FDE: The $1M/Year AI Job, Explained

TL;DR

Every company can now buy the same intelligence — the frontier models ship weekly and everyone runs the same stack (Claude Code, Codex, Cursor, Copilot). So intelligence stops being the moat; the edge moves to deployment: applying general intelligence to one company's specific, messy processes. The person who does that is the Forward Deployed Engineer — a Palantir-coined role that's now the hottest job in tech. It pays because it demands a rare combination: consultant + engineer, art + science, in one person. The work is a repeating loop — audit → evals → deployment — and Vas Moza lays out a 30-day plan to build the proof solo and "do the job before you have the title." Why most AI spend fails: companies bought intelligence and skipped deployment (~95% of GenAI pilots fail; one exec torched a $10M budget in 3 months "token-maxing").

Key takeaways

• Intelligence is commoditized. Same models, same stack, available to anyone who pays. The moat is no longer who has intelligence — it's where, how, and why they deploy it.
• The FDE is the bridge. Between the business reality and the intelligence stack. Whoever builds the best bridge wins — that's the role.
• It's the best-of-both, not the average. Left aisle = consulting (workflows, incentives, risk, adoption, politics). Right aisle = engineering (models, evals, guardrails, harnesses). The $1M hire is genuinely strong on both.
• Documented process ≠ real process. "An email arrives" is really 40+ senders, PDFs/screenshots/Excel, half exceptions, routing logic living unwritten in one person's head. You only get it by sitting on-site.
• Judgment is the job: where does intelligence belong — and where not? Majority deterministic software, LLM only at real judgment nodes, human-in-the-loop before irreversible actions.
• The audit is the wedge — worth 10× its price, "better than McKinsey." Rebrand it ("sprint"), do the first one free to de-risk, get paid on results.
• Build on top of what exists. Never force a migration off their ERP — you'll get thrown out. Integrate, then stage: test → shadow → autonomy → production.
• Measure only three buckets: revenue uplift, risk mitigation, cost savings. No number in one of them = no proof.
• Build for the unhappy paths. "One way it goes right, a thousand ways it goes wrong. Build only the happy path and you're worth nothing."

What an FDE actually is

The term comes from Palantir (originally a military term — engineers deployed forward, to the field). Palantir built an ontology + data-lake platform whose real power wasn't how tech-forward it was — it was how customizable it was. FDEs deployed on-site with enterprise and government clients, learned their workflows, and spun up the dashboards and agents that solved their specific problems. Consulting, but for software. Vas's thesis: if it worked for Palantir in the data age, the AI age demands it 100× more — every company will need customized agents, and the FDE is why AI actually becomes useful to a business.

It runs in three stages, and each one feeds the next:

[14:58]

Stage 1 — understand the business reality

This is where the bulk of the time goes, and where tech people get it wrong. The documented process is almost never the real one. Vas's example: an email arrives. Sounds like a clean trigger. It isn't.

Go on-site (or deep-remote when you must). A one-hour meeting gets you the person's idea of the job; a full day gets you the job — including what breaks and how they quietly fix it. Interview multiple people; the same workflow usually spans several heads.

Stage 2 — FDE judgment: where does intelligence belong?

The early-AI reflex was "slap AI everywhere, let the model figure it out" — token-maxing. That's what produced the hallucinations, the blown budgets, and the ~95% pilot-failure rate. The FDE's real value is deciding, step by step, where intelligence actually belongs.

Stage 3 — deploy, and the audit that opens the door

Stage 3 is build → evals → deployment. But the thing that gets you in the room is the audit.

[27:08]

Every engagement starts with a paid audit that maps each workflow, its exceptions, and an ROI matrix of what's worth automating. Clients have said the audit alone was worth 10× what they paid — "better than McKinsey," because AI is new enough that no one else can map it. Practical moves:

• Rebrand "audit." People react to the word like a tax audit. Call it a sprint, a current-state analysis, current vs. future state.
• De-risk yourself. To the buyer, you are the risk — status quo is safe; you might fail and make them look bad. Do the first audit free, prove value, get paid on results. Your first 1–3 clients teach you more than you give them.
• Sell to the human. They don't want to get fired; they want to get promoted. Help them point to "I drove this value" at review time. Pitch value cost-effectively — not "rip out your ERP."
• Build on top of what exists. A client spent years and millions moving to NetSuite; tell them to leave it and you're gone. Integrate on top, connect the rest (Salesforce, SAP, Concur, Workday, Gong), then stage the rollout: test → shadow mode → increasing autonomy → production.
• Audit trail is non-negotiable. If you can't show the client what the agent did — logged, inspectable traces — they will never trust you.

The 30-day plan: do the job before you have the title

You can't get embedded without being an FDE first — so build the proof solo. One focus per phase; space it out if 12-hour days aren't realistic.

Measure only the three buckets

An agent's value always reduces to the same three numbers. If you can't put a figure in at least one, you haven't proven anything.

My take

This maps almost one-to-one onto what I'm already building — the Content Studio, the Hermes router, the "own your context, invest in the harness" thesis. The reframe that stuck: the money isn't in the intelligence, it's in making it land inside one specific, messy business — and that's a communication job as much as an engineering one. The 30-day plan is the useful part: it's a concrete way to manufacture proof before anyone hands you the title. I turned the whole method into a reusable forward-deployed-engineer skill so any of my agents can run an engagement to the same checklist.

Worth a healthy dose of salt: as the comments point out below, a lot of this is a rebrand of the systems-analyst / business-analyst / solutions-engineer / management-consultant role that's existed for decades — and both guests have something to sell (Verity Agents; a hinted-at paid program). The novel part is real, though: super-intelligence on tap genuinely is new, and "someone has to decide where it belongs" is a genuinely different job than it was five years ago.