The Top of the Iceberg

You've probably seen the headlines by now. "95% of AI pilots fail." "Companies wasted billions on AI that never shipped." "The AI bubble is real." It's become a comfortable narrative for anyone who wants a reason not to move.

Here's the thing about that stat: it's true, and it's also only the top of the iceberg. What almost nobody is writing about is the other 5%.

The other 5% aren't posting case studies. They're not doing press tours. They're not slowing down long enough to explain how they did it, because explaining it means giving away the head start, and the businesses that figured this out first are treating it exactly like what it is: a temporary, compounding advantage in a race that has actually gone zero-sum for the first time in modern business history.

That's the part of this conversation that's missing. It was never really zero-sum before. A well-run regional business could coexist with a well-run national one for decades. Distribution had physical limits. Systems took years to build and were expensive enough that most companies simply didn't bother. Mediocrity was survivable because everyone around you was roughly as slow as you were.

That era is over. The company with better systems and better distribution doesn't just outperform the legacy company without them anymore, it eats its market share entirely, because the tools to build those systems are now cheap and instant instead of expensive and slow. Every large, dominant company in the last fifty years got that way by building superior internal systems before anyone else had the tooling to compete. The difference now is that the tooling is available to everyone, and companies are still failing anyway, because having access to a tool and knowing how to structure a business around it are two completely different problems. Most companies are failing at the second one, not the first.

The AI isn't the bottleneck. That needs to be said plainly, because the "AI doesn't work" narrative is doing a lot of load-bearing work for people who'd rather not confront the real issue. Vas from Varick Agents, an ex-Meta engineer who now embeds forward deployed engineers inside enterprise clients, put it about as clearly as it can be put: he watched one C-suite executive blow through an entire $10 million AI budget in three months that was supposed to last a year, simply by handing AI to everybody and slapping it on everything, and getting token-maxing and hallucinations instead of results. The model wasn't the problem. There was no system underneath it. There was no judgment about where AI actually belonged in the workflow. It was a Ferrari with no driver, and it embarrassed everyone standing around it.

The businesses seeing exponential gains right now aren't smarter than everyone else. They just stopped running post-COVID-era processes on top of post-AI-era tools, rebuilt the underlying structure first, and then layered intelligence on top of a foundation that could actually carry it. That's the entire difference between the 95% and the 5%, and it's exactly the gap one very specific, very new job title exists to close.

Where the Term Actually Comes From

The role isn't new. The Forward Deployed Software Engineer, or FDSE, originated at Palantir, which built an entire multi-billion dollar company on the model of embedding elite engineers with massive government and commercial clients to solve huge, complex data problems, engineers legendary for their autonomy, high stakes, and total ownership of the problem.

What changed in the last eighteen months is that everyone else finally admitted Palantir had the model right the whole time.

In May 2026, OpenAI launched The Deployment Company, a $10 billion venture backed by TPG, Goldman Sachs, SoftBank, and BBVA, structured entirely around embedding engineers inside high-stakes enterprise deployments. Anthropic followed with a $1.5 billion joint enterprise services venture alongside Blackstone and Hellman & Friedman. These aren't consulting expansions. Two of the most valuable AI labs on the planet just bet billions of dollars that the model isn't the product anymore. The deployment is.

Amazon reached the same conclusion at a different scale, announcing a $1 billion investment in June 2026 into a dedicated Forward Deployed Engineering organization intended to embed thousands of engineers with customers, with engagements reportedly running approximately 45 days per client, writing production code and navigating both technical and organizational barriers. Anthropic is running a parallel track under the name Applied AI, with 20 openings listed in early August 2026 alone and a $100 million commitment to its Claude Partner Network. OpenAI has turned this into a distinct organizational function entirely, with more than 40 open roles across the U.S., Europe, Asia, Australia, and the Middle East as of August 4, 2026, describing its FDEs as owners of the complete deployment lifecycle: discovery, technical scoping, system design, development, and production rollout.

This is what it looks like when an entire industry realizes, more or less simultaneously, that intelligence without integration is just a very expensive demo.

