The most useful frame for how an FDE engagement actually runs comes from Colin Jarvis, who leads forward deployed engineering at OpenAI. It has three phases: scoping, validation, and delivery. Practitioners across the industry have landed on the same three cuts under different names, which is a good sign the structure is real.
The frame makes explicit what weaker engagements blur. Validation isn't a sub-step of delivery. It has its own deliverables and its own decision gate. Most failed AI deployments treat validation as a post-launch review nobody acts on. In a working engagement, validation decides whether delivery is worth attempting.
Phase one: scoping
Scoping lasts a few days. The FDE sits with users, watches the work, maps the processes, and leaves with a ranked list of things that could be built and an opinion about which one is most likely to move a number.
The most important question: is this one of the customer's top five priorities? If not, stop. That has nothing to do with technical feasibility and everything to do with organizational physics. A production AI deployment needs sustained senior attention for months. If the project isn't top-five, legal deprioritizes the review, IT applies the standard six-week request cycle, and the line workers treat the rollout as one more initiative to wait out.
The second job is checking that the stated problem is the real one, since sales conversations almost always produce a problem one level removed from the truth. The third is naming the sponsor: the senior person with the political capital to grant exceptions to IT, security, and procurement. Walk out of scoping without a named sponsor and you don't have an engagement. You have a mislabeled sales conversation.
The scoping ladder
Leo Mehr, who built Ramp's deployed engineering team from two engineers to about thirty, compresses this into a slogan: always be scoping.
His example: it's Friday night, and a sales rep arrives with an urgent message. A strategic logo will only close if Ramp builds an SAP S/4HANA integration. The engineering reflex is to open the SAP API docs. The trained FDE reflex is to ask what's driving the urgency (the customer's deadline, or the rep's quarter?) and then climb a ladder of questions. Who would actually use this? Have the workarounds been exhausted? Is there a manual bridge in the meantime? Could the customer's own engineers hit the API? And the most valuable rung: who else in the pipeline needs this? One customer asking is a favor. Five customers asking is a roadmap item.
Skipping the ladder costs real money. Mehr tells of a large customer that needed mobile reimbursements while Ramp's mobile team was swamped. Two FDEs taught themselves iOS and Android and built both apps in a few weeks. Then they asked for the list of Android beta users. The customer mandates iPhones for every employee. The question that would have saved half the build costs nothing to ask. Test the assumptions so basic nobody thinks to say them out loud.
Phase two: validation, the gate
In consulting, the next step after scoping is a statement of work. In FDE work, the next step is asking whether the scoped thing is worth building. Build the evaluation criteria, run the proposed system on a representative slice of real data, and present the results with a recommendation. If they clear the bar the customer cares about, proceed. If not, change the scope or end the engagement honestly.
Here's what the gate is worth, in a composite drawn from real engagements. A vertical AI startup was deploying a store-associate assistant for a four-hundred-store retailer. Scoping set a threshold of seventy percent on a hundred-case eval set. The team was tempted to skip validation and go straight to a one-store pilot, and the customer's own product manager was pushing for exactly that. The FDE lead held the line. The baseline came in at forty-one percent.
The failures weren't the expected ones. The system couldn't tell loyalty members from walk-ins, which changed the right answer to return questions. It also couldn't handle clearance markdowns governed by a rule set that changed every Friday. Both gaps were obvious in hindsight and invisible in scoping. Two weeks of fixes took the retest to seventy-three percent. The pilot launched in twelve stores and reached forty-eight by week ten.
Without the gate, a forty-one-percent system would have been in front of real shoppers, and the customer's confidence would have been very hard to win back. Validation looks expensive: one to three weeks of senior engineering time with no production code to show for it. The real cost of skipping it is four months of delivery against the wrong target, and then doing it all again.
Phase three: delivery
Delivery looks most like ordinary software engineering, with a few differences. Instrument the system from the first commit. Keep a customer-facing dashboard live from day one of production. Design for handoff from week one, because the FDE is leaving and the system has to keep running. The runbook, the alerts, the eval set, and the playbook for the customer's team are part of the deliverable.
The frame also works as a diagnostic for engagements in trouble. If the customer can't name the top-five priority the project serves, you're stuck in scoping. If the eval set exists but nobody on the customer side has signed off on the threshold, you're stuck in validation. If the system is live and the metric isn't moving, you're stuck in delivery. Knowing which phase you're stuck in tells you what to fix.
Thinking about moving into this work?
I wrote a free guide for people considering forward deployed engineering: what the 4.6x growth figure means, how to tell a real FDE role from a renamed sales job, and a 30-day plan that ends with your first piece of proof.