⌂ Index🎧 ListenGlossary Field Notes
🎧 Listen to this chapter

08 — The Assisted / Expert Operating Model

TurboTax Live, the Virtual Expert Platform, and the seasonal labor model behind them — the growth engine and (directly) Consumer Group's world.
Sources: research/SOURCE_INDEX.md — [S24][S25][S2].


What "assisted" is

The Virtual Expert Platform (VEP)

VEP is the connective platform that routes customers to the right expert (AI or human) at the right moment, and gives experts the tools/context to help. It powers TurboTax Live and QuickBooks Live and is central to Intuit's "AI-driven expert platform" strategy. (VEP was created within Consumer Group leadership — see the leadership dossier; onboarding new businesses like Credit Karma onto VEP is active work.)

VEP's technical stack (what's actually inside) [S43][S44]

Five capability layers, per Intuit's own descriptions:
1. Matching & routing — AI models do real-time matching of customer → expert, scoring on the customer's specific need vs. each expert's profile, credentials, current capacity, and past performance. This is the "matchmaker" core (the Connect team's domain).
2. Document intelligence — AI extracts, classifies, and auto-places data from customer tax documents into the correct form fields — the expert reviews instead of transcribing. (The consumer-facing "90% form automation" rides the same capability.)
3. Expert augmentation ("AI on the expert's shoulders") — during a live voice/chat session the platform feeds the expert real-time article recommendations, auto-complete suggestions, and response snippets; afterward it auto-generates call summaries (questions, actions, resolutions) and applies quality scoring / resolution-likelihood prediction. Experts also get a real-time summary of the customer's journey and prior touchpoints.
4. Virtual collaboration — screen sharing, co-browsing, document co-review; what makes a single-session Full Service prep possible.
5. Digital expertise layer — upstream of any human: AI predicts contact reasons, offers personalized self-help, resolves straightforward issues, and decides when to escalate to a human — the AI↔HI handoff logic.

The design thesis: built as configurable, reusable platform capabilities any Intuit business can adopt (hence CK onboarding = configuring these five layers for credit/money journeys, not building new tech). Underneath: ML + NLP + knowledge engineering — and post-GenOS, these capabilities increasingly ride the GenOS stack (file 11 §c2). Intuit's framing: "augmented intelligence that sits on the shoulders of its experts" — automate the admin, keep the human on judgment and reassurance.

Lineage & direction (don't get this backwards): VEP originated on the TurboTax/QuickBooks side — Notarainni created it with the assisted-tax strategy; it has only ever run tax + bookkeeping journeys. Credit Karma is being onboarded ONTO it (the active migration program): Salesforce = the member/case data foundation underneath, VEP = the experience on top, and CK's credit/money member journeys are net-new design work on existing rails.

The underlying data (verified foundations + hypothesized model) [S45][S46]

Verified: Intuit runs a lakehouse + data-mesh architecture (domain-owned data products; Spark Streaming/Flink real-time; Databricks + Redshift access). VEP's stated capabilities imply its data needs: matching requires live expert profile/capacity/performance data; the expert's "customer journey summary" implies an event/timeline store; summaries/QA imply captured transcripts. GenOS's data cognition layer is the LLM↔data grounding bridge.

Hypothesized data classes (Analysis — validate inside): (1) customer/member context (identity, product state, journey events); (2) case/CRM + entitlements from Salesforce (which SKU → what service owed); (3) expert workforce data (credentials, schedules, real-time availability, quality history — the matching engine's supply side); (4) interaction data (transcripts, summaries → QA + training); (5) document/return data (the expert's work surface; §7216's permitted core); (6) knowledge content (articles, tax-knowledge engine, snippets).

The flywheel (Analysis): events → journey timeline → matching consumes low-latency features from both sides → sessions generate interaction data → retrains matching/quality models. Every session improves the next match — a data moat competitors can't rent.

Why "Salesforce first" is technically forced (Analysis): the matcher can't route a CK member without identity, entitlement, and case history — exactly the data being consolidated into Intuit's Salesforce instance. The migration isn't adjacent to VEP onboarding; it builds VEP's supply of member-context data. And the compliance boundary runs through here: CK data + tax data mingling in one routing/context layer demands consent-aware data scoping per member and per direction (§7216 outbound / GLBA-FCRA inbound — file 06).

Validate inside: homegrown vs. vendor workforce management; formal feature store or not; real-time vs. batch journey context; where exactly the consent gates sit.

The seasonal expert labor model [S24][S25]

The assisted model runs on a flexible, largely remote, seasonal workforce:
- Who: credentialed experts — CPA, EA, or practicing attorney — with an active PTIN and typically 2+ seasons / 30+ returns of experience.
- How they work: 100% remote, minimum ~20 hrs/week across 3+ days, scheduled around operating hours on a first-come basis during the season.
- Pay: hourly (rate varies by state) plus a performance-based season-end bonus.
- Training: up to ~17 days (paid, remote, self-led + live).
This lets Intuit scale a human workforce up for the ~10-week peak and down afterward — matching the seasonal demand curve (file 03) without carrying year-round headcount.

Why assisted matters strategically (Analysis)

  1. ARPU + growth — assisted customers are worth multiples of DIY (file 05), and the segment is growing fastest.
  2. Retention & trust — a good expert experience deepens loyalty in a trust-driven category.
  3. The AI moat — the human-in-the-loop is exactly what general-purpose LLMs don't have. As AI commoditizes DIY guidance (file 11), "AI plus a real, accountable expert who stands behind the return" is a defensible position an LLM alone can't match.
  4. Operational complexity — it turns a software company into a workforce-operations company for a quarter each year: recruiting, credentialing, training, scheduling, and quality-managing thousands of experts against a hard peak. That's a real program-management surface.

The AI angle on the expert model

Intuit's direction is AI-augmented experts: AI handles routine lookups/drafting and triage so human experts focus on judgment and reassurance — raising expert productivity while keeping the human accountable. This is the practical shape of "expert-in-the-loop AI" and the bridge between files 08 and 11.

Will agents become the experts? (Analysis, 2026-07-18)

The sharp version: agents will do the expert's work; humans keep the expert's accountability — and the ratio quietly inverts.

← 07-how-tax-software-works09-the-intuit-ecosystem-and-cr →