Case Study · Enterprise AI · 2023–2024
Contact Centre 2030
A forward-looking agent workspace for a US-based telecom carrier — one human supervises 5–10 concurrent AI-mediated chats, with sentiment-triggered escalation.
The fast read
- A 2030-vision platform brief: design the agent workspace for a future where 1/5 of conversation volume is handled by AI.
- Reframed the role from multi-tasking agent to supervisor of AI subordinates — different mental model, different UI.
- Sentiment-triggered handoff designed as an explainable signal, not a binary alert.
- Sized as a 4–8× throughput multiplier for the contact-centre function.
- Under NDA. Client name, internal screens, and KPIs withheld.
Context
The brief, in 2023
One of the largest customer-service workforces in the US, operating at a scale where small per-conversation improvements compound to tens of thousands of agents and billions in cost. The in-call agent workspace — the one surface every customer issue is touched through — was due for a forward-looking redesign.
The brief was deliberately ambitious: design what this workspace looks like in 2030, when AI handles a meaningful share of the conversational load and the human role has shifted from doing the talking to supervising the talking.
The problem
Three layers
For the human agent of 2030
"If I'm responsible for 5–10 conversations at once, all happening in parallel, how do I stay cognitively present in any of them?"
For the business
"How do we capture the productivity multiplier of AI-mediated chat without losing service quality, brand trust, or the human craft that defines our customer experience?"
For the design
"Design the screen, the workflow, the handoff signals, the trust UI, and the supervisor view — for an agent who is simultaneously the operator of multiple AI subordinates and the responsible escalation point for any of them."
My role
Design lead — workflow, blueprint, implementation
Led the design strategy on the engagement: workflow architecture, service blueprints, and the supervisor-model framing. Managed the design team executing the visual and interaction work — the actual screens were drawn by talented designers I directed, not by me.
The other half of my role was translating the design model into something engineering could ship — partnering with developers through implementation, reviewing build, and making sure the intent survived the trip from blueprint to production.
Approach
Reject the multi-tasker. Adopt the supervisor.
The instinctive frame for "one human handling 5–10 conversations" is multi-tasking — and multi-tasking UI is a known failure mode at any density above 2. I rejected the multi-tasker framing entirely and reframed the human as a supervisor of AI subordinates. Each chat has an AI agent operating it; the human is one tier up, intervening only when sentiment, complexity, or business risk crosses a threshold.
The UI follows that metaphor — not a Slack-style chat sidebar with 10 active conversations, but a queue of supervisory exceptions with full-context loading on demand. That single reframe collapsed the entire interaction model from "impossible to design" to "tractable."
Key decisions
Four trade-offs that shaped the design
Decision 01
The supervisor metaphor, not the multi-tasker
The initial instinct in the room was to design for parallel chats. We worked through the cognitive limits together and reframed the brief: the human isn't multi-tasking, they're supervising AI agents. That blueprint decision shifted the whole interaction model — the team then designed the UI to match the supervisor metaphor, not the multi-tasker.
Decision 02
Sentiment as an explainable signal
A binary sentiment alert ("escalate now") would have failed agents in the field. We landed on sentiment as a trending, explainable signal — the receiving human should see why the handoff fired before they say a word. The team designed the signal treatment; we defended the principle together through three stakeholder reviews that wanted it simpler.
Decision 03
Context-loading in the first five seconds
When a human takes over from AI, the first five seconds are everything. We specified the context-load contract together: what information surfaces, in what hierarchy, in what timeframe. The team designed the screen; we reviewed every iteration against the "no reading required" bar before it shipped to dev.
Decision 04
Hybrid mode — human voice, AI hands
The brief treated AI as either fully autonomous or fully absent. The team and I worked through the cost of each end and landed on a hybrid: the supervisor takes the call while the bot works silently in the background — live-transcribing the conversation, pulling relevant context from product manuals and account history, surfacing the answer before the supervisor needs to ask, and post-call handling the busy-work (raising or updating tickets, logging interactions, generating the summary). The team designed the in-call assist sidebar; we aligned engineering on the live-transcribe-to-search pipeline contract.
01 / 04
Outcome
What it became
- Vision design accepted as the platform direction by the client's CX leadership.
- Sized as a 4–8× throughput multiplier for the contact-centre function — based on the per-agent concurrent-chat math.
- Patterns from this work — supervisor metaphor, sentiment as explainable signal, accept-or-edit trust UI — became reference points for downstream work.
- Specific internal KPIs, screens, and rollout plans remain under NDA.
Reflection
"The hardest problem in AI-augmented work isn't designing the AI's output. It's designing the human's role in a world where the AI handles 80% of what used to be the job."
The next decade of customer experience belongs to companies that design human-AI orchestration models — not chatbots. The seat the remaining humans occupy is the harder, more interesting design problem. That's the work I want more of.