Conversation designer shaping how AI speaks, listens, and behaves.

The problem

During peak tax season, demand for expert assistance vastly exceeded Tax Pro capacity, creating critical bottlenecks. Users faced extended wait times exactly when they needed help most, leading to frustration, missed deadlines, and questions about the value of premium support. For the business, linear scaling of Tax Pro costs against growing user volume created an unsustainable model.

What I did

I led model UX design for a multi-agent agentic AI system that would handle high-volume tax questions while transparently managing its own limitations. I conducted extensive user interviews, log analysis, and research to understand user pain points and question patterns. Working alongside product, engineering, and domain experts, I designed the conversational responses, agent behavior patterns, guardrails, and the overall system that guided users from initial inquiry through to their final tax filing.

Key design decisions

I established clear response standards—using middle school reading level to ensure clarity on complex tax topics—and created distinct personas for each agent role. I designed decision trees and guardrails that transparently communicated when the AI was uncertain, automatically escalating complex cases to human Tax Pros rather than providing unreliable advice. The system was built as a five-agent pipeline: Interview Agent (gather info), Form Agent (identify documents), Extractor (verify data), FAQ Agent (answer questions via RAG), and Reviewer (confirm findings).

Outcome

The pilot launched during the 2024 tax season with strong results: 95% user satisfaction rate, handling thousands of tax questions and guiding users all the way through their filing journey. The system successfully reduced wait times while maintaining accuracy and user trust through transparent handling of edge cases and complexity.

Try the live demo →

The problem

When users asked Joule "What can you do?", the first version returned a single generated summary — a wall of generic capabilities with no sense of the user's context. Users couldn't tell what was actually available to them, how it applied to their journey, or how they'd benefit, which left Joule's capabilities underused.

What I did

Leading design alongside product, developers, and architects, I designed a context-aware conversational help system that reveals capabilities incrementally instead of all at once. The experience adapts to the user's journey, their device, the surrounding app where Joule was opened, and Joule's latest abilities, guiding users from a high-level introduction down to the specific things they can act on.

Key design decisions

I defined a clear message anatomy grounded in real reading constraints: a concise intro, a scannable body, and a single clear call-to-action. I set length and formatting guidelines so every response stays easy to read within the chat window, balancing completeness against the limited space users actually see.

Outcome

The framework was adopted as the model for how Joule introduces its capabilities, giving users a guided, incremental way to discover what they can do rather than a one-time generic summary.