How a Forward Deployed Engineer Ships an Agentic AI Workflow in 4 Weeks
TL;DR
A Forward Deployed Engineer is an engineer embedded inside a customer's workflow whose only job is shipping the AI system to production. The role skips the requirements-doc handoff. Four weeks is a real timeline when three conditions hold. Scope is narrow. The customer has already granted data access, and one business owner can decide. Miss any of these and the calendar slips. The 2025 hiring surge at OpenAI, Anthropic, Ramp, and Google Cloud tracks this pattern — buyers now expect production behavior on a defined date. This post is the working recipe: what fits in four weeks and what breaks the schedule. The last section names when to reject the deadline instead of missing it, and where to go instead.
What a Forward Deployed Engineer actually does
Palantir named the function two decades ago: put an engineer into the customer site with commit rights on both sides. The engineer closes the loop end to end. Modern AI vendors adopted the pattern without renaming it. Colin Jarvis, OpenAI's Head of Forward Deployed Engineering, described the job to The Pragmatic Engineer in August 2025: “FDEs work in a ton of ambiguity, and often what the customer describes in scoping doesn't match the data/system reality on the ground.” That gap between the scoping doc and the real production data is the four-week problem. An FDE closes it by writing code against live systems inside week one. Discovery reports are what the failed projects produce in week three. Jarvis said the OpenAI FDE org started 2025 with two engineers and grew fast on shipped systems.
Why this matters now
The math on AI project delivery got worse before it got better. RAND put AI project failure above 80% — roughly double the rate for non-AI IT work. MIT's 2025 study of over 300 enterprise AI implementations found that fewer than 5% of integrated pilots produced measurable financial value. Buyers responded. They shifted spend from vendors who write specs to vendors who ship code inside the customer's VPC. AWS announced a $1 billion Forward Deployed AI engineer investment in November 2025, and Ramp built its own Applied AI FDE function from a standing start in late 2024. Anthropic's FDE listings call for MCP servers and sub-agents in production. Google Cloud posted the same shape of role this year. The listing names customer-embedded delivery experience.
Where FDEs win
The 4-week calendar assumes four working conditions. Scope covers one durable workflow; platform migrations don't fit. The customer has already signed a data-processing agreement and provisioned access. One executive can approve the eval criteria without a committee. And the agentic system will run against a real, funded process at go-live. Demo benches don't count. When all four hold, week one goes to shadowing the actual user and instrumenting the failure points. Week two ships a working end-to-end path against the customer's staging data. Week three is prompt hardening and red-team runs against the sub-agents, with evals updated on the failures. Week four is production behind a feature flag with the operator watching outputs. Skip any of the four conditions and the schedule slips. In practice it lengthens to eight or twelve weeks.
Where FDEs aren't the answer
The Forward Deployed Engineer model does not fit every AI project. A vanilla RAG chatbot over public docs is faster to buy than to deploy; an off-the-shelf vendor closes it. Broad, undifferentiated rollouts across ten thousand seats are staff-aug work. Hire a partner with bench depth and template implementations. Regulated environments with vendor-distance requirements (federal contracts, custody-model banking) can prohibit the FDE model entirely. The customer's compliance team may forbid write access from an external engineer. And if the customer has no live production process to embed against, only a strategy deck about future workflows, an FDE will invent scope on their behalf. That ends badly. One more anti-pattern: retrofitting an FDE onto a fixed-price statement of work. That contract shape optimizes for invoice items over production outcomes.
What to do next
Buyers thinking about a four-week engagement should stress-test three questions before signing. Is scope one workflow, or three? A three-workflow scope needs a twelve-week engagement or a smaller ask. Who signs off on evals? If the answer contains the word 'committee,' the four-week clock cannot start. Does the operator sit next to the engineer during week four? A hand-off email is not deployment. A well-run Forward Deployed Engineer engagement produces a running agent inside the customer's stack. It also produces an eval harness the customer's own team can extend beyond week four. Anything shorter is a demo. Anything longer is a consulting engagement in disguise. Ramp's Applied AI leadership talked about this discipline publicly during their late-2025 hiring push.
Want to see Lynk against your own workflow? Book a build session and we'll prototype it in front of you.
Frequently asked questions
How long does a Forward Deployed Engineer engagement typically last?
A Forward Deployed Engineer engagement typically runs four to eight weeks for a first scoped workflow. Longer contracts exist at Palantir and OpenAI, but the initial ship should be scoped so four weeks is realistic when inputs are ready.
Which companies hire Forward Deployed Engineers in 2025?
OpenAI, Anthropic, Palantir, Ramp, Google Cloud, AWS, Databricks, Salesforce, and Commure all publicly hire Forward Deployed Engineers as of 2025. AWS announced a $1 billion Forward Deployed Engineer investment in November 2025.
What is the difference between a Forward Deployed Engineer and a solutions engineer?
A Forward Deployed Engineer writes production code inside the customer's environment and owns the resulting system's behavior. A solutions engineer supports the sale and produces reference architectures without owning what runs in production.
Can a Forward Deployed Engineer work remotely?
A Forward Deployed Engineer can work remotely when the customer grants full data access and one operator makes decisions. On-site sprints still shorten ambiguity resolution, which is why OpenAI and Palantir maintain travel-heavy Forward Deployed teams.
When does the four-week Forward Deployed Engineer timeline actually work?
The four-week timeline for a Forward Deployed Engineer works when scope covers one workflow and data access is already provisioned. It also requires one executive with authority to approve evals and a live production process ready to receive the agent.