Tax experts know the workflow. Now they can describe it and watch an agent run it.

The challenge

Tax teams answer complex questions by pulling reports from different systems, comparing them manually and investigating discrepancies. An AI agent could take on that sequence, but tax professionals needed a way to define the work and judge whether the system had understood it.

The project began with a one-page vision: let customers and partners create, share and find agents built around the platform's data. The experience needed to help people describe a task, connect its data sources and review the resulting workflow without writing code. During execution, they needed access to the evidence, clear approval points and a way to intervene.

My contribution

I led the 0-to-1 design effort in a 10-person team, working directly with the project lead across frontend, backend and infrastructure.

From a one-page vision to a plan the team could build. The project began with a one-page vision for creating and sharing tax agents. I helped turn it into a product direction through lightning demos, a feature breakdown and value × feasibility prioritization. At sprint reviews, engineers demonstrated completed work while I presented explorations for the next sprint, bringing feedback from product leaders and subject-matter experts into the decisions.

One plan, two ways to shape it. I paired conversational AI with generative UI. As people described their task in chat, a canvas showed the steps the agent proposed. They could refine the plan through conversation or work on the canvas directly, adding and editing nodes, with both kept in sync. Tax professionals could work the way that suited them, and the system's interpretation stayed in view where they could examine and change it before committing to a run.

Designing the run from the agent's real behavior. To design what happened during execution, I needed to see how the agent actually behaved. With playground access from the AI engineer, I tested it firsthand in Azure AI Foundry. Those observations shaped how the run view presented progress, explanations and source reports. People could inspect the evidence behind a result, and cost and time estimates helped them assess a run before starting it.

Approval where reviewers already work. Explicit approval checkpoints reached reviewers in Teams, where they could approve or reject without opening the Builder.

Beyond a single agent. I explored how workflows could reach past the platform's built-in capabilities through Model Context Protocol (MCP) connectors, with selection patterns that helped people find and compare partner tools. I also explored an agent marketplace where people could discover prebuilt agents and open them in the Builder to adapt to their own work. Sharing required clear ownership: workspace permissions governed who could build, and permissions on each agent defined who could run, edit, approve or view it.

Built to hand over. I shaped the interactions around the existing component library so the team could build quickly. Before production handover, I facilitated a closing workshop to document priorities for the next phase.

Methods

• Lightning demos, feature definition and value × feasibility prioritization.
• Subject-matter-expert feedback informing design decisions; direct end-user testing was not available.
• Firsthand testing and observation of agent conversations and behavior in the Azure AI Foundry playground.
• Multi-stakeholder user-flow mapping, including the agent as an actor.
• Design explorations and AI-assisted code prototypes for discovery and alignment.
• Interaction and content design within an existing component library.