AI Agent Assist Dashboard
A live agent workspace: transcript, customer context and AI suggestions designed as human + AI collaboration — suggest, never decide.
Client
Enterprise Contact Center Platform (confidential)
Role
UX/UI Designer — suggestion interaction model, transcript UI, agent controls and feedback loops.
Platform
Web Agent Desktop
Confidentiality: Client identity, transcripts and customer data are confidential. Patterns shown are sanitised.
01Overview
An AI agent-assist workspace: live transcript, customer context panel, and AI suggestions presented as cards the agent chooses to act on. Designed around one principle — context in, human decision out.
02Context
Agents handle high-volume calls while juggling screens. AI could summarise, suggest and nudge — but only if suggestions arrive at the right moment, justify themselves and never act without the agent. The design goal was human + AI collaboration with an obvious line of responsibility.
03User Journey
How a suggestion becomes an action
04UX Decisions
01Suggestions are cards, never autocomplete
The agent stays in control: suggestions surface as actionable cards with a one-line reason, not as text the AI writes directly into the conversation.
02Show the why
Every suggestion carries a short rationale ('customer asked twice about the offer') — trust comes from visible reasoning.
03One action per card
Accept or dismiss. Editing happens inside the card's context, not in a separate flow.
04Keyboard-first placement
The suggestion panel sits next to the transcript with keyboard shortcuts, because agents move fast.
05PII masking by default
Sensitive fields are masked unless the agent has permission — the interface enforces policy.
05Interaction Design
States, transitions and behaviour — not just the happy path.
Suggestion arrival
A quiet card in the side panel with a subtle pulse — attention without interruption.
Accepted
The card collapses into a confirmed action with a timestamp; the transcript shows the effect.
Dismissed
One-tap dismissal with an optional reason — feedback becomes training signal.
ASR failure
Transcript stalls with an explicit state, not silence — the agent knows audio is not being captured.
No suggestion
The panel rests in a designed empty state instead of pretending to be idle.
Model latency
Suggestions that arrive too late for the moment are offered as 'missed opportunities', never injected mid-call.
06Edge Cases
Good UX is what happens when things go wrong. Each case: the problem, the user-facing message and the recovery action.
ASR failure
Edge caseSpeech-to-text drops mid-call.
“Live transcript paused.”
Recovery: Explicit state plus retry; the agent's flow does not break.
No suggestions
Edge caseThe model has nothing useful.
“No suggestions right now.”
Recovery: Designed rest state — not silence, not fake activity.
Stale context
Edge caseCustomer context is outdated.
“This profile may be out of date.”
Recovery: Marked stale with a refresh action before the agent acts on it.
Late suggestion
Edge caseThe suggestion arrives after the moment passed.
“Suggested earlier — view now?”
Recovery: Offered as a retrospective item instead of interrupting.
Sensitive data
Edge casePII appears in the transcript.
“Masked for your role.”
Recovery: Masking enforced by permission; nothing sensitive shown by default.
07Final Experience
Agent workspace
Agent workspace
Transcript, context and suggestions in three stable zones.
Schematic placeholder — add an approved screenshot via data/case-studies.ts
Suggestion card
Suggestion card
A suggestion with its reason — accept, dismiss or edit, never autocomplete.
Schematic placeholder — add an approved screenshot via data/case-studies.ts
Accepted action
Accepted action
Confirmed actions with timestamps; the loop is visible.
Schematic placeholder — add an approved screenshot via data/case-studies.ts
Rest + edge states
Rest + edge states
Designed states for ASR failure, stale context and no-suggestion moments.
Schematic placeholder — add an approved screenshot via data/case-studies.ts
08Learnings
- Human + AI interaction design is about the line of responsibility: suggest, never decide.
- Timing is a UI property — a late suggestion is noise, whatever its quality.
- Feedback loops must be visible for trust to grow.
Next case study
GenAI 'Ask Me' Experience
← /work · case study · agent-assist