← Selected work
Consumer AI0 to 1Regulated product

Case study 01

Making credit data understandable, not just available.

A customer-led AI explanation experience designed to answer a deceptively simple question: what happened to my score, and why?

Company
Capital One, CreditWise
Role
Product Manager
Focus
Discovery, strategy, experience, alignment
Stage
Production experiment

The core insight

The information existed. Understanding did not.

Credit monitoring products can show a score, a factor, an alert, and an educational article while still leaving the customer unable to connect any of it to their own life.

BehaviorContextGrounded data
✦GENERATIVE AI INSIGHTTurn scattered signals into clear, personal context.

Grounded in the information already available.

01 / Problem

A mature product with an information-clarity problem.

Organic reviews and complaint channels revealed a pattern: customers questioned accuracy, struggled to connect score changes to real financial behavior, and wanted guidance rather than another static explanation.

The product had information, but the cognitive burden sat with the customer. They had to hold personal credit details in mind, find generic educational content, and make the connection on their own.

02 / Strategy

Use AI where rules-based generic content stopped scaling.

Today

Monitor

Report the data.

Phase 1

Explain

Make it personal.

Phase 2

Interpret

Find the pattern.

Future

Guide

Help decide what is next.

The first release was intentionally narrow. Its job was to establish the infrastructure, governance, and customer signal needed to earn the right to build the more ambitious product.

03 / Product decisions

Make the experiment useful enough to matter and bounded enough to ship.

01

Put the answer at the moment of confusion

Place the explanation beside the score change, not behind another page or inside a separate chatbot that requires the customer to know what to ask.

02

Ground before generating

Use existing credit data as the source for the first version. Let the model translate and synthesize before asking it to create new financial conclusions.

03

Generate only when the context changes

Refresh the summary when the underlying score changes. This protects latency and cost without creating stale explanations.

04

Instrument trust directly

Add a one-click helpful or not-helpful signal to capture customer response to the explanation itself, not only downstream engagement.

04 / Ownership

The model was only one part of the product.

I originated the customer problem, formed the hypothesis, prototyped the experience, defined the experiment, and led the work from concept toward a controlled production release.

The harder product challenge was organizational and technical: align engineering, machine learning, design, legal, risk, compliance, vendor management, and senior leadership around a customer-facing use of generative AI.

Discovery

Turned unstructured customer feedback into a focused problem statement.

Experience

Built an initial prototype to make the product tangible and accelerate decisions.

Quality

Defined pass/fail checks, quality dimensions, and formatting criteria for output evaluation.

Alignment

Created a path through competing customer, business, technical, and control priorities.

What this work demonstrates

Customer empathy is most valuable when it improves the product.

Finding a meaningful signal in messy feedback.

Choosing an AI use case with a reason to exist.

Making technical and experience tradeoffs together.

Moving an ambitious idea through a constrained system.

This case study reflects public-safe product reasoning. It intentionally omits internal screens, proprietary architecture, and unverified outcome claims.

Next case study

Building Telic Portraits