Forward Deployed Engineer is the hottest role in AI. Every lab is hiring. Almost none of them will succeed.
1,165%
FDE postings, YoY
$300K
Anthropic FDE top base
70%
Will abandon by 2028
THE THESIS
Vendor FDEs ship the demo and leave. Your team inherits the hard part: production hardening, on-call, eval regressions, cost. The role only pays off when whoever embeds also stays for the cliff.
What an FDE Actually Does
Palantir invented the role in 2006. Send engineers into customer infra. Give them production access. Let them ship. Every AI lab copied it.
EMBED
On customer infra. SSO, VPN, repo access. ~25% travel to site.
SHIP PROD CODE
Agents, MCP servers, sub-agents, integrations. In the customer’s repo, not a sandbox.
BUILD EVALS
On real customer traffic. The eval suite is where workflow knowledge accrues.
FEED THE PRODUCT
Patterns from the field flow back into the platform. Customer becomes the R&D edge.
The Demo-to-Prod Cliff
Here is where the role breaks. Vendor FDEs are measured on logos and reference customers, not on uptime. The incentive ends at the demo. Then they leave — exactly when the work gets brutal.
Vendor FDE here — clean data, fresh repo, demo-day pressure.
Your team alone here — 3am pages, eval drift, infra bills, security review.
The PoC that wowed the steerco is not the system that survives a Black Friday. Latency budgets shift. Edge cases multiply. The eval suite that passed in week four starts regressing in week twelve. Cost per request triples once real traffic hits. None of that is in the FDE’s scope.
Why It Keeps Happening
Three forces push every AI lab toward the same pattern.
- The deployment gap. Pilots stall on SSO, ETL, security review — not on prompts. Vendors send FDEs to unblock the demo, not to own the system.
- Evals need real data. The eval suite must live on customer traffic. If it leaves with the FDE, the only thing that compounded just walked out.
- Value moved down the stack. Models commoditise. The integration and hardening layer is where margin lives — and where vendor FDEs aren’t paid to stay.
“Frontier AI isn’t a product yet. CIOs thought they were buying software. They’re actually buying a professional services engagement.”
— Nik Kale, Coalition for Secure AI
Vendor FDE vs Production Squad
Vendor FDE
- Paid by Anthropic / OpenAI / Cursor
- KPI is logo + reference, not uptime
- Exits at pilot — next account waiting
- Owns the demo, not the on-call rotation
- Eval suite leaves with them
Production Squad
- Paid by you, vendor-agnostic
- KPI is system in production + hand-off
- Stays through hardening and cost work
- Owns the cliff your team would face alone
- Evals + runbooks live with your team
The labs see this too. OpenAI didn’t build its own consulting arm — it bought Tomoro and spun it out as a separate entity. Salesforce routes its FDE work through Accenture, Deloitte, IBM, KPMG, PwC, Slalom, and TCS. The vendor FDE writes the cheque-grabbing demo. The production squad ships the system that survives.
The Eval Loop
The hidden centre of the role. Production data lives inside the customer. Whoever owns the eval suite owns the workflow.
const evalSuite = new EvalRunner({
dataset: customer.productionTraffic({ days: 30 }),
graders: ['task-success', 'tool-correctness', 'hallucination'],
guardrails: customer.policy(),
});Three Signals of a Healthy Engagement
SIGNAL 01
Effort falls, doesn’t flatten
Month six should not look like month two. Flat = dependency.
SIGNAL 02
Evals stay with you
If the eval suite walks out with the FDE, the only thing that compounded just left.
SIGNAL 03
Internal engineers ship
By month nine your team should be writing agents without the FDE in the room.
The Bottom Line
The vendor FDE is real and useful. They ship the demo your steerco needs to see. But they leave at the worst moment — when PoC meets production and the work stops being interesting. Bring someone in who’s paid to stay for the cliff. Keep the evals. Own the hand-off.
Agentika operates as an embedded FDE squad for engineering teams shipping agentic AI in production. Let’s talk.