GenAI 'Ask Me' Experience
A grounded conversational assistant for operational questions — suggestion chips, cited sources, graceful refusals and honest limits.
Client
Enterprise Platform (confidential)
Role
UX/UI Designer — conversational UI, suggestion design, source attribution, feedback and failure states.
Platform
Web · Mobile
Confidentiality: Client identity and product specifics are confidential. Interaction patterns shown are sanitised.
01Overview
An 'Ask Me' GenAI assistant for operational questions in a financial-services product: suggestion chips to start, short grounded answers, cited sources, graceful refusals and honest failure states. A chatbot, designed like a product.
02Context
Users repeatedly asked the same operational questions. The answer existed in documentation and agent knowledge — the ask was to make it instantly available without risking hallucinated answers about money and policies. A bare chatbot was never the answer; a grounded assistant with visible limits was.
03User Journey
How a question becomes an answer
04UX Decisions
01Scope is visible from the start
A scope badge and suggestion chips tell users what the assistant can answer before they ask — setting expectations is the first line of trust.
02Short answers, then sources
The answer comes first in one or two sentences; sources expand beneath. Reading the answer must take under ten seconds.
03Citations, always
Every substantive claim carries its source. 'Because the system said so' is never enough in finance.
04Disclaimers where they belong
Where an answer touches financial decisions, a contextual note ('confirm with your advisor') appears inline — not a wall of boilerplate.
05Refusal is a designed state
Out-of-scope questions are answered honestly with an escalation path. The unhelpful path is part of the UX, not a failure of it.
05Interaction Design
States, transitions and behaviour — not just the happy path.
Typing + streaming
Answers stream with a compact marker, so users read as they arrive — no waiting on a blank screen.
Suggestion chips
Dynamic chips restate the top questions; they disappear after use to reduce noise.
Sources
Expandable citations with titles and links, collapsed by default.
Feedback
Thumbs up/down on every answer with an optional reason field.
Ambiguity
One clarifying question, then the most likely interpretation — never an interrogation.
Long generation
A progress note ('still working — this one needs more steps') instead of an infinite spinner.
06Edge Cases
Good UX is what happens when things go wrong. Each case: the problem, the user-facing message and the recovery action.
Out-of-scope question
Edge caseThe question is outside what the assistant can answer.
“I can't help with that yet — but here's who can.”
Recovery: Escalation path with a human contact; no pretend answers.
Ambiguous question
Edge caseThe intent could be two things.
“Did you mean your policy renewal or a new quote?”
Recovery: One targeted clarification, then proceed.
Service down
Edge caseThe model backend is unavailable.
“The assistant is temporarily unavailable.”
Recovery: Fallback to search or human path; never a fake answer.
Sensitive question
Edge caseThe question touches personal financial advice.
“This needs a human — here's how to reach one.”
Recovery: Contextual disclaimer plus escalation, built into the flow.
Timeout
Edge caseGeneration exceeds the patience budget.
“That took too long. Ask me again?”
Recovery: Restart with a clearer prompt — the user keeps their place.
07Final Experience
Ask Me — first view
Ask Me — first view
Scope badge and suggestion chips set expectations before the first question.
Schematic placeholder — add an approved screenshot via data/case-studies.ts
Grounded answer
Grounded answer
Short answer first, sources beneath, disclaimer in context.
Schematic placeholder — add an approved screenshot via data/case-studies.ts
Graceful refusal
Graceful refusal
Out-of-scope is a designed state with a human escalation path.
Schematic placeholder — add an approved screenshot via data/case-studies.ts
Feedback loop
Feedback loop
Thumbs on every answer — the system learns in the open.
Schematic placeholder — add an approved screenshot via data/case-studies.ts
08Learnings
- GenAI UX is mostly about limits: scope, sources, refusals and latency are the product.
- Trust in AI interfaces comes from visible grounding, not confident prose.
- The unhelpful path needs as much design as the happy path.
Next case study
Loan Status & EMI Journey
← /work · case study · ai-ask-me