There is a comforting story that the AI age would be won by whoever owned the smartest model. It is already wrong. A new frontier model ships almost every week, and the moment it does, every company can rent it for the price of an API call. Walk into fifty enterprises and you find the same stack — the same coding agents, the same model-agnostic editors, the same copilots. If everyone can buy the same intelligence, intelligence cannot be the moat.
So the edge moves one step downstream: not who has the intelligence, but where, how, and why they use it. A general model is worth nothing to a specific company until someone bends it around that company's real processes, its real data, its real exceptions. That bending is the job. It has a name — the Forward Deployed Engineer — and it is the most valuable role the AI age has produced precisely because the intelligence around it has become free.
The failures prove the point better than the wins. The industry spent a year believing that if you pointed a good-enough model at a problem and spent enough tokens, it would sort itself out. The results are in: an oft-cited MIT figure puts the share of generative-AI pilots that fail at 95%. On Greg Isenberg's podcast, Veric Agents' Voss describes executives who burned through an eight-figure model budget meant to last a year in three months — and moved the needle on nothing, because they sprayed intelligence everywhere instead of deciding where it belonged. "Token-maxing" was never a strategy. Judgment is.
That judgment is the real deliverable. Take any ten-step back-office workflow and the honest answer is that most of it should stay deterministic software — if-this-then-that, API calls, a database write. Maybe three steps genuinely need a model's non-deterministic judgment, and one of them needs a human to approve before anything irreversible happens. Knowing which three, and defending that choice to the business, is worth more than any model. It is the difference between an 80%-accurate system nobody trusts and a boring one that actually ships.
This is why the role resists commoditization even as its raw material becomes free. It sits on two skills that rarely live in one person: the consultant's ability to read a business — its incentives, its politics, its undocumented exceptions — and the engineer's ability to build the thing and own it when it breaks in production. Models will keep getting cheaper and more interchangeable. The person who can walk into a bank or a ministry, find the workflow worth rebuilding, and make a frontier model survive contact with twenty-year-old systems will not.
The strategic takeaway for anyone reading this as a career bet: stop optimizing for access to the best model. Everyone has it. Optimize for the scarce thing on top of it — the judgment about where intelligence belongs, and the reps to prove you can install it without breaking the business. That is the whole thesis of this publication, and it is why the FDE, not the model, is the story.
Sources: Framing and figures drawn in part from "FDE: The $1M/Year AI Job Explained" (Greg Isenberg, with Voss of Veric Agents), YouTube: https://www.youtube.com/watch?v=zXysLUTLjw4 — watch it on The Broadcast. MIT pilot-failure figure as widely cited in 2025–26 coverage. Analysis and conclusions our own.