AI Architecture · Multi-tenant · Staff IC
Designing Multi-Tenant AI for Professional Accountants
I redefined how Intuit's GenAI behaves for professional accountants — moving it from a generic single-user chatbot to a context-aware, multi-tenant intelligence engine. By advocating for the accountant persona and defining three core AI architectural principles, I enabled cross-client workflows that reduced manual client health checks from hours of navigation to seconds of natural language query.
My contribution
- Identified the "identity crisis" gap — that the existing AI was providing amateur-level responses to expert professionals — and built the case to the business to fix it
- Defined three foundational AI architectural principles (Sovereign Context, Scalable Authority, Permission Persistence) that now govern how Intuit AI behaves across multi-client contexts
- Designed the three-level delivery roadmap, navigating real technical constraints while continuously pushing for the multi-tenant future state
- Invented the "@" mention interaction — the hero feature that enabled true cross-client AI queries from the IAS dashboard
The challenge: an AI built for the wrong person
Intuit's GenAI framework was originally built for Small Business Owners — people who benefit from basic guidance like tax term explanations and simple categorization tips. When professional accountants, who manage hundreds of clients simultaneously, used this same AI, the results were jarring.
The AI couldn't distinguish between the accountant and their client. It gave CPAs advice meant for beginners. It had no concept of "I'm working across 200 client files right now." And it forced accountants to manually hop into each individual QuickBooks file — one at a time — to perform any real professional work.
I identified this gap, named it the "Identity Crisis," and argued to the business that we were actively alienating our most powerful users.
The framework: three architectural principles
Rather than designing surface-level UI fixes, I defined three foundational principles that would govern AI system behavior at the architecture level.
1. Sovereign Context — thread isolation
"Each context has its own conversation space."
Accountants frequently share their screen with clients. If the accountant's IAS conversations appeared in the QBO chat history, clients would see information never meant for them. By keeping IAS and QBO chat threads completely separate, accountants could use both spaces freely.
2. Scalable Authority — entry-point scope
"Access scales based on location."
In the IAS Dashboard, the AI has "Universal" authority — it can look across the entire client portfolio. Inside an individual QBO client file, that scope automatically narrows to single-client only.
3. Permission Persistence — RBAC sync
"Omni is an interface, not a loophole."
The AI strictly inherits each user's existing Role-Based Access Controls. It cannot see or act beyond what the human user is already authorized to do.
The delivery: three levels of execution
Level 1 — FoundationMaking the AI recognize who it's talking to
I worked with the team to establish the accountant as a distinct persona so the AI would stop giving CPAs beginner-level advice. This required both a UX intervention and a data model change.
Level 2 — Iteration 1: Client dropdown in the input fieldShipping useful functionality while tech was developing
Engineering was actively building multi-tenant infrastructure, but it wouldn't be ready for months. I designed a pragmatic first iteration: a client dropdown selector built into the input field. Once selected, the client name converted to static text — making the scope explicit and immutable for that conversation.
Design
Intuit Intelligence — Embedded in IAS Dashboard
The AI panel sits alongside the full Accountant Suite — giving accountants instant access to intelligence without breaking their workflow context. The "My practice" scope chip anchors the query to the right data environment.
Design
Intuit Intelligence — Client context dropdown
The "My practice" chip expands to a searchable client list. Selecting a client scopes the AI query to that file — a pragmatic shipping path while multi-tenant infrastructure was being built.
Level 3 — Iteration 2: The "@" mention — the actual solveDesigning for how professionals actually think and work
The dropdown worked, but it felt clunky. I challenged the "one context at a time" limitation by designing the "@" mention feature — letting accountants reference specific clients inline within their query, just like tagging someone in Slack.
From the IAS dashboard, an accountant types "@Client A @Client B @Client C — compare Q3 revenue trends." The AI respects each client's permission boundary, synthesizes cross-client data, and returns an insight in seconds — a task that previously took 45–90 minutes of manual file navigation.
Design
Intuit Intelligence — @ mention client picker
Typing "@" triggers a client picker dropdown. The accountant selects clients inline — each becoming a scoped reference in the query — enabling multi-client synthesis without leaving the IAS dashboard.
Impact
The three architectural principles I defined became the governing framework for how Intuit AI behaves for all professional users — a reusable foundation that any future AI feature in the accountant ecosystem builds on.
Cross-client workflows that previously required hours of manual file navigation were reduced to seconds of natural language query.
What I learned
The hardest part of this project wasn't the design — it was the advocacy. Convincing a business that its most powerful AI feature was alienating its most valuable users required framing the problem in business terms: these accountants manage the majority of data in the Intuit ecosystem, and if they don't trust the AI, adoption fails.
