FDE: The $1M/Year AI Job Explained

Forward-deployed engineers (FDEs) connect general AI capabilities to a company’s specific workflows, and the role was described as paying from $150,000 in base salary plus considerable equity to as much as $1 million a year. The “million-dollar” FDE combines software skills with communication and business understanding: they can discover what a company needs, build a system end to end, and explain its value. As access to similar models and tools spreads, Voss argues, competitive advantage increasingly comes from where, how, and why a business deploys them.

The work starts with understanding how employees actually do their jobs, often by interviewing or observing them and examining systems such as ERPs and CRMs. Voss says FDEs spend a large majority of their time on site, where they can uncover tacit rules, exceptions, handoffs, and bottlenecks that written procedures miss. Palantir popularized the model as software-based consulting, pairing a software stack and data connectors with on-site workflow discovery and customized tools for enterprise, military, and government clients. Voss expects the AI era to require customized agents at a scale 100 times greater than the data era’s need to unify and visualize information.

A core FDE decision is where an LLM adds useful judgment and where deterministic software, existing automation, or no change is safer and more cost-effective. A strong implementation often combines deterministic steps, LLM reasoning, and human review; one suggested pattern uses AI for intake, validation, and drafting, then requires a person to approve further action. FDEs audit workflows, build evaluations, and deploy and monitor systems against adoption, KPIs, and SLAs. Audits map steps, exceptions, repetitive work, and judgment points, then prioritize high-volume opportunities by expected return. Historical examples can support evaluation—such as 10,000 emails or 5,000 presentations—but non-deterministic tasks are hard to score perfectly and need continuing human feedback. In one report, 41 of 50 runs passed; five failures involved missing data and four involved the wrong record, while unsafe cases should go to a person.

The speakers advise understanding the workflow before choosing or benchmarking a model, and suggest newcomers first become proficient with one model and one agent-building platform. Implementations should work with existing systems such as NetSuite, Salesforce, SAP, Concur, Expensify, Gong, and Workday, and can progress from testing to shadow mode, greater autonomy, and production. One cited example had C-suite executives use a $10 million Claude budget in three months, although it was meant to last a year, without materially moving the business. The speakers argue that AI spending alone may not improve results unless engineering connects the technology to actual business workflows.

Enterprise buyers want measurable, cost-effective value and may see an FDE as a personal risk if a failed change reflects badly on them. One recommendation is to reduce that risk with a free initial audit, demonstrate value, and charge afterward; the first one to three customers may be more valuable for learning than revenue. Demand for people who can apply AI inside companies was described as strong, while relatively few people have that expertise. The recurring cycle is audit, evaluation, deployment, observation, and improvement: fixing one workflow can expose upstream or downstream bottlenecks and lead to further opportunities.

Greg’s flexible 30-day preparation plan is meant to help candidates “do the job before you have the title,” rather than serve as an embedded customer engagement. Its first week is to build an agent for one granular workflow in areas such as finance, HR, procurement, logistics, or sales. The agent should complete the task reliably without depending on a user to prompt it perfectly, and include looping, tool use, guardrails, deliberate context and memory, and a full audit trail; clients need to inspect its traces to trust it. In week two, candidates make the demo recoverable with JSON schemas, validation, exception handling, and planned failure modes. Greg emphasizes designing for unhappy paths, contrasting one way a process can go right with roughly 1,500 ways it can go wrong.

Week three adds retries, a golden dataset for evaluation, ongoing improvement, and economic testing, including whether cheaper or less advanced models such as Gemini Flash and Muse Spark can handle particular tasks. Candidates should measure time saved, risk mitigated, revenue uplift, or cost savings, and know the agent’s performance, failure modes, and costs. The final week is about defending the system: its business problem, why AI fits, architecture, iterations, evidence, risks, and economics. Greg cited accuracy improving from 70% to 95% and recommends explaining the work both as an engineer and as a VP. He says pitching an agent to businesses can test whether its solution and economics meet customer needs; Voss adds that hands-on, self-directed learning may be advantageous while universities take time to teach these emerging skills.