Intuit · Expert Co-Pilot · 2024–Present

Atlas AI-Powered Workspace

The evolution of Expert Co-Pilot — from AI assistance to AI agency.

Role

Lead Product Designer

Team

Innovation, Engineering, Data Science

Timeline

2024–Present

Platform

Native App · Internal AI Platform

Experts are operators, not orchestrators

Today's experts drive every action across disconnected tools — IEP, Slack, Outlook, internal systems — while AI helps task-by-task. The ceiling is the expert's bandwidth, not the AI's capability.

  • Context Switching Experts juggle IEP, Slack, Outlook, and internal tools. Each switch costs time and focus.
  • Fragmented Context Customer information lives in silos. Experts manually stitch together the full picture.
  • AI as Assistant Only Current AI helps with single tasks but can't execute across systems — experts remain the bottleneck.

Give AI agents hands

Atlas is a unified workspace where agents do the work and experts manage — built as a native app that connects to dependencies, powered by a framework any team can extend.

01

New Mental Model

Agents do the work. Experts manage. The shift from operator to orchestrator.

02

Native App Architecture

Connect to dependencies. Don't build within them. Ship fast with full control.

03

Framework-Driven SDLC

Load an expertise file. Atlas does the rest. Any team can extend — no sprints needed.

One workspace, every system

Atlas sits at the center, connecting to integrated products via API. Experts give intent — Atlas executes across systems. A Framework Layer underneath lets teams load .md expertise files to extend Atlas to new domains without feature builds.

Integrations

Slack Outlook Calendar IEP Internal Tools
Atlas

Framework Layer

tax_pro.md bookkeeper.md csm.md payroll.md + any team

One command, full execution

Tax Pro

"Follow up with Sarah about her missing 1099"

Atlas drafts the email, schedules a 15-min call in Outlook, and logs the action in IEP — one command.

Bookkeeper

"Prep me for the Henderson review"

Atlas pulls customer context from IEP, surfaces recent Slack threads, and assembles a summary brief — expert reviews and goes.

CSM

"Reschedule my 2pm and notify the client"

Atlas moves the calendar event, drafts a Slack DM to the internal team, and sends a client email — all confirmed with one approval.

New Team

"Payroll team loads their expertise file"

Atlas immediately supports payroll workflows — no feature requests, no sprint cycles. Author expertise, not code.

Before & After

Today With Atlas
Experts drive every action manually Agents execute, experts approve
AI helps with single tasks in isolation AI works across systems in one flow
Context lost between tools Continuous context across every tool
Each team builds features from scratch Teams author expertise files, not features
SDLC bottleneck: build → test → deploy per feature New SDLC: write .md → load → validate

Workshopping a new AI-augmented PDLC

Atlas didn't start as a product spec — it started as a question: what if the product development lifecycle itself was redesigned around AI? I structured a series of cross-functional workshops with PM, PD, and design partners to prototype the workflow live, with the Atlas concept emerging from the practice.

Phase 01 · 3-Day Workshop

Define the workflow

PM, XD, and PD co-designed a new AI-augmented process — from Discover → Diverge → Build → Cycle. Shipped three decks: Vision, Exec Pitch, and Workshop Playbook. Established a two-tier context model (shared templates + private agent expertise) and a push → review → learn feedback loop.

Phase 02 · Week 1 Planning

Run it live

Put the workflow into practice on a real initiative. Parallel agent exploration became a team learning exercise — not me teaching, but all three disciplines discovering AI capabilities alongside each other. FigJam board with color-coded activities kept the session legible.

Phase 03 · This Week

Validate Steps 6–9

1.5-day workshop testing the build half of the flow. PM plays Orchestrator (prototype → Jira stories with AI assist). PD plays Executor (builds from stories, Slacks when blocked). I respond to design pings and push UI PRs.

Key Finding · The Pivot

The PM writing requirements first was the bottleneck.

Ideas were too exploratory for detailed specs. So the team pivoted: PM writes one paragraph, XD prototypes immediately, prototype becomes the living artifact.

From that pivot, the AI Orchestrator + Executor pattern emerged — far more ambitious than the manual tag-team we'd originally imagined. A working Expert Copilot prototype and a refined 9-step process flow came out of a single session.

The 9-step flow

Step 01 Paragraph brief PM
Step 02 Prototype immediately XD
Step 03 Parallel agent exploration Team
Step 04 Review & align on prototype PM · XD · PD
Step 05 Refine living artifact XD
Step 06 Orchestrator breaks down work PM + AI
Step 07 Jira stories from prototype PM + AI
Step 08 Executor builds from stories PD + AI
Step 09 Design pings → UI PRs XD
Discover / Diverge Orchestrator phase Executor phase

What's in scope

In Scope for Exploration
  • Native app prototype with core integration points
  • Framework definition and .md expertise file spec
  • Cross-product task execution model (Slack, Outlook, IEP)
  • Expert testing with tax pros, bookkeepers, and CSMs
Not Yet
  • Production-grade infrastructure decisions
  • Full BU-specific rollout plans

Proposed Roadmap

01

Align on vision & principles

This presentation

02

Define framework spec

Expertise file structure + principles

03

Build native app prototype

Core integrations + 1 expert type

04

Expert validation sessions

Test with real workflows

05

Expand & iterate

Load additional expertise files

Back to all work