbent SaaS vendors for decades:

Migration cost was always the real moat. You stayed with Salesforce, Oracle, or SAP because leaving would be worse than staying, even if staying was painful. Palantir’s CTO claimed on their Q4 2025 earnings call that their platform can now complete complex SAP ERP migrations from ECC to S4 in as little as two weeks, where that same work previously took years. Amazon used AI agent coordination to modernise thousands of legacy Java applications in a fraction of the expected time. If those claims are even directionally correct, the lock-in argument for legacy monoliths weakens every time migration gets cheaper and faster.

And this is where IBM’s rough week and the FDE boom meet. IBM’s hardware revenue fell because customers are starting to think AI will eventually free them. They are delaying new mainframe purchases, shifting capital to AI infrastructure, and waiting for migration tools to mature enough to make the move. That belief may be premature. Gartner thinks so. But the belief itself is already shaping buying decisions, which is why IBM just had its worst day in 38 years.

The paradox is that the same AI industry promising to free enterprises from legacy lock-in is also building a new version of the same trap. The FDE does not arrive to set you free. They arrive to tie you to a specific vendor’s model, tooling, and stack. The switching costs are different from IBM’s. They are not COBOL and batch schedules. They are prompt libraries, eval suites, guardrail configurations, and orchestration logic. But they are switching costs all the same, and they pile up faster than the mainframe version ever did because technology advances faster and integration goes deeper.

The question for any business watching this unfold is not whether to use AI. It is whether you are building on ground you control. Mainframe customers in the 1960s did not know they were being locked in. They were solving real problems with the best available technology, and lock-in was a side effect that took decades to become apparent. The AI equivalent is happening now, at a pace that compresses decades into quarters.

Get your tanks off my lawn

The AI labs’ direct FDE efforts also collide head-on with recently announced partnerships. For example, in February 2026, OpenAI launched “Frontier Alliances”, a multi-year partnership with BCG, McKinsey, Accenture, and Capgemini to sell its enterprise products. Three months later, it launched the Deployment Company, a $4 billion vehicle that competes directly with those same partners. Anthropic followed the same path, signing deals with Deloitte and Accenture before creating Ode as a parallel channel that can bypass them entirely.

The Telegraph reported that the labs are now targeting the $1 trillion global management consulting market head-on. Lian Jye Su, an analyst at Omdia, described the pattern, “we are looking at AI companies now…deciding that they just want to be Palantir.”

The consultancies are unsurprisingly not standing still. Accenture spent $1 billion acquiring Faculty, and all five major firms are hiring FDEs of their own. But the structural disadvantage is clear. A Deloitte partner selling Claude implementations is a reseller. An Ode engineer who sells Claude implementations has unique insider knowledge of the model. The labs can offer what the consultancies cannot: direct access to the people who understand the model’s quirks because they built them. That advantage only grows as models change faster, because every update widens the gap between the vendor’s engineers and everyone else’s.

What this means for the rest of us

Here is the challenge for a small or medium business watching this unfold. You are not Blackstone. You do not have $300 million to commit to a joint venture. You are unlikely to be the kind of company that OpenAI’s Deployment Company or Ode with Anthropic will prioritise, because those ventures are targeting firms at BBVA and John Deere scale, enterprises with thousands of employees and the revenue to justify a six-figure FDE engagement.

You are also, however, a company being told by every vendor, conference speaker, and LinkedIn thought leader that you need to adopt AI or be left behind. AI agents will transform your operations. AI will make you competitive. AI will give you the productivity of a company ten times your size. But making any of this work in production, with your data, your systems, and your unique constraints, remains genuinely difficult. So what do you do?

A close friend who runs a 40-person services company recently asked me whether he needed an FDE. He had seen the salary figures and the Google Trends chart that goes vertical. My answer was no, but not unequivocally so. Small- to medium-sized enterprises may not have such large or sensitive use cases. Still, the very existence of this role, its explosive growth, the half-million-dollar compensation, and the billions being committed to building entire companies around it represent an industry-wide admission that the technology is not yet self-serve.

This is where we are finding an increasing need for what we do at Better Than Good. We think of ourselves as Mess Assessors rather than the rather grandiose Forward Deployed Engineers, but we do the same basic work, only helping smaller companies without huge consultancy budgets. We still bridge the operational gap between theoretical capabilities and a working system. If that sounds like something you need, we should chat.

Palantir recognised the underlying need 15 years ago, when it sent engineers to sit inside intelligence agencies because the software could not be deployed any other way. The rest of the industry is now reaching the same conclusion. The vocabulary has changed, from “Delta” to “FDE” to “Deployment Company.” The underlying challenge has not. The last mile of AI deployment is a real person, sitting in your office, learning your business, writing your plumbing, and making things work reliably. That will remain human for the foreseeable future.