Ambiguity and escalation in AI conversations
2026
Ambiguous questions force AI to guess, and wrong guesses lose trust fast. I built a framework that teaches the assistant when to ask back, and when to hand off to a human expert. Five teams across Intuit have adopted it.
Solo Product Designer partnering with 3 Backend, 6 Frontend Developers, Researchers, Content and PM
Role & team
Product area
QuickBooks mobile App on iOS & Android
Adopted by 5 teams in QuickBooks
The impact
The challenges
An assistant that answers questions can afford to be occasionally wrong. An assistant that acts on someone's books cannot.
And intent is rarely clean. "Create invoices from this file" — which rows? All three projects, or one?
In a financial product, guessing is not an option.
The process
Our first instinct was simple: when the AI isn't sure, ask. But testing showed the cost, every clarifying question added a round trip, and a cautious AI asks constantly. The assistant became a form with extra steps.
So we inverted the logic: let the AI act on its best interpretation, and reserve confirmation for when it matters. I consulted with every product team to draw that line, and it became the principle:
If it writes to the books, or reaches a customer, the AI confirms first.
The solutions
The principle plays out in two moments: how the AI presents an action for review, and when it steps aside for a human.
Before
No hierarchy, no attention, no action
AI reasoning, transaction details, and next steps collapsed into one block of text. Nothing stood out, and the action users needed to take was buried where no one would find it.
Worse: the invoice was already created before the user confirmed anything.
After
Confirm before commit
The AI translates its understanding into a reviewable card. Users see exactly what will happen and correct it line by line. Clear CTA makes confirmations a single tap.
Escalate when the AI isn't the answer
Client relationship judgment calls or product questions route to the right human, user's accountant, an Intuit expert, or support, with a conversation summary passed along so users never repeat themselves.
(In development.)
The payoff
Question to committed action in one thread - not because the AI is always right about intent, but because the design catches it when it isn't.
The impact
Five product teams adopted the framework — Invoicing, Accounting, Expenses, Reports, and the QuickBooks Agent — rendering everything from plain text to charts, tables, and mixed responses.