Case study - Health ops · Search & data mapping

A two-week investigation, answered on the call

When a member asks "is this procedure covered?", the answer starts with a five-digit CPT code - and nobody could get from that code to a coverage answer without leaving the tool they were in. I designed a CPT search, a code landing page, and the admin layer that keeps the mapping honest, replacing a four-tool manual lookup that generated over half a million tasks a year.

My role
Product designer - research, IA, UI, phasing
Team
Product, engineering, content, claims & benefits SMEs
Timeline
Analysis mid-2025 · Phase 1 delivery Q4 2025 - Q1 2026, Phase 2 scoped
Tools
Figma, Confluence, data warehouse, Claude for first-pass mapping

Overview

Context · Problem · Constraints

01

Context

Half a million tasks a year, resting on a number the tools didn't recognise.

The claims & benefits team fields member calls and processes roughly 500,000 tasks a year. A large share of them exist for one reason: the CPT or HCPCS code on a claim - the standard code describing a procedure - doesn't map to anything a health assistant can see while the member is on the phone.

Neither the health-assistant platform nor the internal benefits admin platform had any concept of a CPT code. It was just a number sitting on a claim. Connecting it to "what does this member's plan actually cover" was entirely human work.

The data already existed. The entry point didn't.

02

The problem

No path from a CPT code to a coverage answer inside the tools people already use - so a lookup became a two-week task.

Health assistants left the platform mid-call to cross-reference external sources. Whatever they couldn't resolve on contact became a task for the ops team. Members waited.

7–14 daysto process a claims or benefits task - for what is, in principle, a lookup
542,027claims tasks - the largest actionable category, 12.34% of all tasks
4 toolsconsulted to answer one question, none of them ours
MembersCall about a specific procedure and, if it can't be answered live, wait out a follow-up task.
Health assistantsLeave the platform mid-conversation to piece together an answer from external sources.
Ops & claims teamsInherit every unresolved call, plus claims work a straightforward code-to-benefit lookup could settle.

03

Constraints

DomainCPT codes aren't one-to-one with benefits. The same code can map to several benefits depending on context, so the model had to be many-to-many from the start.
TrustAn AI-generated mapping is a reasonable first draft, never the final word. A wrong mapping misinforms a coverage conversation - SME review had to be part of the tool, not a side process.
Scale vs. timeThe CPT universe is enormous and one quarter is not. Mapping everything wasn't realistic, which forced an explicit prioritisation call.
No new toolAnything built had to live inside the platforms people already use. A separate CPT lookup tool would just be one more tab - which is the problem, not the fix.

Discovery

Research · Synthesis

04

Research

18 months of task data, 40 million claims, and an afternoon watching the actual process.

Where the volume actually sits

Starting from a warehouse pull - task data from January 2024 to June 2025, and every 2025 claim. Setting aside generic follow-ups, which are 42.4% of tasks and no signal on their own, claims tasks were the largest single category at 12.34%: bigger than outreach, rising-risk care or case management.

Four steps to answer one question

Shadowing the claims & benefits team through a real CPT-related task exposed the chain behind every answer.

  1. 01Look the code up in an external CPT reference tool
  2. 02Search the web for plain-language context
  3. 03Cross-check the relevant medical policy document
  4. 04Translate all of it into an answer about this member's benefit
Process walkthrough - task to answer, four tools deep

The finding that made the project possible

Across roughly 40 million claims, the 1,500 most common CPT codes accounted for about 80% of all volume. The whole code set never needed mapping - a focused map of the highest-frequency codes covers the large majority of real member questions. I built a proof-of-concept mapping of the top 1,000 codes against the current benefit set, using Claude for the first pass, for a human to verify before anything shipped.

1,500 codescover~80% of 40M claims

05

Synthesis

Two opportunities, two different kinds of answer.

Claims tasks exist because nothing connects a claim to its benefit

Make that connection automatically and a large share of claims tasks never need to be created - scoped at up to a 40% reduction.

Benefits tasks exist because a code can't be looked up in place

Making CPT codes searchable, with results scoped to the member's plan, was scoped at up to 50%. The ceiling is lower on purpose: some questions genuinely need a human to read medical policy, and that step doesn't disappear.

Guiding principle

Don't build new intelligence - remove the need to leave the tool.

Phase 1Make CPT codes searchable and visible, and give ops the tools to maintain the mapping.
Phase 2Connect an incoming claim to its most relevant benefit automatically - once Phase 1 validates the mapping.

Design

Exploration · Iteration · Validation

06

Directions considered

Rejected
Fully automate the mapping, skip human reviewAn AI-only mapping without SME sign-off is exactly what erodes trust the first time it's wrong on something coverage-related.
Rejected
Build a dedicated CPT lookup toolA branded version of the same context-switch. It doesn't remove a tab, it adds one.
Chosen
Embed CPT search and code pages into the platforms in useBuilt on the coverage and benefit data that already existed. This addressed the root problem: the data was fine, the entry point was missing.

