Project ROAR: Designing the Registry of All Registries for KPMG's AI Agent Ecosystem
Overview
Designed ROAR (Registry of All Registries), a centralized platform that lets KPMG employees discover, evaluate, and submit AI agents across every line of business. ROAR acts as the firm's single source of truth for AI agents, replacing fragmented team-level lists and word-of-mouth with a searchable, AI-assisted discovery experience.
The platform includes a structured search and filter system, an AI-powered answer layer that interprets natural language queries and surfaces relevant agents with cited sources, detailed agent profile pages, and a submission flow for teams to contribute new agents to the registry.
The Problem
As KPMG's AI agent ecosystem grew rapidly, employees had no reliable way to know what agents existed, which were production-ready, or whether an agent for their specific workflow had already been built. Teams were duplicating effort, building agents that already existed elsewhere in the firm, or defaulting to generic tools because they couldn't find purpose-built ones.
The problem wasn't just discoverability. It was trust. Employees needed to know not just that an agent existed, but whether it was certified, how it had performed, what skills it offered, and how it fit into their existing workflows. A basic directory wouldn't be enough; the experience needed to feel like a knowledgeable colleague recommending the right tool.
Research
Conducted research with employees across Advisory, Audit, Tax, and IT functions to understand how they currently found and evaluated AI tools. Mapped pain points in existing discovery patterns, which largely relied on Slack messages, email forwards, and informal team knowledge.
Key findings: employees wanted to filter by their line of business and function rather than browse a flat list; trust signals like agent IQ scores, usage counts, and certification status were critical to adoption decisions; and the ability to see related agents and "works best with" pairings reflected how people actually used these tools in practice, rarely in isolation.
Design Solution
Designed the registry around three core experiences. Search and browse lets employees filter agents by Line of Business, Function, and tags like Popular, New, and Suggested, with a sort layer for Most Popular and Most Recent. Results surface agent cards with enough context to assess relevance without clicking through.
The AI answer layer sits above search results as a conversational card. When a query is ambiguous or complex, the AI surfaces a synthesized recommendation with cited sources, related agents, and expandable detail, designed so employees get an answer, not just a list.
Agent detail pages give employees everything needed to evaluate and act: IQ score, engagement count, last updated date, skills and example prompts, documentation links, a Works Best With section for workflow pairing, and an Add to Team action. The Submit Agent flow lets any team register a new agent with basic info, business classification, capabilities, and skills, making the registry self-sustaining.
Testing and Iteration
Tested the search and filter experience with employees across multiple functions to validate that the filter taxonomy matched how people actually thought about their work. Iterated on the AI answer card format to balance depth with scannability, early versions were too dense and were refined to a collapsed default with expandable detail.
Tested the agent detail page information hierarchy with users evaluating agents for the first time, confirming that IQ score and engagement count were the first things people looked for when assessing reliability. Refined the Submit Agent flow through rounds of feedback with agent owners to reduce drop-off on the capabilities section.
Outcome
ROAR established a single, trusted source for AI agent discovery across KPMG's global practice. The AI answer layer significantly reduced time-to-find for complex, multi-criteria queries. The Submit Agent flow gave individual teams a clear path to contribute to the ecosystem, shifting the registry from a curated list to a living, firm-wide resource.
The design patterns developed for ROAR, agent cards, IQ scoring displays, AI answer cards with cited sources, and the skills taxonomy, are being applied across other AI product surfaces in KPMG's enterprise portfolio.
Key Learnings
Trust signals matter more than completeness. Employees would rather see 20 well-documented, scored agents than 200 with no context. Quality of information per agent drove engagement more than the size of the registry.
AI search needs a citation layer. Users trusted the AI answer card significantly more when it showed where the recommendation came from. The Related Sources panel wasn't decorative, it was load-bearing for adoption.
Submission UX determines registry health. If submitting an agent is painful, teams won't do it. Reducing friction in the Submit Agent flow directly impacts how current and comprehensive the registry stays over time.