Visa Assist

Executive Summary

About Visa Assist

Visa Assist, Visa's first client facing Generative AI solution, enhances knowledge search and discovery within a client facing portal through a conversational UI. Designed to tackle inefficiencies in traditional search processes, this GenAI powered tool transforms a fragmented, time-consuming workflow into an intuitive, client-first experience.

Interface visuals and content in this case study have been adapted to preserve confidentiality while sharing valuable insights into the project’s success.

What was the motivation?

Since the public launch of ChatGPT, Visa, like many companies, have invested heavily into the space. While the company has announced its GenAI fund and continues to advancing its technology stack, there was an opportunity to explore GenAI B2B solutions. A small team proposed a solution for an existing client platform that connect to tools, resources, and support in an effort to improve the platform experience, acting as a launch pad for further GenAI solutions.

As a team, the goal was to make Visa Assist the primary tool for knowledge search & discovery by leveraging the capabilities of GenAI.

Team

The team was comprised of myself (design), product management, data analysts, front-end engineering, and AI Platform engineers. We partnered with our Global Design Research, Marketing, and Client teams.

My role

I was the design lead responsible for the user experience. I collaborated daily with product management and engineering on design, releases, product scope, and technical improvements of Visa Assist.

Responsibilities

Product Design: Concept Exploration, Design Strategy, User Experience, Prototyping, Usability Testing
UX Research: User Interview, Survey, Analysis, Synthesis
Product Management: Epic & Story creation, scope

My contribution

Defined Visa Assist's design principles for the user experience and conversational design. Establishing these foundations drove alignment on a user first approach across product management and engineering.
Designed new features and enhancements for Visa Assist's end-to-end experience including the landing page, prompt writing, responses, and sources.
Co-created epics & stories with product management and coordinated with engineering throughout the product development lifecycle.
Lead user research during the month long employee pilot, created a research plan including an outreach strategy, user surveys, and user interviews. I synthesized those insights and presented them to our product and engineering partners during a team workshop.
Contributed to the product roadmap by proposing new features, enhancements, and concepts based on our research and on-going design work which were presented during recurring steering committee meetings that included senior executives.

Outcome

Established foundations for conversational UI patterns adopted by other GenAI products within Visa.
Over the course of the employee pilot, daily use of Visa Assist grew steadily with 100K+ questions asked, achieving less than 1% factual errors in responses, added 13k+ documents to the library, and was used internationally.
Released Visa Assist to all internal employees across the global organization.
Visa Assist was presented during Visa Payments Forum US and EU, Visa's largest client conference.
Steering committee green lit the continuation of the product through the start of a closed client pilot.

Img 1:

Visa Assist Screen Samples

The product

RAG: Robots are generating

Visa Assist's MVP product experience was based on conversational UI patterns seen in GenAI solutions in the market. The essential UI includes an input field and text areas for prompts & responses. Above the input field were a row of buttons that acted as prompt starters. If selected, the input field would be filled and that row would reveal a new set of buttons.

A user submits a prompt to Visa Assist, the back-end analyzes, builds context, generates the response, and returns the content to the user. Visa Assist uses a RAG (Retrieval-Augmented Generation) framework and uses the content in the client platform as its library.

Each response has at least one source for reference, allowing the user to access the source material. The can copy and rate responses.

Img 2:

Redacted Visa Assist MVP UI

The Problem

Too much information, not enough time

Limitations from conventional search

Visa’s internal client platform serves as a go-to resource for businesses navigating Visa’s knowledge base. In it's existing state, clients could only use the existing search process, sifting through the tens of thousands of documents. Although some information lived in long-term documentation, other pieces of information were in documents that had a limited lifespan.

Time sensitive questions

Clients frequently turned to account executives (AEs) for answers but would be at the mercy of AE's schedules. When we interviewed our platform stakeholders, they reported that in the previous year, an estimated 100k questions were asked by clients and responses would take an average of four hours to find the answer.

Information our clients and partners need is scattered across systems, buried in documentation, and captured in the minds of our teams.

Ideation

Paving the path

At the end of the pilot, I led multiple synthesis exercises with my product partners, clustered findings, compared insights against assumptions, created personas, and identified themes.

We presented the findings to our product and engineering team during a two day workshop by framing our insights starting with our assumptions, followed by our learnings, and finishing with how might we questions.

During the second half of the workshop, I led ideation workshops with my partners, defining key moments of the product and explored enhancement opportunities. My engineering partners appreciated the opportunity to co-create ideas and we were able to leave the workshop with two concepts.

Img 4:

Related Searches concept card

Enhancement opportunities

Leveraging research insights

I took the research insights identify key moments of the product and explore enhancement opportunities. I focused on the first time user experience and response interactions.

Landing Page

In designing the landing page for Visa Assist, wanted to address a usability challenge I identified in the MVP version of the product. The original experience started users directly in a conversational UI with introductory text as the first "response."

User research showed this confused new and returning users, creating a blank canvas problem where there was no clear entry point or sense of direction. By designing the landing page, I not only introduced the product and set expectations, but also provided a flexible foundation that could accommodate future enhancements and features.

Img 5:

Sample of landing page flow

Elevating Sources

Many UXR participants told us that they appreciated that the responses from Visa Assist included sources (at the time called references). In its current form, sources lacked prominence and I explored ways to bolster its presence.

Img 6:

Sample Source card design

Rewrite Button

One of assumptions we had was that users came to Visa Assist with a sense of knowledge curiosity, we learned users instead, come to Visa Assist with a question in mind. This insight was used to explore alternatives to the prompt assistant row above the input field in the MVP version.

I designed a rewrite function that refined responses. This allowed users to choose options like adding examples, simplifying, elaborating, or more technical.

Img 7:

Rewrite feature documentation

Version 2

If I had more time

From the beginning, I noticed significant issues with response formatting in Visa Assist. Readability was poor, the design did not align with our system standards, and the unstructured formatting of responses made for a subpar user experience.

As my time with the team was nearing its end, I designed and proposed an version 2 response UI. I audited the existing product and identified specific areas of improvement. I focused on prioritizing the response hierarchy by removing the background of the response container, reducing the width, adjusted source cards, creating a dedicated type scale for responses, and mapped text components to the back-end.

Unfortunately, the challenge wasn't the front-end implementation but the back-end limitation. At the time, the GPT API implementation could not fully support the proposed formatting. Given the large scope of back-end and our lean technical team, we had to prioritize other features.

Img 8:

Sample of proposed response formatting

Reflections

Managing Expectations vs. Technical Capabilities

Throughout my time working on Visa Assist, it was clear to me that user expectation and trust for GenAI tools is very sensitive. This project further deepened my appreciation for my AI-platform partners. I had many conversations getting their feedback on the scope of my proposals. This forced me to design solutions that balanced our system's generative capabilities.

Building Trust in a Complex Ecosystem

Ensuring users trusted Visa Assist was paramount. I collaborated closely with legal and technical teams to develop ethical guidelines and maintain data privacy. This focus on transparency, from clear sources to user feedback loops, was key to building credibility.

Emerging technology

This project expanded my perspective on what it means to design for AI-powered experiences. Unlike traditional interfaces, conversational tools rely as much on user trust as they do on usability. By focusing on clarity, accessibility, and empathy, we created something that didn’t just work—it resonated.