The Numbers Behind the Hype

The growth curve here is one of the steepest job-market movements in recent memory. FDE job postings on Indeed jumped 729% year over year, from 643 postings in April 2025 to 5,330 in April 2026. Job postings overall reportedly grew more than 1,165% year-over-year, with New York overtaking San Francisco as the top hiring hub, largely on the back of fintech demand. New York now accounts for roughly 35% of FDE postings, against 11% for San Francisco.

On compensation, the median across all postings sits at $183,000, typically ranging from $160,000 at the 25th percentile to $215,000 at the 75th. That's the floor.

The ceiling is where the title earns its name. Vas laid it out directly on the podcast: FDE compensation runs from roughly $150,000 base plus considerable equity on the low end, up to $1 million a year for the best people, and he was explicit that he wasn't exaggerating. Senior FDEs at Anthropic and OpenAI reportedly clear $785,000 and above. Staff-level FDE compensation reportedly reaches $725,000 at frontier labs more broadly. Palantir, the role's originator, runs a bit more modestly by comparison, with average total compensation around $238,000 and staff-level FDEs clearing $630,000 or more.

Vas was also clear that the role isn't uniform. It varies wildly by company: some FDE roles are technically light, essentially configuring workflows and writing SQL on top of an existing platform, while others require shipping real production code on-site with a client. The pay follows how well a person combines the consulting side with the engineering side, which is precisely the intersection almost nobody occupies naturally.

What Isenberg and Vas Actually Said About the Job

Greg Isenberg's episode with Vas, "FDE: The $1M/Year AI Job Explained," starts from a single premise: every company can now buy the same frontier intelligence, so the real advantage has moved entirely to deployment. Vas traces the role back to Palantir, walks through the judgment required to decide where AI actually belongs in a workflow, and lays out the audit → evals → deployment loop that turns a raw model into measured business value.

That loop is the actual mechanics of the job, and it's worth understanding in plain terms. Audit means learning how the work really happens today, not the documented version of the process, which is almost always fictional by the time you actually sit with the people doing it. Evals means building a real test set and measuring the system honestly, including its failures, not just celebrating a pass rate. Deployment means putting the system into real operations with permissions, monitoring, and a rollback path, and then repeating the loop as production reveals edge cases the audit missed.

Isenberg's companion guide compresses the entire thesis into two sentences that stand as the cleanest definition of this role available anywhere: Forward deployed engineers bridge the gap between business problems and working AI systems. Companies pay $150K to $1M for the people who can do both.

That's the whole game. Not "knows AI." Bridges the gap between a business problem and a working system. Two different skill sets almost nobody has been forced to develop together until this exact moment, because until now, you didn't need both in one person. You hired engineers, and separately, you hired consultants. Now the model layer is commoditized, and the only thing left to compete on is who actually gets the thing built, deployed, and adopted inside a real, messy business.

The Part Almost Every FDE Explainer Misses

Every article about this role right now, including most of the ones covering Isenberg's episode, is written about the version of this job that exists inside Palantir, OpenAI, Anthropic, and AWS. Big labs, enterprise clients, seven-figure packages, case studies about logistics rerouting and government data pipelines.

That's real, but it describes maybe a few thousand jobs at a handful of companies with enterprise-scale budgets. What almost nobody is talking about is that the exact same audit-evals-deployment loop, the exact same "bridge business problems to working systems" mandate, is needed by literally every small and mid-size business in the country. A 40-person fitness equipment distributor has an integration wall. A dental practice has one. A DTC fashion brand doing seven figures a year has one. None of them are hiring a Palantir FDE at $400K a year. All of them desperately need exactly what a Palantir FDE does, at a scale that actually fits their business.

This is the gap I've built my practice around, and I think it's the single biggest unclaimed territory in this entire conversation. Call it what you want. I call it an AI Architect and Business Solutions Architect. The function is identical to a Forward Deployed Engineer: go into a business, diagnose what's actually broken, prescribe a system, and personally deploy it so it runs and performs its function, whether that's saving time, cutting costs, or scaling revenue.

