There is a job in AI right now paying up to $1M a year.
Not a founder. Not a researcher with a PhD. An engineer who shows up at a company, sits next to the people doing the work, and makes AI actually stick.
It's called a forward deployed engineer. FDE for short.
Greg Isenberg just did a full breakdown with Vas from Varick Agents, an ex-Meta engineer who now builds AI agents for enterprises. Vas laid out the entire playbook: what the role is, why it pays so much, and a 30-day plan to do the job before anyone gives you the title.
I watched the whole thing so you don't have to. Here's the episode, and below it, everything worth stealing.
What a forward deployed engineer actually does
Here's the core idea, and it explains the salary.
Every company can now buy the same intelligence. Claude Code, Codex, Cursor. Same models, same APIs, same prices. Intelligence is a commodity.
So the edge moved. It's not "who has AI" anymore. It's "who deployed it where it actually matters."
That's the FDE's whole job. You get deployed inside a client's business (often literally on-site), you learn how the work really gets done, and you decide where AI belongs and where it doesn't. Then you build it, prove it works, and defend the numbers.
Half consultant, half engineer. Vas calls it a rare "art plus science" combination, and that rarity is exactly why the comp is insane.
Palantir invented this, and now everyone wants it
The role comes from Palantir. Vas lived in New York for years and a bunch of his friends were Palantir FDEs, so he's seen it up close.
Palantir's model: they built a customizable ontology (a software stack full of connectors and data links that pipes enterprise data into one interface). Then they sent engineers on-site to clients, the military, the government. Those engineers learned each client's workflows and spun up dashboards and agents tuned to that specific business.
Consulting, but for the software age. That's what made Palantir take off. The platform wasn't the magic. The customization per client was.
The thesis now: if it worked for Palantir in the data age, the AI age will demand it 100 times more.
How much FDEs make
Greg asked the question directly: someone in New York wants this job, what do they make?
Vas's answer: "A lot of money. This is the hottest role in technology right now."
→ $150,000 base with considerable equity on the low end.
→ Up to $1M a year at the top. His words: "I'm not joking."
The catch is the role varies wildly. At some companies you're barely writing code, you're configuring workflows and writing SQL on top of an existing platform. At others you're shipping production code on-site with a client. The pay follows how well you combine the consulting side with the engineering side.
Why 95% of AI pilots fail (and why that's the opportunity)
Vas brought up the MIT stat: 95% of generative AI pilots fail.
He's watched it happen. One C-suite exec he talked to blew through an entire $10 million AI budget in 3 months. It was supposed to last a year. They gave AI to everybody, slapped it on everything, and got token-maxing and hallucinations instead of results.
The fix is what Vas calls FDE judgment: deciding where intelligence belongs and where it doesn't.
Say a client has a 10-step workflow. Maybe only 3 of those steps need actual judgment from an LLM. The other 7? If-then statements and API calls. Deterministic software. Cheaper, faster, and it doesn't hallucinate.
Most people building with AI right now skip this step entirely. They hand the whole workflow to a model and pray. It's the same lesson from the AI audit business breakdown: the money isn't in the AI, it's in knowing where to point it.
The documented process is never the real process
This was my favorite part of the episode.
Ask an employee "what's the first step of your workflow?" and they'll say "an email arrives."
Sounds like a clean trigger. It's not. The email arrives from 40+ senders. No two are formatted alike. The data is in a PDF, or a screenshot, or an Excel file, or buried in a forwarded thread. Half the cases are exceptions. "Sarah already signed off on this one." "It ignores the second attachment." There's no consistent subject line.
And the real routing logic lives in one person's head. Nobody wrote it down. If you don't sit with that person, they won't even remember to tell you.
That's why FDEs go on-site for the full 8 to 10 hours instead of scheduling a 1-hour call. In a meeting, people describe the job they think they do. Sitting next to them, you watch the real one. Even McKinsey consultants sit in mines with the miners for the same reason.
Build for the documented process and you've built a system that doesn't map to reality. That's where the 95% failure rate comes from.
The loop: audit → evals → deployment
The actual FDE motion is a loop you run over and over:
→ Audit. Find the workflow worth rebuilding. High ROI, real pain, exceptions everywhere.
→ Evals. Turn non-determinism into evidence. Build a golden data set, measure accuracy, bake in human feedback so the system improves.
→ Deployment. Build on the client's existing systems. Don't rip out the ERP they've used for 20 years. Start in shadow mode, increase autonomy, then go to production.