07

Iteration & collaboration

Designed with engineering, ops SMEs and content team.

The mapping model changed shape mid-design

The starting assumption was close to one code, one benefit. Research showed codes routinely serve several benefits depending on diagnosis context - so the schema and the UI both moved to many-to-many: a code page listing many services, a service page listing many codes, rather than forcing a false relationship. Getting that structural call right early is what made both the search results and the admin tool coherent later.

Review designed in, not bolted on

Because the first-pass mapping is machine-generated, the admin surface treats verification as a primary job: ops SMEs and content review and refine definitions, usage guidance, and service links before anything becomes member-facing.

Who owned what

Engineering owned the schema changes needed to store the mapping. Ops SMEs and content wrote and validated the plain-language definitions. Product sequenced the phases. I designed and connected the three surfaces so none of them required leaving the platform.

08

Validation

Against real volume, not a small sample.

The prioritisation was checked directly against roughly 40 million 2025 claims and confirmed to cover about 80% of volume. A working proof-of-concept map of the 1,000 most common codes was built and spot-checked before Phase 1 was scoped.

Ongoing validation is structural rather than a one-off study: every machine-suggested mapping passes SME and content review inside the admin tool before a health assistant or a member ever sees it.

Solution

A search box, two pages, one admin tool

09

Search, where people already are

Typing a code, a keyword or a synonym into the existing platform search returns everything relevant in one list - the code's own overview page, the healthcare services that use it, virtual categories, and benefit programs. No guessing which system holds the answer.

Global search - code overview, services, categories, programs

10

The code page - coverage at a glance

The highest-value screen in the project.

Searching a code lands on a page listing every related service as a card, each taggedcoverage available, restrictions apply, or we are missing information - alongside the code's plain-language definition, clinical usage notes, and links out for anything needing a deeper check. It turns "is this covered" from an investigation into a glance.

  • Coverage available
  • Restrictions apply
  • We are missing information
Code page - 56 related-service tiles and status badges

11

The same link, read backwards

Every service detail page now shows its related CPT codes inline, with a count and a link through to the full list - each code in plain language. The loop closes from either direction.

Service page - code count, and the full code list behind it

12

The admin layer that keeps it true

None of this works without someone maintaining it.

A new codes library in the benefits admin platform: a searchable, filterable table of every mapped code, its type, and how many services it's assigned to - plus a per-code editor for definition, description, usage guidance and service links. This is where the machine-generated first pass gets verified before it reaches anyone.

Codes library - every mapped code and its service count
Code detail - editing a code and the services it links to

Impact

Outcomes · Metrics · Reflection

Outcomes & impact

Phase 1 was in delivery and Phase 2 scoped but not started, so these are the numbers the work was scoped against rather than measured in-market results.

Up to 40% fewer claims tasksProjected once Phase 2 links an incoming claim to its benefit automatically.
Up to 50% fewer benefits tasksProjected from Phase 1's search alone, with the remainder still needing a human medical-policy check.
A tractable scope from a huge problemThe 80/20 finding turned "map every CPT code" into 1,500 codes and a shippable quarter.
No new tool, no new workflowThe answer landed inside the platforms people already had open.

14

How we'd know it worked

Chosen to survive changing customer and call volume.

Contact ratio per benefits taskPrimary metric - normalises for a changing number of supported customers and seasonal call volume, unlike a raw count.
First call resolutionSecondary - the closest proxy for "answered while the member was still on the phone".
Tasks completed on timeSecondary - catches whether the remaining tasks got easier, not just fewer.
Raw task countExplicitly not recommended as a primary metric - it moves with customer count whether or not the tool is working.

15

Reflection

It was a wayfinding problem, not a data problemThe two-week wait wasn't caused by missing information or missing intelligence. The fix was making an existing search box recognise a code, and giving that code somewhere useful to land. A good reminder not to reach for a bigger technical solution than the problem calls for.
Model the real relationship, not the convenient oneDesigning for many-to-many, context-dependent mapping instead of the one-to-one we assumed mattered more than any interface polish.
Let the model draft, design the reviewUsing AI for the first-pass mapping across high-volume codes validated a pattern worth reusing: machine does the bulk draft, and the review step is a first-class part of the tool - especially where being wrong has consequences for a member.
Volume data is a prioritisation toolOne query against real claims turned an impossible scope into an obvious one. Worth doing before any screen exists.

16

What's next

01Phase 2 - connect an incoming claim straight to its most relevant benefit, using diagnosis context to disambiguate when a code maps to several.
02Expand the mapped set beyond the initial 1,500 as Phase 1 usage shows which codes actually come up in member calls.
03Revisit the deprioritised raw task count once there's enough in-market data to see whether it tells a consistent story alongside contact ratio.