Getting to answers faster with AI-powered conversational search and summarization

Role: Product & Design Lead (Discovery & Strategy)

Scope: AI-powered conversational search and summarization for Hayes (healthcare and clinical insights)

Timeline: 3 months

Impact: Reduced manual effort, improved information retrieval speed, and established an AI foundation across products

Stakeholders: UX Design, Research, Product Management, Engineering, GTM teams

Overview

Healthcare professionals who use Hayes spend a lot of time searching through clinical studies, medical policies, and appeals data to find the information they need. Hayes helps healthcare organizations make informed decisions by providing clinical research and policy guidance. The speed and accuracy of this work can impact treatment and policy decisions, appeals outcomes, and patient care.

I led the exploration and definition of an AI-powered solution to help users find answers about clinical trials faster, while creating a foundation that could grow over time.

The Situation

The business launched a top-down strategic initiative to introduce AI-powered capabilities into the Hayes product. Leadership wanted to bring AI to market within three months to stay ahead of competitors, strengthen the product's market position, and support future sales opportunities.

The challenge was that no one knew where AI would create the most value for users. The engineering R&D team had already developed a generative AI chat experience, and the business was eager to find a meaningful use case for it within Hayes. One hypothesis was that AI could help users quickly extract answers and insights from clinical documents through a conversational interface, but this assumption had not been validated.

Quick problem discovery

Before investing in design and development, I led a rapid discovery effort to validate where AI could have the greatest impact. Within one week, we interviewed a few Hayes users and synthesized insights from Productboard feedback and customer success teams.

The research revealed a consistent pattern: reviewing documents, finding relevant information, and extracting answers were among the most time-consuming and manual parts of the workflow. These findings validated our hypothesis and identified a clear opportunity for AI to reduce effort, accelerate decision-making, and improve the overall user experience.

A problem AI might help solve

Hayes users, including policy writers and clinicians, review large amounts of clinical research to create policies and support appeals decisions. For example, they may evaluate whether a previously denied medical procedure, treatment, or medication should be approved based on the latest clinical evidence and medical guidelines.

Their work required them to:

Search across many documents and studies
Manually pull out relevant information
Turn findings into summaries and recommendations

This created several challenges:

Slow turnaround times
High mental effort
Greater risk of missing important information

The Hayes existing search experience did not support the way the policy writers and clinicians actually worked. Instead of helping them quickly find answers and make decisions, it required them to spend too much time searching for information.

Decision making inefficiency

The core issue wasn’t just “search”—it was decision-making inefficiency.

1. Fragmented Information Retrieval

Users had to manually scan multiple documents to find relevant answers.

2. High Effort Summarization

Even after finding information, users still had to interpret and condense it themselves.

3. Accuracy & Trust Risks

Manual workflows increased the likelihood of manual error, missed or misinterpreted details.

4. Lack of Workflow Integration

Search results didn’t align with how users actually completed tasks like:

Writing policy updates
Conducting appeals review
Insurance coverage verification

Why This Was Hard

High-stakes domain (healthcare decisions require precision)
Ambiguous starting point (no clear “right” AI solution or pattern yet)
Low trust in AI outputs without clear AI transparency
Cross-functional complexity (many stakeholders including design, product management, engineering, research, GTM teams)

We were dealing with ambiguity

We weren't simply designing a feature—we were defining how AI would operate within the product and determining the boundaries of the information it could search and analyze. At the time, there were many unknowns. We didn't know what AI was truly capable of, how much of the user burden and manual labor it could streamline, or whether it could remove enough burden to create meaningful value for users.

Turning ambiguity into clarity

Aligning on the Problem

To create alignment, the product designer and I co-led a workshop with stakeholders across the business. Together, we identified the biggest challenges Hayes users faced in their analysis flow, why those challenges mattered, and the outcomes we hoped to achieve by solving them.

We also explored how user behavior would change if we were successful and how those improvements could create value for both customers and the business.

By the end of the workshop, we had aligned on:

A shared vision for the future experience
A hypothesis “If we provide AI-generated answers with clear sourcing, users can complete complex workflows faster, with less effort and greater confidence. “
User and business outcomes
Success metrics
Technical capabilities and constraints

Creating a Shared Vision

The workshop helped us move from assumptions to a shared understanding of the opportunity. Rather than focusing on AI features, we aligned around the experience we wanted to create for the users and the outcomes required to achieve it.

This vision became the foundation for every decision that followed.

