0% completed
What a Forward Deployed Engineer Does
On This Page
What the Role Is
What the FDE Built
How the Six Weeks Were Spent
Two Kinds of Output
How the Work Is Measured
What the Role Requires
Comparing Product Engineering and Forward Deployed Engineering
Using This in an Interview
Key Takeaways
Practice Questions
A regional car insurer receives about 1,200 new claims a day. Each claim is a form, a set of photos, and sometimes a scanned police report. Adjusters route each one by hand to one of six queues: glass, minor damage, major damage, injury, theft, or possible fraud. At about 5 minutes a claim, routing takes about 100 adjuster-hours a day, and the average claim waits 2 days.
The insurer bought an AI model to do this routing, but its own demo sent 1 claim in 4 to the wrong queue, and the project stopped.
The software company then sends one engineer. Six weeks later, 70 percent of claims are routed automatically in under an hour, and routing work has fallen to about 30 adjuster-hours a day. That engineer is a forward deployed engineer.
What the Role Is
A forward deployed engineer (FDE) is a software engineer from a product company who works directly with one customer and builds whatever that customer needs for the product to deliver results. The code runs in the customer's environment and solves the customer's specific problem. Palantir created the title, and AI companies now hire FDEs because a capable product often produces no value until someone connects it to the customer's data, workflows, and people.
The FDE works for the AI company, alongside its product team and the insurer's 4 engineers, who will operate the system afterward. This fixed period of work with an agreed goal is called an engagement, and here it lasts six weeks.
What the FDE Built
The demo sent each claim to the model as plain text and asked for a queue name. The FDE spent the first week reading 300 claims with the adjusters and found three gaps. Photos decide most glass and minor damage claims, and the demo never sent them. The police report decides most theft claims, and it arrives as a scanned image. Adjusters also send claims to fraud for reasons they had never listed, like a policy bought less than seven days before the accident.
The system that went to production has five parts:
- Intake. A small service reads each claim, its photos, and any scanned report from the claims API.
- Extraction. A separate step converts scanned reports to text. The model then turns the form, photos, and report into a short list of facts: damaged parts, injuries, and whether a police report exists.
- Rules. Fixed code checks the conditions the adjusters named, like a policy younger than seven days. A known rule that must never vary belongs in code, not in a model.
- Routing with a confidence check. The model suggests a queue and a confidence score, a number that says how sure it is. Above a per-queue threshold the claim is routed automatically; otherwise an adjuster gets it with the suggestion attached.
- Decision log. Every automated decision is logged with its inputs and score, as the regulator requires.
None of these parts is unusual. The difficulty was learning which ones were needed.
How the Six Weeks Were Spent
Most FDE engagements follow this sequence:
| Week | Step | What happened |
|---|---|---|
| 1 | Discovery | Read 300 claims with the adjusters and found the photo, report, and fraud rules |
| 2 and 3 | Prototype | Ran on 2,000 past claims with known queues; routed 70 percent automatically and matched the adjusters on 97 percent of those |
| 4 and 5 | Pilot | Enabled glass and minor damage only, about half of all claims, with adjusters checking a sample |
| 6 | Production and handoff | Enabled all six queues; the insurer's engineers took over operation |
Testing on past claims gave the error rate before any real claim was routed, and the two-queue pilot delivered half of the value at low risk.
Two Kinds of Output
An FDE ships a working system to the customer and reports product gaps to the product team, since one customer's problem is often many customers' problem. This engagement produced three reports:
- The product read scanned police reports poorly, so a separate conversion step was needed.
- The regulator required a log of every automated decision, and the product had none.
- The API allowed one confidence threshold for all requests, while the insurer needed one per queue.
The product team built the last two into the product, and the next three insurance customers used them without custom code. This feedback is part of the FDE's job, not an extra task.
How the Work Is Measured
An FDE is measured by the customer's results in production, not by the amount of code written. Here, three numbers matter:
- Share of claims routed automatically: from 0 to 70 percent.
