Grokking the Forward Deployed Engineer Interview
Vote

0% completed

​

Finding the User and the Decision

The User

The Decision

The Cost of a Wrong Choice

Watching the Work

Three Common Mistakes

Using This in an Interview

Key Takeaways

Practice Questions

Title image: the claims manager's request is narrowed to one user, the adjuster, and one decision, which queue a claim belongs in.
Title image: the claims manager's request is narrowed to one user, the adjuster, and one decision, which queue a claim belongs in.

"We bought the model three months ago. We want AI to take over our claims."

That is how the insurer's claims manager described the project. The request is too vague to build. A claim passes through routing, assessment, payment, and sometimes a fraud check, and the request could mean any of them.

Before the FDE writes any code, the request must become two concrete things: one person whose work will change, and one choice that person makes many times a day.

The User

The user is the person whose daily work changes when the system goes live. This person is often not the one who asked for the system. The claims manager asked for the project and controlled the budget, so the claims manager was the sponsor. The adjusters did the work the system would change, so they were the users. The insurer's 4 engineers would run the system after the engagement, so they were its operators.

The sponsor knows why the project matters and what result would justify the cost. Only the users know how the work is really done. The operators know which systems the code may touch and how it will be maintained.

Asked who spends the most time on claims before assessment, the claims manager named the adjusters. About 40 adjusters each spend about 2.5 hours a day sorting new claims, which is 100 adjuster-hours a day. The project was no longer "AI for claims". It became "the work adjusters do before a claim reaches its queue".

SponsorUserOperator
In the insurer engagementThe claims managerThe adjusters who route claimsThe insurer's 4 engineers
KnowsWhy the project matters, the budgetHow the work is really doneWhich systems the code may touch
Main question for themWhat result justifies the cost?What do you look at, and what do you decide?How will you run and change this?
Risk if ignoredThe project loses fundingThe system solves the wrong problemThe system stops working after handoff

The Decision

A decision is one choice, made repeatedly, that has a fixed set of possible outcomes. A good first project automates or supports one decision, not a whole department.

Routing fits well: the same choice, about 1,200 times a day, between six fixed queues. Assessing damage and deciding a payment are also decisions, but they are made less often and carry far more risk.

The FDE described each candidate decision in the same five parts:

  • Input: what the person looks at before choosing. For routing, that is the claim form, the photos, and any scanned police report.
  • Options: the possible outcomes. For routing, these are the six queues.
  • Volume: how often the choice is made. For routing, 1,200 times a day.
  • Time per decision: about 5 minutes today.
  • Cost of a wrong choice: what happens when the choice is wrong.
The five parts of the routing decision: the input, the six options, the volume, the time per decision, and the cost of a wrong choice.
The five parts of the routing decision: the input, the six options, the volume, the time per decision, and the cost of a wrong choice.

The Cost of a Wrong Choice

The last part of the decision needs the most care, because it controls how much automation is safe. The FDE asked the adjusters: "When a claim ends up in the wrong queue, what happens next?"

A glass claim sent to minor damage loses about 1 day before an adjuster moves it. An injury claim sent to minor damage may make the insurer miss a legal deadline. A real fraud case routed to glass may be paid before anyone checks it. Three facts followed:

  • Errors do not all cost the same, so the system cannot use one rule for every queue.
  • A mistake costs least in glass and minor damage, so these are the safest queues to start with.
  • A wrong automated choice for injury or fraud costs far more than the 5 minutes an adjuster spends reading the claim.

The final system's per-queue confidence thresholds followed directly from what each kind of mistake costs.

Watching the Work

Asking tells you what people believe they do; watching tells you what they actually do. Watching two adjusters for a morning, the FDE saw them open the policy record in a second system to check when the policy began. Nobody had mentioned this step. This explained the fraud rule: a policy bought less than seven days before the accident sends the claim to possible fraud. The system would need the policy start date, which the claims form did not contain.

Three Common Mistakes

  • Treating the sponsor as the user. The claims manager did not route claims, so a design from that description would have missed the policy check.
  • Choosing a decision that is too large. "Decide whether to pay a claim" combines routing, damage assessment, and fraud review. A first version that tries all three usually finishes none.
  • Skipping the cost of a wrong choice. The demo's single error rate, 1 claim in 4, hid that some errors were cheap and some were dangerous.

Using This in an Interview

In a decomposition or customer simulation round, the interviewer often gives a request as vague as "take over our claims". Before proposing any system, name the user and the decision, and ask what a wrong choice costs.

"Before I design anything, I want to know whose work changes. Adjusters spend about 100 hours a day routing claims into six queues, so I would start with that decision. What happens when a claim reaches the wrong queue? If glass and injury differ, that changes how much we automate."

Expect two follow-up questions:

  • "Why not automate the payment decision, since that is where the money is?" Payment combines several decisions, while routing is made most often with the fewest options.
  • "What if the sponsor insists on a bigger scope?" Agree on the larger goal, then show that routing is its first step and delivers a measured result within weeks.

Key Takeaways

  • A vague request must become one user and one decision before any design starts.
  • The sponsor, the users, and the operators each know different things, and each must be asked.
  • Describe a decision by its input, options, volume, time per decision, and the cost of a wrong choice, which decides how much the system may automate.
  • Watch the work as well as asking about it, because users leave out steps that have become automatic.

The next lesson, Questions for the First Ten Minutes, lists the questions for the first conversation.

Practice Questions

1. The claims manager asked for the project. Why was the claims manager not treated as the user?

<details><summary>Show answer</summary>

Because the claims manager did not do the work that the system would change. The adjusters knew how each claim was actually routed, including the policy check that nobody had mentioned.

</details>

2. A glass claim sent to the wrong queue loses about 1 day. An injury claim sent to the wrong queue may miss a legal deadline. How should this difference affect the design?

<details><summary>Show answer</summary>

The system should require more confidence before it automates the riskier choice. This is why the final system used a separate threshold for each queue, with glass routed at a lower score than injury.

</details>

3. What did watching two adjusters show that interviews did not?

<details><summary>Show answer</summary>

That adjusters opened the policy record to check when the policy began. This explained the fraud rule for policies younger than seven days, and showed that the system needed data from a second system.

</details>

Reading Progress

0%


Vote for new content

On This Page

The User

The Decision

The Cost of a Wrong Choice

Watching the Work

Three Common Mistakes

Using This in an Interview

Key Takeaways

Practice Questions