Designing for the future of human-AI collaboration
XDesign Hackathon Winner 2026
AI-led Customer Service
0-1 Product Concept
ROLE
Product Design & Strategy
INDUSTRY
B2B AI
THE PROBLEM
A customer's bank account is frozen. And they been on hold for 12 minutes.
Today, AI agents handle up to ~90% of B2C calls, but when they encounter a compliance edge case, they abruptly hand off to a human agent who must step in and resolve the issue.
The human agent starts from zero.
So when the human finally joins, they know nothing, not why the customer is frustrated, not what the AI already tried, and not even that this is a compliance issue and not a billing one.
This is the designed outcome for most contact centres today. AI is deployed to contain calls, to resolve what it can and pass the rest to a human.
But the handoff itself has no design. It's a technical event, not a human one. And the cost is measurable.
THE BUILD
0 to 1, in 6 hours, working in true startup mode
We built features that truly redefines the AI Agent x Human collaboration.
Pre-recorded personalised hold message from the specific unavailable agent. buys time, keeps trust
Priority queue, top 3 calls by confidence and urgency, full AI action history per call
Context-switching stepper, between two high frustration calls.
Agent wellbeing monitor, emotional load tracking across a shift, proactive break suggestions
Live co-pilot, case details, recommended actions, inline policy and compliance checks, drag-to-confirm
Non-human caller detection, interaction pattern recognition for personal AI agent callers, confidence score, join-to-verify checkpoint
Pattern flagging when AI agent repeatedly escalates the same call type, auto-added to agent knowledge
Customer frustration tracking directly addressing context-switching for the human agent
Canvas mode, spatial clustering of all active calls by confidence and customer type, semantic background lighting below threshold
Features we wanted to build
AI moderation on the live call. AI stays on after the agent joins, passes context to both sides, exits only when the agent confirms, we explored this and stepped back. It added friction for the customer and complexity for the agent. The co-pilot handles it more cleanly, invisibly.
Adaptive communication mode for elderly. When the system detects low digital confidence (slow response pace, repeated hesitation, language mismatch), it automatically simplifies the AI's language, slows the interaction pace, and flags to the supervisor that this customer may need a more patient, guided handoff.