- Time to the right queue: from 2 days to under an hour for automated claims.
- Adjuster time spent on routing: from about 100 to about 30 hours a day.
At 5 minutes a claim, 1,200 claims take 100 hours, so removing 840 of them saves 70. The FDE wrote perhaps 3,000 lines of code, and the value came from choosing which 3,000 lines to write.
What the Role Requires
Every FDE job uses six skills:
- Discovery: learning what users do and what data exists.
- Scoping: choosing the smallest version that delivers value, like the two-queue pilot.
- Practical coding: writing working code quickly inside systems you did not build, like the claims API.
- Data work: turning forms, photos, and scans into facts a model or rule can use.
- Applied AI judgment: deciding which steps go to a model, to fixed rules, or to a person.
- Communication: explaining progress and tradeoffs to adjusters, managers, and engineers.
Two failures are common. Some FDEs build a large custom system only they understand, which stops working after they leave. Others act as consultants, producing slides while the customer waits for something that works.
Comparing Product Engineering and Forward Deployed Engineering
| Product software engineer | Forward deployed engineer | |
|---|---|---|
| Works for | All of the product's users | One customer at a time |
| Requirements come from | Product managers and usage data | The customer's users, directly |
| Code runs in | The company's own product | The customer's environment, connected to the product |
| Length of the work | Ongoing ownership of a product area | A fixed engagement of weeks or months |
| Measured by | Product quality and delivery | The customer's results in production |
| Also responsible for | Code quality and operation | Reporting product gaps to the product team |
Both roles need strong engineering. The difference is whose problem the engineer solves and how close they work to the people who have it.
Using This in an Interview
FDE interviewers want evidence that you can be trusted in front of a customer and that you ship. When you describe your work, give the customer's result in numbers, name your scope decisions, and say what you learned from the users.
"The customer's demo misrouted 1 claim in 4, so I spent the first week reading 300 claims with the adjusters. We piloted on the two largest queues first. After six weeks, 70 percent of claims were routed automatically, and routing work fell from about 100 to about 30 adjuster-hours a day."
Expect two follow-up questions:
- "Why not route all six queues from the start?" The two largest, simplest queues gave half of the value while the team measured real errors.
- "What would you do differently?" Name one real lesson, such as testing the scanned reports in week one.
Key Takeaways
- An FDE works directly with one customer and builds what the product needs to deliver results there.
- Most of the difficulty is discovery, learned from the people who do the work today.
- Good engagements test on past data, then pilot on part of the work before full production.
- An FDE delivers a working system and product feedback, and is measured by the customer's results.
The next lesson, Forward Deployed Engineers, Solutions Architects, and Software Engineers, compares the FDE role with the two roles it is most often confused with.
Practice Questions
1. The insurer's demo misrouted 1 claim in 4. Why would a stronger model alone not have fixed it?
<details><summary>Show answer</summary>Because the demo was missing information, not intelligence. It sent only the form text. The photos, the scanned reports, and the unwritten fraud conditions decide most claims. No model can use evidence it never receives.
</details>2. Why were the fraud conditions written as fixed rules instead of being left to the model?
<details><summary>Show answer</summary>Because the rules were already known and must never vary. A condition like "policy younger than seven days" has one correct answer. Fixed code gets it right every time and is easy to explain to the regulator. A model adds uncertainty to a decision that has none.
</details>3. An FDE spends three months building a 40,000-line custom system, and the customer's engineers cannot operate it. What went wrong?
<details><summary>Show answer</summary>The work was measured by what the FDE built, not by what the customer can keep running. A system that only its author understands stops working soon after the author leaves. The handoff should be planned from the start, as the insurer's engineers took over in week six.
</details>Reading Progress
0%
On This Page
What the Role Is
What the FDE Built
How the Six Weeks Were Spent
Two Kinds of Output
How the Work Is Measured
What the Role Requires
Comparing Product Engineering and Forward Deployed Engineering
Using This in an Interview
Key Takeaways
Practice Questions