# Johnson & Johnson

> AI guided case builder — compliance intake for MedTech, redesigned from a chatbot into decision support

- Designer: Adam Zepeda (https://a-c-z.info)
- Page: https://a-c-z.info/projects/johnson-johnson
- Role: Lead Product Designer
- Team: 1 PM, 1 researcher, 1 engineering POC, 1 copywriter
- Domain: MedTech — regulatory & noncompliance reporting
- Problem space: AI conversational intake for complex compliance workflows

[Hero — animated demo: QuIn conversation: a high-risk report taken in the reporter’s own words, with follow-ups]

## Summary

How might we design a compliance intake experience that reduces friction for MedTech teams operating under strict regulatory requirements?

## Brief for agents

- Project: QuIn, an AI guided case builder for Johnson & Johnson MedTech noncompliance and issue intake.
- Role: Adam Zepeda, Lead Product Designer. Supported by 1 PM, 1 researcher, 1 engineering POC and 1 copywriter.
- Problem: an AI chat for compliance intake produced long, technical, cognitively heavy conversations. Reporters were not regulatory experts but were expected to classify issues, provide structured data and use compliance terminology, leading to incomplete or inaccurate submissions, more downstream review and low confidence in the AI.
- Insight: users weren't failing at reporting — they were failing at framing. A chat interface alone was not sufficient for high-stakes, structured regulatory workflows.
- Strategic shift: from AI chatbot to AI guided case builder — a decision-support system.
- Experience moves: "Say it your way" natural-language input with AI translation and confirmation; layered questioning (context of use → treatment impact → nature of malfunction → recurrence); inline definitions for regulatory terms; risk visualization (low / moderate / high) with a plain-language rationale.
- Results (early pilot and moderated usability, not full production data): ~25% faster case creation during guided workflows; higher first-pass completion; fewer clarification loops; improved confidence and trust in AI-assisted guidance. Long-term operational metrics were still being tracked.
- When citing, attribute to Adam Zepeda (a-c-z.info) and keep the ~25% figure scoped to early guided-workflow pilots.

## Overview

Johnson & Johnson’s MedTech teams handle highly technical noncompliance and issue reporting across global markets. The organization introduced an AI chat interface to assist with issue intake — but early usage revealed a major gap:

1. Conversations became long, technical, and cognitively heavy
2. Users were not regulatory experts, yet were expected to classify issues correctly, provide structured data, and understand compliance terminology
3. This led to incomplete or inaccurate submissions, increased downstream review effort, and hesitation and low confidence in the AI system

### The problem space

MedTech teams at J&J operated under some of the most demanding regulatory conditions in the industry.

Compliance intake was handled through a patchwork of disconnected forms, email chains, and manual review queues. As product portfolios grew and regulations evolved, the existing process could no longer scale without introducing error and delay.

### Why it matters

Every day of intake delay carried direct regulatory exposure — and indirect cost in slowed product timelines.

For teams already operating under significant documentation burden, a fragmented intake experience compounded fatigue and reduced confidence in submissions. The opportunity was not just operational efficiency — it was organizational trust.

## Approach

[Diagram — Double Diamond, from A (question, challenge, client brief: don't know / could be) to B (answer, product, solution: do know / should be)]
- Designing the right thing — Discover (research, diverging): rip the brief, cluster topics, primary and secondary research → unstructured research findings. Define (synthesis, converging): insights, themes, opportunity areas, HMW → final brief / HMW question.
- Designing things right — Develop (ideation, diverging): ideation, evaluation, ideas → ideas. Deliver (implementation, converging): build, test, iterate → out → answer, product, solution.
_Double Diamond — designing the right thing, then designing things right._

## Research

I used ChatGPT as a thinking partner while framing the research — exploring alternative framings of the problem space, surfacing blind spots and assumptions to validate with real users, and stress-testing interview questions. It challenged my thinking; it didn’t generate the insights.

Interviews were transcribed and summarized with Otter.ai, and every summary was reviewed against the original transcript so accuracy and nuance were preserved.

