The hottest job in enterprise AI right now is not a sign that AI is working. It's a sign that it isn't.
The role is the forward deployed engineer, the FDE. Job postings for it grew roughly 729% year over year through this spring, and the money behind it has gone from a hiring trend to a balance-sheet commitment. If you run technology inside a large organization, this role is going to show up in a vendor pitch or a consultant's staffing plan within the next two quarters. Worth understanding what it actually is, and what its sudden arrival is telling you about your own AI program.
The 95% problem
Start with the number that explains everything else. MIT's research on enterprise AI found that 95% of enterprise generative-AI pilots produce little to no measurable impact on the bottom line. Not because the models are weak. Because the work of making them function inside a real company, with its messy data, its integrations, its compliance reviews and its politics, barely gets done.
That gap has a name now. Call it the integration wall. The bottleneck in enterprise AI has moved from model capability to deployment, and the FDE exists to close that specific gap. An FDE is an engineer who embeds with a customer, learns the business from the inside, and builds the last mile that turns a capable model into something that actually runs in production.
Here's the part that matters for a leader. FDE demand is a derivative of enterprise AI struggle, not enterprise AI success. As long as most pilots stall before they reach P&L, someone gets paid to close that distance. And because the gap is organizational rather than technical, better models don't shrink it. They widen it, by increasing what's possible faster than any enterprise can operationalize it.
The money is the tell
The clearest evidence this is not a passing fad is who is funding it, and how.
Over a few months this year, the labs and hyperscalers put roughly $9 billion into dedicated deployment ventures. OpenAI stood up a Deployment Company capitalized at over $4 billion with institutional backers. Anthropic and Blackstone launched an implementation venture reported at $1.5 billion. Microsoft committed $2.5 billion to its own AI deployment unit, with about 6,000 embedded engineers. Amazon went its own way, funding a $1 billion internal forward deployed organization from its own balance sheet rather than a joint venture. Accenture stood up a Microsoft-partnered FDE practice and talked about thousands more.
You don't structure billion-dollar ventures with private-equity partners around something you expect to fade by autumn. These are multi-year bets. The labs have concluded that the money in enterprise AI is not only in the model. It's in sitting next to the customer until the model works.
What an FDE actually is
Skeptics call it a rebranded sales engineer, and in weak implementations they're right. But the real version predates the hype by more than a decade. Palantir built its business on this model, and its alumni now seed FDE teams across the industry. Its August results showed why it keeps pressing the point: revenue up 93% year over year, US commercial revenue up 149%, and a CTO happy to jab that everyone else just has sales engineers in a nicer costume.
The work is engineering, not sales. These roles carry equity, not quotas. An FDE spends roughly half the time customer-facing, doing discovery and reading the room, and the other half building. The lifestyle cost is travel, commonly 30% to 50% of the time, and that cost is precisely why the supply of people willing to do it stays thin while demand climbs. Teams deploy in small pods embedded inside the customer's environment for days or weeks at a stretch.
The scarcity is not hypothetical. One recruiting analysis this summer counted roughly 17,000 FDEs in the US but only about 2,000 who have repeatedly shipped enterprise AI into production, even as the share of companies planning to hire them jumped from under 10% to about 70% in half a year. That gap between the badge and the real capability is the whole game.
The skill that separates a real FDE from a strong generalist is applied LLM work: retrieval pipelines, agents, and above all evaluation frameworks. If you remember one technical detail, make it evaluation. The ability to prove a deployment works, with instrumented evals rather than a convincing demo, is the thing these teams are actually selling.
In Canada, the banks already do this. They just don't call it that
This is where the Canadian picture gets interesting, and where it diverges from the US headlines.
No Big Five or Six bank posts a role titled forward deployed engineer. Neither do the big insurers. They hire the same capability under names like senior AI engineer or AI solutions engineer, partly because provincial regulation restricts who can use the word "engineer" in a title. So the raw job counts understate what's really happening.
The demand is real, and it runs through partnerships. The anchor is RBC and Cohere, whose "North for Banking" platform, launched in early 2025, is a secure generative AI system built specifically for financial services. Cohere's Toronto FDE postings name finance first and RBC among their clients. PwC Canada is building FDE pods and openly cross-training its finance people in tech and its tech people in finance. Its AI lead calls this the path with the biggest impact for clients.
The pattern for a Canadian financial institution is clear. You will consume FDEs through a vendor or a consultancy long before you post the title yourself. Which makes the real question not "should we hire FDEs," but "how do we buy this capability well."
Build, rent, or absorb
There are three ways to bring FDE capability into a regulated enterprise, and the right one depends on where the work sits relative to your core.
Rent it through the labs and their ventures. Fastest access to top talent and frontier tooling. You pay a premium, and the engineers wear the vendor's badge, which matters when the deployment touches sensitive systems and data residency.
Partner through a consultancy or a domain platform. The RBC-Cohere and PwC model. This is the default for financial services because it wraps the capability in a security and procurement envelope you already trust, and the platform carries the regulatory work you can't outsource casually.
Absorb it under your own titles. Build the capability in-house as senior AI engineering roles. Slowest to stand up, but the only option that keeps the deployment muscle, and the institutional knowledge, inside your walls. For anything close to your core, this is where you want to end up.
Most large institutions will do all three at once. The mistake is doing them by accident instead of on purpose.
How to tell a real one from sparkling sales engineering
Barry McCardel, who runs Hex and came out of Palantir, put the warning plainly: slapping a fashionable title on your field team is easy, building a genuinely forward deployed culture is not. Without the culture, what you get is sparkling sales engineering.
When a vendor pitches you an FDE engagement, three questions separate the real thing from the costume. Does the deployment work actually feed the product roadmap, or does it die in your account? What can the field team build without waiting on headquarters approval? And does the vendor treat early pilots as R&D they're willing to lose money on, or as billable hours from day one? The honest FDE model tolerates money-losing pilots because the learning compounds into the product. The imitation bills you for the privilege of being a reference customer.
Two quick proxies while you're reading the posting or the statement of work: how much travel the role carries, and whether evaluation engineering shows up as a named requirement. Both tell you fast whether you're looking at deployment or decoration.
Why this should sound familiar
I wrote recently about harness engineering, the discipline of building the environment, constraints and feedback loops that let AI agents do reliable work inside your codebase. The FDE is the same insight, one layer out. Harness engineering makes the model reliable inside your software. The forward deployed engineer makes it reliable inside your enterprise. Both start from the same admission: the model was never the hard part. The environment around it was.
There's a fair worry that FDE becomes the next prompt engineer, a title that booms and collapses inside two years. It won't, and the reason is structural. Prompt engineering was a thin layer over a moving interface. FDE work bundles production engineering, systems integration and enterprise change management, an old and thick set of skills that AI made more valuable, not less. The title may dilute. The work won't.
So treat the FDE surge as a diagnostic, not a trend to chase. If the market is spending $9 billion to put engineers next to enterprise customers, that is the clearest signal yet that the value in AI has moved from having the model to landing it. Your pilots aren't failing because you picked the wrong model. They're failing at the integration wall. The only question worth asking is who's going to help you climb it, and whether they're the real thing.