Log in←Next →
Field reading8 min

The FDE Operating Loop

By the end of this lesson you will be able to run the discover, build, deploy, measure, feed back loop and pick the right first build.

Try itThe FDE operating looplap 1 · step 1 of 5
Question

What problem, for whom, measured how?

At Brightline

Sit with Priya's intake team; pull fax and EHR timestamps. Baseline: 71 hours median.

In a typical week

Monday: review pilot metrics with the ops lead

Where FDEs get stuck

Discovery forever: timebox it, build something by the end of week 1

Step
Discover
Output
Problem statement with a number
Lap
1
Laps by week 2
2–4
Five steps, run as a loop, several times in the first two weeks. Each step has one output you can show someone, and one way FDEs get stuck there. The example is Brightline Health's referral intake; the day line shows where the step lands in a typical week.

Dana Ruiz says Brightline's referral intake is "too slow" and asks what you'll build first. Find out how slow, then build the smallest thing that tests what's most likely to fail.

 1. Discover  → what problem, for whom, measured how?
 2. Build     → smallest thing that tests the riskiest assumption
 3. Deploy    → real workflows, real users, real data
 4. Measure   → against the metric agreed in step 1
 5. Feed back → results to the customer, patterns to your product team
      └──────► back to 1 with sharper questions

The loop is short on purpose. Each lap should take days, not months, and each one should leave you with something you can show: a number, a script, a pilot, a result, an update. Step through the diagram to see what each step produces at Brightline and where FDEs most often get stuck.

1. Discover

The output is a problem statement with a number in it:

"Tier-1 support agents spend about 6 minutes per ticket finding the right help-center article. Cut that to under 2 minutes for the top 20 ticket categories, without lowering CSAT."

That statement has five parts: who (tier-1 agents), the pain (6 minutes per ticket), the target (under 2 minutes), the scope (top 20 categories) and a guardrail (CSAT doesn't drop). Missing any one of them causes a predictable problem later: no baseline means you can't show progress, no guardrail means you can "win" by making things worse elsewhere.

Ask: "Walk me through the last time this went wrong." "What happens if we do nothing?" "How would you know in a month that this worked?"

2. Build

The riskiest assumption is usually data (can we get it, is it usable?), feasibility (is Claude accurate enough?) or adoption. If it's data access, your first build is a script that pulls and profiles a week of data, not a UI. If it's model quality, it's a small eval set and a prompt. If it's adoption, sit with the users and mock the workflow on paper before you write anything.

At Brightline, nobody has ever exported faxes from the fax server, and about 15% of referrals are handwritten. Your teammate wants to start with the review-queue UI "so Dana sees progress." What do you build first?

A script that pulls a week of real faxes and profiles them: how many, which layouts, how many handwritten, how many multi-patient packets. Data access and data quality are the riskiest assumptions here. A polished UI on top of an export that doesn't exist yet would look like progress and test nothing. Show Dana the profile instead: it's real progress with a number in it.

3. Deploy

Start with a pilot group (5 users, not 500), use shadow mode (humans review suggestions first), and keep it easy to roll back. A small group means problems hurt few people and you can talk to every user. Shadow mode means a wrong extraction is caught before it reaches a patient record. A rollback plan means one bad day doesn't end the project.

4. Measure

Track the step-1 metric plus guardrails (quality, cost, latency) against a baseline: 2.1 minutes means nothing unless you know it was 6. Read 20 failures a week; averages hide them. Twenty real failures tell you more about what to fix next than any dashboard, and they're the best source of new eval cases.

5. Feed back

A weekly written update to the customer; written patterns and feature requests to your product team. Use the same day and the same four headings every week, so the sponsor learns where to look:

WEEKLY UPDATE — Brightline intake pilot — week 3
Shipped:        Extraction for the 5 most common fax layouts; review queue live for 5 staff
Metric:         Median intake 71 h → 22 h (pilot group, n=96); wrong-field rate 3.1%
Next:           Handwritten referrals; split multi-patient packets
Need from you:  EHR sandbox credentials from Sam by Friday

Then the loop restarts. Lap two's discovery is sharper because you have data: "why are 3.1% of fields wrong?" is a much better question than "what's slow?"

Anti-patterns

Each step has a typical way to stall. The diagram shows one per step; the full list is below.

Anti-patterns and their fixes
Anti-pattern Fix
Discovery forever Timebox it; build something by end of week 1
Build in a cave Demo every week, even rough
Pilot purgatory Agree expansion criteria before the pilot
Vanity metrics ("users love it") Tie every metric to the problem statement
Silent FDE Weekly written update, same day, same format

Pilot purgatory deserves special attention because it kills more engagements than bad code does. A pilot that users like but that has no agreed expansion criteria never ends: there's always one more feature to try. Before the pilot starts, write down with the sponsor what result would justify rolling out further, for example "median intake under 8 hours for four consecutive weeks with no increase in wrong-patient errors."

Week 4 of a pilot: users say they love it, but you never agreed what "success" means. The sponsor asks "should we roll out?" What do you say?

Don't answer with feelings. Propose the expansion criteria now, tied to the problem statement and its baseline ("median intake under 8 hours for four weeks, wrong-field rate under 2%"), show where the pilot stands against them, and agree a date to decide. Next time, agree the criteria before the pilot begins.

Go around the loop several times in the first two weeks. Speed beats perfection at any step: a rough answer on Wednesday that you can correct on Friday is worth more than a polished one in three weeks.

Cheat card

  1. Discovery ends in a problem statement with who, pain, baseline, target and a guardrail.
  2. Build first whatever tests the riskiest assumption: data access, model feasibility or adoption. Often it's a profiling script, not a UI.
  3. Deploy to a small pilot in shadow mode with a rollback plan, and agree expansion criteria before the pilot starts.
Problem: <who> spend <baseline> on <task>. Cut to <target> for <scope> by <date>,
         without <guardrail> getting worse.

Interview line "My first build is whatever tests the riskiest assumption. Here that's whether we can get the faxes out at all, so it's a profiling script on a week of real data, not a UI."