The requirements document ran to forty-seven pages. A large shipping-and-dispatch company's VP of operations had specified a custom BI dashboard with fourteen metrics, drill-downs, and alerts, at an estimated three months of development. Palantir's team took it at face value and spent four months scoping against it.
Then one engineer, on site mostly because their family lived nearby, asked a dispatcher a simple question: what's the first thing you do with this information on Monday morning? Check whether trucks are late, the dispatcher said, and call to reroute inventory if they are. The forty-seven pages compressed into a Slack alert. The team built it, start to finish, in four hours.
The story comes from Vinoo Ganesh, an FDE from Palantir's earliest cohort who later ran business engineering at Citadel and is now co-founder and CEO of the verifiable-AI startup Kepler. His summary: "Customers describe solutions, not problems." The requirements document was the customer's proposed solution. The Monday morning was the problem.
The most common reason an FDE engagement fails has nothing to do with the technology: the team picks the wrong problem. Three months in, there's a working system that nobody uses. The system works and the deployment still fails.
Why the stated problem drifts
Sales conversations almost always produce a stated problem that sits a little to one side of the real one, even when it's on leadership's top-five list. The person describing it is one level removed from the work. They may have a political reason to frame it a certain way. They've already filtered it through what they think your product can do, and they speak the language of the executive summary instead of the language of the workflow.
Good interview technique (The Mom Test is the place to start) helps, but interviews have a ceiling. People describe their work as it's supposed to happen, and the deployment usually lives in the gap between that and what actually happens.
Ganesh's second story shows how wide that gap can get. A data-quality engineer at a customer had blocked a CSV-to-Parquet migration for a year. The pipelines were straining under roughly a terabyte a day of CSV drops. She was adamant that Parquet didn't work.
On site, watching her run the workflow end to end, the team finally saw what no meeting had surfaced. Her quality checks consisted of downloading CSVs to a Windows machine and double-clicking them to spot-check the data. Parquet files had no viewer she could double-click. Her objection had always been about the workflow the format would break, a requirement that appeared in no document because nobody, including her, thought of it as a requirement.
The team built a Parquet viewer overnight. She approved the migration within two days, and pipeline execution fell from seventeen hours to about two. A year of blockage dissolved after an afternoon of watching.
What to watch for
Ganesh's method rests on a short list of tells, each marking a spot where the real workflow diverges from the described one:
- any task a user repeats, whether several times an hour, a day, or a week
- copy-paste between tools
- switching between tools or tabs
- pulling out a phone mid-task
- the flat "well, I have to" tone when you ask how they feel about a piece of their work
His opening question on site is tuned to surface them: what's the most annoying part of your morning? And his standing instruction to FDEs who'd rather scope from the conference room fits in four words: "Residents get the truth." Your badge at the customer's site and your contractor email address, he says, are your data-mining permits.
Four stakeholders, two tests
Getting the real problem shipped takes a map of the customer, who is at least four people:
- The sponsor, the senior person whose top-five priority your project serves. An enthusiastic executive who can't direct procurement, IT, or security is a fan.
- The blocker, often the head of IT or risk, whose job is to prevent exactly the kind of change you represent. Ignore them and you'll learn, six weeks before launch, that you face a security review that takes ten weeks to clear.
- The line user, whose daily work the system is supposed to change. They can stall a deployment simply by not using it, and most failed deployments lose them in week two.
- The regulator, internal or external, who has to bless the system before it touches production data or decisions. Their early support is worth more than their late approval.
Two tests then decide whether a candidate problem is worth taking. The gravel-road test asks whether you can build a working version in two to four weeks. If you can't, the problem is too big and has to be decomposed, because the customer is picturing a transformation while you're scoping a deployment. The asymmetry test starts from an uncomfortable fact: you know you can ship, and the customer, having watched many large projects fail, doesn't believe anyone can. The way through is to take more risk than a standard contract asks, by doing a small piece of work at your own cost or tying most of the payment to a milestone you control.
The bottleneck wasn't inspection
A composite drawn from real engagements in contract manufacturing shows all of this at once. The customer made precision components for aerospace primes. The stated problem was "we want AI to automate our quality inspection." The sponsor, a VP of operations, was clear at kickoff that inspection was the bottleneck, and late-delivery penalties were running into seven figures a year.
The discovery week broke the brief in the first interview. Inspectors said the inspection itself took fifteen minutes per part. The report took forty-five. After each inspection they dictated findings onto a paper form, walked it to a shared workstation, transcribed it into the quality management system, attached photos from a separate camera, and filed the record. Documentation ran a week behind, customers complained, and by the time the complaint climbed to the VP it sounded like slow inspection.
The two problems had very different first slices. Automating inspection meant computer vision, approval against AS9100 quality standards, and a multi-quarter rollout with poor odds. Compressing documentation meant a voice-to-structured-record pipeline, a photo workflow, and an export into the existing quality system: gravel-road work with a three-week first slice. The metric was documentation hours per inspector-shift. Baseline overtime was four hours per inspector per week, and the target was under one hour with records still audit-ready.
The blocker was the quality manager, whose career rested on AS9100 audit readiness and who had been burned by an earlier deployment that traded compliance for speed. Showing her the audit trail in week one, including how every voice transcript was archived next to its structured record, turned a near-veto into a co-designer. The line user was a senior inspector with twenty-three years at the company who openly doubted "the iPad people" could build anything that respected the work. In week three she ran the first slice on her own shift, and her overtime dropped to forty minutes.
Finding it took five interviews, two shop-floor observation sessions, and one ride-along. The customer's sales contact had described a different project entirely.
Nobody had written the real problem down. Someone had to go and watch it.
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.