Redesigning the expert workspace for an agentic world — from zero to 12,000+

Every customer interaction with Intuit's 12,000+ experts runs through IEP — the platform that powers every call, every chat, every resolution. And for years, experts were fighting it. They called it the "swivel chair dance": toggling across a dozen apps, stitching context together manually while a customer waited on the other end.

They weren't struggling because they didn't know the job. They were struggling because they were fighting their software. The question wasn't how to improve IEP. It was: what does an expert workspace look like when the software understands you, instead of you having to understand the software?

This started as a GED project — a design provocation asking what a truly AI-native expert workspace could be. The mental model was a shift from a static workshop to a dynamic garage: today, IEP is a static workshop where experts spend half their time finding the right wrench. The vision was a dynamic garage — if you're there to fix a tire, the right tools appear on the workbench. The expert's focus shifts from hunting to doing.

The provocation aligned with a broader three-year platform vision, and moved quickly from thought exercise to an official roadmap. Nine months later, every US CG expert was on it.

An omni-present AI layer at the bottom of every screen — always available, never intrusive. During live calls, it detects entities in real time (a home purchase, a medical expense, a life change) and surfaces structured context without interrupting the conversation. In chat, it drafts replies. And when an expert needs to work alongside it — cross-referencing, multitasking, digging deeper — Copilot can flex into the left panel, giving it dedicated screen real estate without ever leaving the workflow.

The expert stays in the conversation. The AI handles the cognitive load.

Listen4You was an existing capability — but it only answered questions. I extended the framework to make it agentic: it could now use tools, take actions, and give experts something to act on rather than just something to read.

The shift was fundamental. A lookup that surfaces a status code is information. An agent that detects a multi-state filing risk mid-conversation, flags it, and drafts the correction — that's a co-pilot doing the work on your behalf. The expert stays focused on the customer. The system handles the cognitive overhead running in the background.

This was the difference between a reference book and a partner who's already three steps ahead.

Experts weren't just switching between apps — they were switching between entirely different interaction modes. A voice call handed off to a chat thread. An async follow-up buried in a separate queue. Each context switch meant rebuilding the picture of who the customer was and where things stood.

The Omni-Channel Inbox unified all of it. Think of it like Intercom built for tax experts: every channel — voice, chat, async — surfaces in a single prioritized workspace with consistent UX across all of them. Switching from a phone call to a follow-up message doesn't mean starting over. The context travels with the customer.

Experts pick up exactly where they left off, regardless of channel. Less reconstruction, more resolution.

CLIs are powerful because they unlock complexity that GUIs can't surface — but only if you know the commands exist. The same problem applies to AI: experts couldn't use capabilities they couldn't find mid-call.

Two patterns solved this. The / launcher gives experts a full menu of AI actions on demand — hyper-personalized to the current engagement, context-aware, and available from anywhere in the platform. And dynamic suggestion chips surface proactively: they watch the conversation in real time and float the most relevant action at the moment it's needed, without the expert having to ask.

The insight was that discoverability is capability. A feature invisible under pressure doesn't exist. Together, these two surfaces made the entire AI layer reachable — and taught experts that the system could do far more than they'd realized.

Tools in IEP today are static and fragmented — isolated apps that only serve topical information. Experts still have to stitch the picture together, figure out the action plan, and execute it themselves. The vision was different: what if every tool were an agent, each optimized for a unique end goal?

Instead of a row of wrenches, experts get a workspace that understands intent. Take the EFE Agent: it doesn't just pull federal and state filing statuses from source — it interprets them. A rejected return doesn't surface as a raw status code; it comes with a plain-language explanation of why it was rejected and a prioritized list of next steps to resolve it. The agent reads the situation and hands the expert a path forward, not a puzzle to decode. A History Agent that reconstructs the full customer record without tab-switching. A Price Check Agent, a Scheduling Agent, an Info Gathering Agent — each purpose-built to go beyond retrieval and toward resolution, each surfaced when relevant, never when it isn't.

The result shifts experts from tool operators to decision-makers. The system handles retrieval and routing. The expert handles judgment.

Nearly 20 rounds of research kept pointing to the same thing: trust was the real barrier to adoption. Not capability, not UX — trust. Experts who felt the system was a black box didn't use it. Experts who understood why it was suggesting something did.

So we built transparency into every layer. Reasoning traces that show the logic behind a suggestion — the IRS rule applied, the data source cited, the calculation run. Citations that link claims back to source. Actions that always leave the expert in control of the final call. The AI never just asserts. It shows its work.

Rebuilding the foundation of a skyscraper while the building is full of people. That's the only way to describe shipping AI-Native IEP — a platform redesign under live load, with the stakes of tax season and 12,000 experts' daily work riding on every release.

V1 · Oct 2025 — Six weeks to design and build from scratch. Daily expert testing in the Immersion Studio, iterating overnight. A war room week with engineering to translate vision into something real.

V2 · Jan 2026 — Centralized main stage, persistent panel, Intuit Assist in the Command Center, telephony status — each decision shaped by expert VOE from V1. Not additions. Refinements earned through use.

V3 · Apr 2026 — Slash Command, Biz Tax expansion with four new AI-native engagements, proactive guidance staging experiment. Every US CG expert now on AI-Native IEP.

Automation without agency erodes trust. We never wanted experts babysitting an algorithm — we needed them firmly in the driver's seat. If they don't feel in control, they won't trust the system. If they don't trust it, they won't use it. AI should be the best possible co-pilot, not a replacement for the pilot.

Research changes what you're even designing for. We started with Google Maps — turn-by-turn guidance. After 19 rounds, experts taught us something different: they already know 90% of the route. What they need is help with the hazards, the speed traps, the unexpected detours. The research didn't validate our direction — it changed it entirely.

The more advanced the technology, the more human the design needs to be. This isn't primarily a technical problem. It's a trust problem, an agency problem, a transparency problem. The systems that win aren't the most capable — they're the ones experts believe in.