That last point matters more than people think. You're pitching humans at a company, and humans don't want to get fired. They want to get promoted. "Switch your entire stack" is a terrifying pitch. "I'll make the stack you already trust more efficient" is a no-brainer.
Each loop you complete makes the next one clearer. That's the compounding.
One way to go right, a thousand ways to go wrong
Vas said something that applies way beyond FDE work:
"There's only one way that something can go right, but there's a thousand different ways something can go wrong. If you're only building for the way it goes right, you're worth nothing."
An agent that handles the happy path is a demo. An agent that handles the 1,500 ways the workflow breaks is worth, in his words, a million times more.
This is the difference between the AI demos flooding your feed and the systems companies actually pay for. Same models. Completely different value.
The 30-day plan to become an FDE
Vas started as a Meta engineer with zero consulting experience. He compressed a year of learning into a 30-day roadmap. Here's the structure:
Week 1: build an agent that completes a real loop. Pick one real back-office workflow (finance, HR, procurement, logistics, sales). Get it in granular detail. Build an agent with tools, guardrails, deliberate memory, and a full audit trail. His bar for "agent": you can prompt like an idiot and it still gets the job done.
Week 2: make it recover. Defined JSON schemas instead of free-form text. Schema validation. Failure modes and exception handling. This is where you build for the unhappy paths.
Week 3: make it measurable and economically viable. Retry logic. A golden data set for evals. Swap cheaper models in for subtasks where a frontier model is overkill. Then measure against the only three buckets a business cares about: revenue uplift, risk mitigation, cost savings.
Week 4: defend the system. Rehearse it twice. Once as an engineer (architecture, decisions, how accuracy went from 70% to 95%). Once as a VP (the problem, the outcome, the evidence, the risk). If you can't defend it in both rooms, you're not done.
By day 30 you have a production-grade agent with known failure modes, measured costs, and a story you can sell. You did the job before you held the title.
Why founders should care (even if you never want the job)
Here's my take, because most of you reading this aren't trying to get hired anywhere.
This playbook is a services business hiding in plain sight.
You don't need Palantir on your resume to audit a local company's workflows, find the 3 steps that need an LLM, and charge for the outcome. The businesses drowning in exceptions and manual email routing are everywhere, and the people who can do both the listening and the building are still rare.
And if you're building a SaaS, this is your competition brief. AI agents are eating the seat-based software model, and the winners are the ones who go deep on one workflow's ugly exceptions, not the ones with the prettiest demo.
One more thing worth stealing: Vas's company is model-agnostic, but his advice for individuals is the opposite. Pick one ecosystem (OpenAI's agent stack, Claude's agent SDK, whatever), and get very, very good at it before you branch out. Depth first. Benchmarks later.
FAQ
What is a forward deployed engineer?
A forward deployed engineer (FDE) is an engineer embedded directly with a client, often on-site, who learns how the business actually works and then designs, builds, and deploys AI systems tuned to that specific company. The role mixes consulting-grade communication with production engineering. Palantir popularized it; the AI wave made it the hottest role in tech.
How much does a forward deployed engineer make?
Per Vas from Varick Agents, FDE compensation runs from about $150,000 base plus considerable equity up to $1M a year for the best people. The top of the range goes to the rare engineers who combine strong consulting skills with the ability to ship production AI systems.
Do you need a computer science degree to become an FDE?
No. The role varies wildly by company. Some FDE roles need real production coding chops. Others are technically light: configuring workflows, writing SQL, building on top of an existing platform. What every version needs is judgment about where AI belongs in a workflow, plus the communication skills to extract how work really gets done.
How do you become an FDE in 30 days?
Vas's roadmap: week 1, build an agent that completes one real back-office workflow end to end. Week 2, harden it with schemas, failure modes, and exception handling. Week 3, add evals, a golden data set, and cost optimization, then measure revenue uplift, risk mitigation, and cost savings. Week 4, rehearse defending the system to both engineers and executives.
Is the FDE playbook useful for founders?
Very. The same loop (audit a business, find the steps that need intelligence, build for the exceptions, prove the ROI) works as a standalone AI services business or as the discovery process for your next SaaS. The skill that pays $1M as a salary pays even better as an owner.
Steal playbooks like this every week
I interview bootstrapped founders making $100K to $10M a year and get them to hand over the exact numbers and systems behind their growth.
No fluff. Just operators showing their work.
