Designing proactive help for TurboTax Live

Project: Proactive Interventions – TurboTax Live
Role: Principal Designer, Intuit
Platform: TurboTax Online (Mobile Web)

The Problem With Reactive Help

Most help systems wait. A user gets stuck, hunts for a help button, and enters a support flow that was designed as much around the company’s convenience as the customer’s. By the time help arrives, the frustration is already baked in.

TurboTax Live promised something better: real tax experts, available when customers need them. But even with that resource in place, customers still had to recognize they needed help — and then choose to ask for it. The system was reactive by design, and that gap between available help and used help was costing customers and the business.

The question: what if the product could detect the moment a customer needed help and offer it before they had to ask?

Two Problems, One Solution

Making proactive help work required solving two challenges in parallel. Get either one wrong and a disruptive pattern like this doesn’t just fail — it actively makes things worse, adding noise and friction to an already stressful experience.

• Problem 1: Knowing when a customer needs help
• Problem 2: Delivering help in a way that drives customer connections

Both had to be right.

Part One: Listening and Detection

The detection model draws on two layers of signal.

Page and scenario context. Where is the customer in the filing flow? Some parts of a tax return are objectively harder than others, and that context shapes the baseline likelihood that help is needed at any given moment.

Behavioral cues. On top of that, the model watches for real-time UI signals: scrolling up and down repeatedly, clicking and unclicking answers, panning around the screen, extended time-on-question. In combination, these paint a picture of a customer who isn’t moving forward with confidence.

Progression and smart pacing. The model also factors in where a customer is in their overall session and how recently help was already surfaced — keeping the experience from over-triggering and becoming its own source of friction. The decision to act isn’t just about detecting struggle; it’s about reading the full context to choose the right moment.

Part Two: The Help Experience

Knowing when a customer needs help was only half the challenge. Research surfaced a clear set of biases that prevent customers from engaging with human help even when they need it:

• “Is there actually a human on the other end?” — Skepticism that the contact will lead to a real person
• “Are they qualified?” — Doubt about the expert’s credentials
• “This will take too long” — Anticipating having to explain everything over and over
• “My question probably isn’t worth it” — Uncertainty about whether the issue warrants asking
• “Will they understand my situation?” — Worry that the help won’t translate

The intervention design was built to overcome each of these directly:

• Inline, contextual placement. The prompt appears on the page where the customer is stuck, worded to their specific situation — immediately signaling this isn’t a generic pop-up.
• A real, slightly imperfect expert image. Authentic over polished. Small imperfections read as human, directly addressing the “is anyone actually there?” doubt.
• Ratings and social proof. The expert’s rating is visible, answering the expertise question before it becomes a barrier.
• Chat as the contact channel. Low-barrier, low-commitment. Customers can engage without signing up for a full conversation.
• An animated, chat-style appearance. The intervention arrives like an incoming message — familiar, attention-getting, and reinforcing the idea that a real person initiated it.

The Impact

The MVP results validated the core hypothesis: personalized, contextually-triggered help dramatically outperformed static alternatives — both in customer engagement and downstream revenue impact. The targeting model’s ability to improve coverage without over-triggering proved to be the critical balance, and the business case for continued investment was clear.

The specifics stay internal, but the signal was strong enough to drive meaningful annualized revenue impact and shape how TurboTax thinks about proactive help going forward.

What This Work Is Really About

Proactive Interventions is a systems-level design problem that required thinking across detection logic, behavioral modeling, research insight, and interaction design simultaneously. The bias research was particularly clarifying — it reframed the challenge from “how do we surface help” to “how do we overcome the specific reasons customers won’t accept help even when they need it.”

That shift in framing is what made the design decisions coherent rather than decorative. And getting it right meant the difference between a feature that builds customer confidence and one that erodes it.