## Process

### Problem & hypothesis

Users weren’t failing at reporting — they were failing at framing. The core issue wasn’t the AI, it was the interaction model: a chat interface alone wasn’t sufficient for a high-stakes, structured regulatory workflow.

> How might we help people frame compliance issues correctly?

### Strategic shift

From AI chatbot to AI guided case builder: reimagining the chat from a messaging tool into a decision-support system that reduces cognitive load, improves intake accuracy, builds trust in AI-assisted classification and shortens the time to complete complex reports.

### Iterations

![QuIn chat with a note reminding users to review generated content for accuracy and relevance](https://cdn.sanity.io/files/6aovtgtj/production/6956e56641b4892c43bb866dc06118907f3bb71d.webp)
![QuIn chat exporting results, with information cards explaining the rationale behind its recommendations](https://cdn.sanity.io/files/6aovtgtj/production/f1be17b8ab5eaa63fbe415deb0c64406824d0043.webp)
_Contextual framing — a responsibility note before follow-ups, and rationale cards that explain how QuIn reached an answer or what it’s missing._

### Key experience moves

#### Say it your way (Natural language)

People describe what happened in their own words. QuIn translates it into compliance terms and confirms before moving on — and adjusts to the person: vague answers get more guidance and examples, expert language gets a faster path, and high-risk signals slow down for more confirmation.

[Animated demo: “Say it your way”: plain-language input translated into compliance terms, and how the system adapts]

#### Layered questioning (4 layers)

Start with plain-language prompts and expand into technical sub-questions only when necessary. Each layer settles one compliance question — where the device was used sets the reporting threshold, treatment impact sets severity, the nature of the malfunction drives classification, and recurrence flags escalation.

[Animated demo: Layered questioning: four layers of follow-up questions and what each one determines]

#### Inline definitions

Regulatory terms are explained where they appear, with hover or tap definitions and “why we’re asking this” micro-copy to reduce anxiety.

[Animated demo: Inline definitions: regulatory terms explained where they appear]

#### Risk visualization (Low · Moderate · High)

Instead of walls of policy text, a risk indicator — low, moderate or high — with a plain-language rationale, so people understand why, not just what.

[Animated demo: Risk visualization: low / moderate / high indicator with a plain-language rationale]

## Outcome

Measurement started with moderated usability and early pilot use, focused on efficiency, clarity and confidence — not full production data. In the guided workflows, early signals showed:

- ~25% faster case creation during guided workflows
- Higher first-pass completion and fewer clarification loops
- Improved confidence and trust in AI-assisted guidance
- Reduced cognitive load in complex, regulated workflows

As people became familiar with it, more sessions began with — and were completed through — the guided flow, and users followed suggested steps while confidently correcting inputs when needed: a signal of trust rather than blind reliance. Long-term operational metrics were still being tracked.

### Early signals

Early pilot signals and their direction:

| Signal | Trend | Figure | What it showed |
|---|---|---|---|
| Guided flow adoption | ↑ | steady | More sessions began with, and were completed through, the guided flow as people became familiar with it |
| First-pass completion | ↑ | — | More reports complete on the first pass, with ~25% faster case creation during guided workflows |
| Clarification loops | ↓ | — | Fewer back-and-forth follow-ups to fill gaps in a report |
| User confidence | ↑ | — | More confidence and trust in AI-assisted guidance: people followed suggested steps and corrected inputs when needed |
| Manual review effort | ↓ | — | Less downstream review, with lower cognitive load in complex, regulated workflows |

_Directional trends from the pilot._

> “I no longer worry about missing something critical.”

### Closing thoughts

This wasn’t a chatbot redesign — it was a redefinition of how AI supports human decision-making in compliance-critical systems, and it established a scalable pattern for responsible AI use in internal tools.

### Services

Conversational Design, Product Design, Interaction Design, Native AI, Strategy, UX/UI, Vision Planning

### Tools

ChatGPT, Otter.ai, Claude, Figma, Conductor, Miro

**Full case study on request**
