Designing a human-in-the-loop tax review agent

A focused AI review partner that surfaces likely errors at the right moment, assembles evidence, and preserves the tax expert's final say.

Role

Product designer

I led the product design for QR Agent. I translated expert research and platform constraints into a human-in-the-loop review workflow that helps tax experts find and resolve errors without giving up control.

Scope

Interaction design, prototyping, behavioral research

Team

PD · PM · UXR · 2 SWE · CS · SME

Timeline

2025: 4-6 months from research to pilot

Status

Live in production to all accounts with eligible Schedule C returns

Overview

TurboTax connects customers who want professional help with its network of tax experts. QR Agent is an internal review tool for those experts.

QR Agent moved quality review from an exhaustive, end-stage manual task into an in-flow, evidence-backed check at the moments experts already review state and federal returns.

I proposed a dedicated review hub where every action is reversible: errors ranked by model confidence, one-click navigation to the affected form, delivered in a voice that sounds like a partner working the queue with you rather than an authority handing down verdicts.

82.4%

Agent precision

Share of production errors correctly detected and flagged by the agent.

~25 minutes

saved per return

Average time saved per return with a verified issue versus manual review.

~$6M

projected savings

Projected savings if observed time reductions scale across all eligible returns.

The problem

Experts spend too much time manually reviewing mostly error-free returns.

During peak tax season, conducting manual reviews leaves less time to investigate the issues that actually require judgment.

The opportunity

Transform tax quality review from an exhaustive, end-stage manual process into an in-flow, AI-assisted experience without taking control away from tax experts.

Enter password