AI-Enhanced Accessibility Testing — Concept Prototype

The CTO presented a single idea: replace the manual accessibility testing workflow — where users navigate screen by screen, evaluating each individually — with an AI agent that handles the process autonomously. No specification, no scope, no defined output. The challenge was to take that idea and turn it into something tangible, presentable, and credible.

Working closely with the BA, we mapped the existing manual testing workflow end-to-end — understanding what users currently do, where the friction sits, and what an AI agent would realistically need to replace. This discovery phase defined the scenario: what the user would instruct the agent to do, how the agent would communicate its actions, and what a meaningful summary of results would look like. The output of discovery was a clear 4-screen narrative that became the backbone of the prototype.

The concept was structured as a linear 4-screen flow, designed to feel like a natural conversation between the user and an AI agent:

Screen 1 — Instruction: a chat-like interface where the user describes the testing task to the AI agent, setting scope and parameters.

Screen 2 — Proposal: the agent responds with a suggested approach and a clear call to action to initiate the process.

Screen 3 — Execution: the agent proceeds autonomously with the testing, communicating progress to the user.

Screen 4 — Summary: results are presented with identified accessibility issues, suggested solutions, and recommended next steps.

Each screen had its own content, layout and interactive elements — designed to feel like a realistic near-future product feature, not a wireframe.

The prototype was built as a standalone HTML file with CSS and vanilla JavaScript — no frameworks, no external dependencies, no developer involvement. Cursor was used to generate and iterate on the code from detailed prompts, with Figma MCP providing direct access to production design system styles and components. This AI-assisted workflow compressed what would typically require developer collaboration into a solo 3-month process — from discovery through to a conference-ready interactive prototype.

All interactions were hardcoded — the prototype followed a fixed scenario rather than generating dynamic responses. This was a deliberate constraint: the goal was to demonstrate the concept convincingly, not to build a functional AI system.

This project sits outside the typical case study format — there are no user metrics, no shipped feature, no post-launch iteration. What it demonstrates instead is a different kind of design skill: the ability to take an undefined idea, give it structure and narrative, and make it tangible enough to stand on a conference stage.

Building the prototype with Cursor and Figma MCP — without developer involvement — was not just a practical choice. It was a proof of concept in itself: that AI-assisted workflows genuinely expand what a designer can execute independently, compressing timelines and closing the gap between design intent and working product.

The manual-to-AI testing workflow concept remains theoretical. But the process that produced it is already part of how I work.