AI-Powered Research Repository

The Problem

We had years of valuable user research scattered across Slack and Google Drive. Although documents were organized by product and file name, finding the most relevant and up-to-date insights required knowing where to look.

As a result, teams were either going directly to my UX Researcher for answers, relying on outdated findings, or making decisions without consulting user research at all. This created unnecessary work for our researcher and meant the business wasn't fully leveraging the knowledge we had already invested in gathering.

To better understand the problem, I spoke with internal stakeholders and identified three recurring pain points:

1. Findability:“I don't know where to find the most recent insight for the question I have.”
2. Idea validation:“I want to make sure my idea is grounded in user research.”
3. Understanding specific users:“I wish I could run my idea by a specific user persona.”

I saw an opportunity to use AI to turn our fragmented research archive into an accessible, searchable resource that could put user insights directly into the hands of the teams making decisions.

1. Choosing the Right Platform

At the time, the company had signed enterprise contracts with ChatGPT and Google (Claude contract was still in the works). Therefore, I evaluated two options: ChatGPT GPTs and Gemini Notebook, considering three key criteria:

1. Accuracy: Responses should be grounded in our existing research, with minimal hallucination.
2. Maintainability: The solution should be easy for my UX Researcher to keep up to date.
3. Adoption: The tool needed to fit naturally into existing workflows and be widely used across the organization.

Gemini Notebook offered a more straightforward connection to our research stored in Google Drive, making it easier to maintain and naturally grounding responses in the available source material. However, adoption across the company was low.

ChatGPT, on the other hand, was already widely used by teams across the organization. While it required more deliberate safeguards against hallucinations and manual updates to its research knowledge base, its existing adoption made it the stronger choice.

I prioritized adoption over ease of maintenance. A technically simpler solution would have little impact if people didn't use it. I chose to build a custom GPT and designed its instructions to prioritize documented research, favor the most recent findings when insights conflicted, and cite its sources so users could trace answers back to the original material.

This approach required my UX Researcher to manually update the GPT as new research became available. Given our company's strict enterprise workflow restrictions, I considered that a worthwhile trade-off for a solution with greater potential for adoption.

2. Designing for Different User Needs

Rather than creating a single GPT for the entire business, I made the decision to build two separate research GPTs: one for Financial Services and another for Insurance.

Although both product verticals served families, their users had distinct behaviors, motivations, and needs. The teams also operated independently. Separating the repositories allowed each GPT to focus on the research most relevant to its audience and made the tool more useful within existing team workflows.

I then mapped each of the three stakeholder pain points to a specific capability:

Together, these capabilities transformed our research archive from a passive collection of documents into an interactive resource that teams could consult as part of their everyday decision-making.

3. Results

The Research GPTs gave teams across the company a new way to engage with user research. Instead of tracking down documents or routing every question through our UX Researcher, stakeholders could ask questions in natural language, pressure-test ideas against existing findings, and explore concepts from the perspective of specific user personas.


Driving Adoption

Building the tool was only half the challenge. For it to have a meaningful impact, teams needed to know it existed, understand how to use it, and incorporate it into their workflows.

To drive adoption, I had our UX Researcher introduce the GPTs to each product team and present them at a company-wide All Hands. We also pinned the tools in our #user-research Slack channel and established an ongoing communication practice: whenever new research became available, we updated the repository, announced the findings in Slack, and highlighted the updates during team meetings.

This helped make the GPTs part of an ongoing research practice rather than a one-time tool launch.


Extending the Framework

The approach also proved reusable beyond user research. I applied the same framework to create a Copy GPT for Marketing and UX, helping teams develop messaging that was both user-friendly and compliant—a particularly important consideration in the fintech and insurtech industries.


Key Takeaway

This project demonstrated how I approach AI as a product and organizational challenge, not just a technology experiment. By starting with real stakeholder pain points, evaluating solutions against adoption and business constraints, and designing around the way teams actually worked, I created a repeatable framework for making specialized knowledge more accessible across the organization.