On May 11, 2026, OpenAI announced the launch of the OpenAI Deployment Company — a standalone business unit backed by more than $4 billion from 19 investors including Goldman Sachs, TPG, Bain Capital, and McKinsey. Its stated mandate: embed specialized engineers directly inside enterprises to help them actually use AI. To seed the effort, OpenAI acquired Tomoro, an applied AI consulting firm, to bring 150 Forward Deployed Engineers on board from day one.
Think about that for a moment. The world's most valuable artificial intelligence company whose technology is reshaping every industry on earth, raised four billion dollars to help people use its products. Not to build the next model, but to show up at your office and make sure the software works inside your building.
What OpenAI is doing makes strategic sense and will likely be profitable on its own. I say it because it is, quietly, the most expensive admission in the history of enterprise software that there is a profound gap between what AI promises and what enterprises can actually do with it — and that the biggest AI labs have decided the cheapest fix is to send human beings to close it.
The industry has celebrated this decision. I think we should examine it more carefully.
How We Got Here
Palantir invented the Forward Deployed Engineer around 2009, not by design philosophy but by necessity. Their intelligence agency clients could not openly state what they needed. Their platforms were too complex to hand over with documentation alone. So Palantir sent engineers — Shyam Sankar, now CTO, famously described them as people who "absorb pain and excrete product" — to live inside customer environments, map fragmented data into usable architectures, and build software that kept running after the engineers went home. The model drove what investors have calculated as 640% returns and established Palantir as a category unto itself.
By 2025, every major AI lab had imported the model. OpenAI, Anthropic, Google DeepMind, Databricks, Scale AI — all hiring FDEs at extraordinary speed. Job postings for the role surged between 800% and 1,000% in the span of a year. Total compensation of an FDE at OpenAI and Anthropic reached $350,000 to $550,000. In April 2026, EY launched FDE roles in the UK and Ireland explicitly to "scale AI into production" for enterprise clients. Andreessen Horowitz published a widely-read piece calling the FDE "the hottest job in startups" and argued that services-led growth — trading short-term margin for deep customer integration — was the new competitive moat.
The industry largely agreed. I am skeptical.
Two Kinds of FDEs
Not all FDEs are the same, and conflating them is costing enterprise buyers both money and time.
The first kind is, in plain terms, a confession. It exists because the product cannot deploy cleanly in real enterprise environments — because it cannot handle legacy authentication, because the data model was never production-ready, because the API surface was designed for demos rather than operational integration. McKinsey's 2026 Global AI Survey found that 73% of enterprise AI deployments fail to achieve their projected ROI. S&P found that 42% of companies abandoned at least one AI initiative in 2025, with an average sunk cost per abandoned project of $4.2 million. Across almost every study, the dominant root cause isn't the model — it's the integration last mile. The pilot impresses the sponsor. Then the team discovers the actual data is messier, the Active Directory is twelve years older than the vendor assumed, and the change-advisory board wants three months of paperwork before anything moves to production. When you send an FDE to navigate that gap, it’s really an expensive apology masked as a forward deployed engineer. When a16z calls this "trading margin for moat," that is a VC-flattering way of describing what most enterprise CFOs would call technical debt — socialized onto the customer.
The second kind of FDE is something genuinely different, and it is worth defending. This one does not exist because the product failed. It exists because enterprise transformation is inherently complex in ways that no product alone can fully anticipate from the outside. In regulated industries — financial services, healthcare, telecom, government — the gap between what AI can do in theory and what an organization has the data, governance, and operational maturity to use in practice is wide. Bridging that gap requires someone who understands both the platform and the customer's world: their compliance posture, their content architecture, their data infrastructure, their internal decision-making patterns. This FDE is not apologizing for the software. They are doing something the software was never designed to do alone.
Palantir understood this distinction from the beginning. When their engineers embedded with the NHS in England to build the Federated Data Platform, they were not patching a broken product. They were navigating genuinely complex data environments where the stakes were clinical, the systems were fragmented across decades of legacy architecture, and the outputs had to be trusted by clinicians who had never worked with AI before. That kind of work earns its price tag.
BBVA's partnership with OpenAI offers a more recent illustration of what the second kind of FDE looks like at scale. What began as a ChatGPT Enterprise deployment has expanded into a system serving 120,000 employees across 25 countries, with more than 4,000 custom AI tools embedded in daily operations and individual workflows showing up to 80% efficiency gains. The FDE work there was not compensating for product gaps. It was driving transformation outcomes that required deep expertise in both the platform and the bank's operational environment — regulatory requirements, workflow constraints, cultural readiness. That is a different engagement entirely.
OpenAI clearly knows the difference. Their $4 billion Deployment Company is a bet that demand for both kinds will be enormous for years to come. They are probably right about the demand.
The question for enterprise buyers is which kind they are actually buying.
Where Real Complexity Lives
For a CIO or CMO navigating this landscape, the more useful question is not whether your vendor has FDEs. It is what problem those FDEs are actually solving. Here is the terrain where the second kind earns their keep — and where the right platform and the right expertise, working together, can deliver outcomes that most organizations can only dream about today.