The difference is scope, not skill. Instead of one FDE embedded for a year solving one enterprise's data pipeline, it's one person running rapid-cycle audits across many small and mid-size businesses, each needing a smaller but no less real version of the same integration work.

What the Role Actually Requires, and Why Almost Nobody Has All of It

Strip away the compensation numbers and the labs, and the job description is consistent everywhere it's described. You need three things at once, and this is exactly the intersection Vas describes when he talks about how hard these people are to find: you'll have consultants you then have to train on the technical side, or engineers you then have to train on the soft-skills side. It's genuinely difficult to find people who are the best of both.

Understanding the Business

Not the abstract, case-study version. The kind you only get from sitting inside real operations, sales, and margin long enough to know exactly where the money leaks. This is the piece almost every purely technical candidate is missing, and it's exactly why so many enterprise AI projects never move the P&L. Nobody diagnosed the actual business problem before reaching for the model.

Understanding the AI

Real technical depth: agent orchestration, evals, structured outputs, failure handling, not just familiarity with a chat interface. Enough to build the system yourself, end to end, without a lossy handoff to someone else.

Organizing the Business Around Systems That Actually Run

This is what separates an FDE from a consultant. A consultant hands you a report. An FDE builds the thing, deploys it, and stays until the people who have to use it every day have actually adopted it. The job isn't done until it runs without you standing next to it.

Almost nobody has genuine depth in all three simultaneously. That's not a knock on the market, it's just true, and it's exactly why demand is scaling exponentially while supply grows linearly, and why the compensation for the rare people who do have all three keeps climbing rather than settling.

Why This Matters Beyond One Job Title

The Forward Deployed Engineer conversation is really a proxy for a much bigger shift in how value gets created in a business right now.

For decades, the scarce resource was capability. Wanting a custom system meant a development team, months of time, and real budget. That scarcity is gone. Capability is cheap and abundant. AI startups building on top of that abundance are compounding at rates legacy companies structurally cannot match, not because their technology is secret, but because their systems and distribution are simply better, and in a winner-take-all market, better systems don't just win, they absorb the market share of everyone who doesn't have them.

What's scarce now is judgment. Knowing which problem is actually worth solving. Knowing which system to build and which to skip. Knowing how to sequence a deployment so a business doesn't break itself trying to modernize. That judgment is exactly what the audit-evals-deployment loop is designed to produce, and it's exactly the thing that doesn't show up in a model's parameter count.

That's why this role pays what it pays, whether you're talking about a $785,000 senior FDE at a frontier lab or a fractional AI architect working directly with founders. The scarcity was never the AI. The scarcity is the person who can look at a real, messy, specific business and know exactly what to build, and then actually build it.

Where I Fit Into This

I've spent the last three years building exactly this profile, mostly before there was a name for it.

Fifteen years inside the commercial fitness equipment industry, starting on a warehouse floor, eventually running marketing and operations for multiple companies in that space simultaneously. That's where the business instinct came from, the hard way, under real budgets with real revenue on the line, not from a case study.

The last three years have gone entirely into AI: building custom agents, automations, full-stack applications, and AI operating systems for businesses across fitness, healthcare, food service, martial arts, and now fashion. Not theorizing about it. Shipping it, in production, for businesses that had a specific problem and needed a specific system built and deployed to fix it.

That combination, business fluency earned inside real operations plus technical depth in AI plus the ability to actually deploy and hand off working systems, is precisely the three-part skill stack every source above independently describes. I just apply it at a different altitude than Palantir or OpenAI. I apply it to the businesses nobody else in this conversation is talking about: the founders and mid-size operators who need this exact function and will never make it onto a frontier lab's hiring page.

I'm not at the seven-figure outcome this role is named for yet. But the roadmap is the same one Vas lays out on the podcast: run the audit, build the evals, deploy the system, and let the case studies compound. That's the work. I'm doing it now, one deployment at a time, with the explicit goal of becoming exactly what this title describes.

If you're a founder sitting on the wrong side of that 95% statistic, wondering why the AI pilot never actually moved a number, that's the gap I work in. Not more intelligence. Deployment.