TPA Agent —
autonomous tax gathering

Company Intuit
Segment TurboTax / Futures
Role Strategic Design + Research
Discipline Agentic AI Design
UX Research
Vision Design
Strategy
TPA Agent — autonomous tax gathering

Tax prep has a problem

Users hunt for documents, hand over credentials to inconsistent API flows, or photograph paper forms on their kitchen table. Fragmented, manual, and repeated every single year.

What if an AI agent could just handle it?

The agent does the work. The user stays in control.

Two capabilities that together could save a TurboTax Live Full Service customer 40 minutes of effort:

Identification via TurboScan — an email scanner that surfaces W-2s, 1099s, and tax signals without the user lifting a finger. No questionnaire. The agent reads the room.

Acquisition via Intuit Autopilot — a browser extension that navigates to financial institutions and retrieves documents autonomously. No credentials passed manually. The agent does the fetch.

A live task list shows exactly what's happening. Conflicts surface as confirmations. When the agent discovers something significant — a new home purchase — the experience pauses to acknowledge it personally.

The question nobody wanted to ask:
will users actually want this?

The technology worked. The use case was real. The 40-minute saving was measurable.

Email access and autonomous browsing aren't neutral capabilities. I partnered with the research team to run qualitative studies pressure-testing the vision — not to validate it, but to find out where it broke.

For most users, the risk-to-value ratio wasn't there. The time saved didn't offset the vulnerability of handing an AI access to their email and financial accounts. Trust hadn't been established — and without it, even a frictionless experience felt intrusive.

But a distinct cohort responded differently. Younger users and early adopters were genuinely excited — not despite the autonomous technology, but because of it. Two groups. Opposite reactions. The same feature. User readiness for agentic AI isn't a binary — it's a spectrum, and it's moving.

The decision not to ship

The research led to a recommendation to postpone the in-product experiment based on mixed signals from users compared to the investment needed to ship in product.

The early adopters weren't an edge case — they were a leading indicator. As agentic AI normalizes and the generation raised on broad app permissions becomes the primary tax-filing demographic, the risk-to-value calculation shifts. I wonder if someday, TPA Agent (or a form of it) will become reality.

More projects
Let's get acquainted — build something fun · thai.thao@outlook.com