Data, personalization, and the activation gap. The most persistent frustration I hear from enterprise marketing and technology leaders is some version of this: we know who our customers are, we have years of behavioral and transactional data, and we still cannot serve them a relevant experience at the moment of interaction. The problem is not ambition or even technology. It is that the data is scattered — CRM here, analytics there, consent signals somewhere else, behavioral data not connected to anything actionable. Getting those signals unified into a coherent customer profile and activating it in real time, at the moment of interaction, requires work that sits at the intersection of data engineering, content infrastructure, and real-time decisioning. No product ships with that solved. The right FDE makes it possible. The wrong vendor hires an FDE to pretend their product handles it.
Multi-site orchestration and governance overhead. Consider an enterprise running 80 brand sites across 12 markets, each with different regulatory requirements, different editorial workflows, and different audience segments. The operational complexity alone is significant. Adding AI-driven content operations to that environment — autonomous drafting, workflow routing, validation, approval — requires someone who can map governance requirements to system configuration with real precision. A single misaligned configuration in a regulated industry is not a UX problem; it is a compliance incident. The FDE who understands both the platform and the regulatory context is one of the most valuable people in that organization. The FDE who is there because the platform has no governance layer is just expensive cover.
Agentic workflows and the trust gap. This is early, but it is coming faster than most organizations anticipate. The shift from AI-assisted content workflows to autonomous agentic operations — where AI agents draft, route, validate, and publish with human oversight built in — requires someone to configure the guardrails, define what good output looks like in a specific regulatory and brand context, and establish the approval chains that let an organization actually trust what the agent is doing. This is not something any platform delivers out of the box. It requires judgment that sits between platform capability and organizational context. For enterprises in regulated industries, the person who can build that trust infrastructure will be among the most consequential contributors to AI adoption over the next three years.
The Prerequisite Nobody Is Discussing
Almost every FDE conversation focuses on infrastructure, models, and data pipelines. What is conspicuously missing is the content layer.
For any of the above to work — personalization at scale, agentic content operations, real-time decisioning — the content itself has to be structured, governed, and architecturally ready for multi-channel delivery. You cannot personalize content that is not tagged for personalization. You cannot run agentic workflows over content that is not versioned and permission-controlled. You cannot serve real-time experiences from a content system that requires a developer to change a headline.
The integration wall that kills 70% of enterprise AI projects is not only a data problem or an authentication problem. A significant and underappreciated portion of it is a content architecture problem. Enterprises that adopted a pure headless CMS during the last wave of architectural enthusiasm now face marketing teams that cannot make changes without a developer queue, content that is not structured for AI consumption, and no clear path from their content system to their data layer. The FDE walking into that environment is not accelerating transformation. They are doing archaeology.
A Framework for Buyers
So what should a CIO or CMO actually do with the FDE boom?
When a vendor tells you they have Forward Deployed Engineers who will ensure your success, ask two questions.
First: what has their FDE team learned in the field, and how has it changed the product? If there is no answer — if the FDE work is siloed from the roadmap, if pain does not become capability — you do not have a forward deployed engineer. You have an expensive customer success manager with a more impressive title. The defining characteristic of a legitimate FDE model is that customer friction drives product evolution. The feedback loop is the point. Without it, you are funding a workaround, not building toward a solution.
Second: what problem would the FDE be solving for you specifically? If the answer is "helping you get the product deployed and working," probe harder. A platform built for enterprise environments should deploy in enterprise environments. If the FDE's job is to bridge the distance between your first-party data, your content infrastructure, and the AI outcomes you are trying to achieve — if they are navigating genuine transformation complexity, not product gaps — that is a different and potentially very valuable engagement.
The companies worth betting on are the ones where the FDE is at the frontier of what is possible, not papering over the gaps in what exists.
Where This Is Going
The FDE will not disappear. But the market will eventually separate the two kinds, and the economics will follow.
AI labs and early-stage companies deploying genuinely novel technology will need FDEs for the foreseeable future. The gap between model capability and production deployment is real and significant. That is legitimate. OpenAI's $4 billion bet is probably right about the near term.
But for the category of software that was already supposed to be production-ready — content platforms, experience orchestration systems, enterprise workflow tools — the companies hiring FDE armies to compensate for product gaps will eventually face a reckoning. Enterprise buyers will recognize the pattern: they are paying services prices for software returns, and the math will stop working.
The more interesting frontier — and the one that will define which platforms matter in five years — is the synthesis of structured content, unified first-party data, and AI decisioning into a closed-loop experience engine. This is the terrain where real personalization happens: not the rule-based segmentation that passes for personalization in most enterprises today, but genuinely contextual, adaptive experience delivery that responds to who a customer is, what they need, and where they are in their relationship with a brand, in real time and at scale. Getting there requires content infrastructure, data infrastructure, governance infrastructure, and the human expertise to configure all three. It is where the second kind of FDE — the transformation architect — will do their most consequential work yet.
Most enterprises can only dream about that outcome today. The ones that get there will not be the ones who hired the most FDEs. They will be the ones who asked the right questions at the right time, built on the right foundation, and put genuine expertise to work on the right problems.
The FDE boom is telling us something. The question is whether we are listening to the right part of the signal.