Effort/Value Matrix on our design jam ideas

Exploring Solutions Together

With alignment in place, we facilitated a design jam with stakeholders, designers, product managers, and engineers.

The goal was simple: explore how AI could reduce the burden of searching through large volumes of clinical evidence and help our users make faster, more confident decisions.

Participants sketched potential solutions, and we reviewed each concept together.

Prioritizing for Impact

We evaluated ideas based on two factors:

User impact
Design and engineering effort

Several concepts emerged as clear favorites, and we combined the strongest elements into a single solution direction.

Along the way, we debated whether features such as providing prompt guidance, historical searches, saved searches, and rating search results, were necessary for the initial release.

Phased approach

To guide prioritization, we created three versions of the experience.

Phase 1 (“Roller skates”) → Core summarization + feedback
Phase 2 (“Electric scooter”) → Ratings, saved results, history
Phase 3 (“Sports car”) → Prompt guidance + deeper interaction

The Sports car
Our long-term vision included prompting guidance and a prompt builder, and deeper summary interactions.

The Electric scooter
The near term vision included ratings, saved searches and search history

The Roller skates
Our MVP focused on validating the core value proposition: helping users find trustworthy answers faster.

The MVP included:

A few suggested prompts
A prompt entry area
AI-generated responses
Source citations for verification

Key Decisions

1. Start with a Focused MVP (Not a Full AI Assistant)

We scoped Phase 1 to summarization and conversational search within a contained experience.

Why: We needed to validate usefulness before expanding scope

Tradeoff: Limited functionality, but faster learning

Choosing to Learn First: Rather than investing months building advanced functionality, we launched the roller skates version first.

This approach allowed us to validate whether AI could meaningfully reduce research time and improve confidence before investing in a more sophisticated experience.

We aligned around a clear hypothesis:

If we provide AI-generated answers with clear sourcing, users can complete complex workflows faster, with less effort and greater confidence.

2. Separate AI from Product Core Search (Initially)

We introduced AI in a dedicated surface instead of embedding it directly into existing workflows.

Why: Reduced risk and allowed controlled testing

Tradeoff: Added friction (two product searches) vs. fully integrated experience

3. Design for Iteration, Not Perfection

We structured the roadmap into progressive milestones:

Phase 1 (“Roller skates”) → Core summarization + feedback
Phase 2 (“Electric scooter”) → Ratings, saved results, history
Phase 3 (“Sports car”) → Prompt guidance + deeper interaction

Why: AI quality and trust improve through iteration

Tradeoff: Early versions feel incomplete but enable faster validation

4. Prioritize Trust Signals Early

We emphasized:

Source visibility
Feedback mechanisms

Why: Adoption depends on user confidence in AI output

How I Led the Work

I led the initiative from discovery through execution strategy:

Discovery & Alignment

Facilitated cross-functional workshops (design, product, engineering, research, GTM)
Synthesized user pain points into clear opportunity areas
Defined the initial product vision

Vision + Strategy

Translated ideas into a phased execution plan
Scoped MVP to balance speed and value
Defined milestones aligned to learning goals

Execution Support

Partnered with designer to shape the user experience
Ensured alignment across teams as we moved toward delivery

Design solution

Phase 1 delivered:

AI-powered summarization
Conversational search interface
Feedback capture for continuous improvement

The experience allowed users to:

Ask questions directly
Receive synthesized answers
Reference supporting sources

Watch it work!

Open the video above to watch me interact with our shipped AI Phase 1 solution.

Early Results & Learnings

What worked:

Users saw immediate value in reduced manual document search and analysis effort
Summarization accelerated information processing

What didn’t work:

Users were not saving time! They were double checking the AI output which required manual labor.

Where we saw friction:

Users were hesitant to fully trust AI outputs
Citations lacked precision (needed section-level linking)
Users expected AI answers at the top of results (Google-like behavior)

These insights directly informed new Phase 2 priorities.

What This Enabled

Established an AI interaction pattern across the product
Created a foundation for scalable AI capabilities
Enabled faster iteration based on real user feedback

What This Demonstrates

Leading 0→1 product exploration in ambiguous spaces
Translating AI potential into practical, testable solutions
Balancing speed, scope, and user trust
Driving alignment across cross-functional teams
Designing for learning and iteration, not just delivery

Final Takeaway

This work wasn’t just about adding AI—it was about redefining how users access and act on information.

By starting small, prioritizing trust, and iterating quickly, we turned a complex, manual workflow into a foundation for faster, more intelligent decision-making.