Design a Virtual Call Center

Problem statement

Design a virtual call center: a central hub that accepts inbound calls and chats, places them in queues when no suitable agent is free, routes each contact to an available remote agent based on skill set, priority, and load, bridges the real-time media or messaging session, and exposes live operational dashboards to supervisors. The problem requires an inbound queue with priority or skill-based routing, agent availability and presence management, real-time call or chat bridging, and reporting on calls handled and wait times. Remote agents work from home or distributed offices over ordinary broadband, which means presence, connectivity, and media quality are untrustworthy by default.

This is not a CRUD application with a notification bolted on. It is a real-time orchestration system with three clocks running at once. The first clock is the contact lifecycle: a caller dials in, traverses an IVR, enters a queue, waits, is assigned, connects, talks, and hangs up, all while accruing service-level time. The second clock is the agent lifecycle: agents sign in, go available, take contacts, enter wrap-up, take breaks, and sign out, all while their browser or softphone can crash at any second. The third clock is the business: service level agreements such as answering 80 percent of calls within 30 seconds, occupancy targets, abandon-rate ceilings, and compliance obligations like recording consent and payment-card redaction. The architecture is judged on whether these three clocks stay coherent under partial failure.

Why the problem is distinctive

An e-commerce backend can retry a failed write a second later. A call cannot retry a second of silence: every extra second of dead air, ringing into nothing, or dropped audio is directly experienced by a human and directly measured by the business. The design therefore separates three properties that weaker answers merge. Contact progress is a durable workflow: queued, assigned, connected, completed, or abandoned, and it must survive service restarts. Routing correctness is an ownership problem: exactly one agent owns a contact at assignment time, enforced with a lease and a fencing token, because double-connecting a caller to two agents is a visible incident. Media continuity is a latency-critical path that must tolerate agent network degradation without dropping the contact, which means the media plane must be decoupled from the signaling plane so that one can fail over while the other holds state.

Public operating baseline versus design assumptions

Public evidence shows the category is operationally real and enormous. Genesys reports that its platform powers more than 25 billion customer interactions annually across more than 100 countries. Amazon Connect, launched in 2017, is the same service Amazon uses for its own retail contact centers and publishes per-minute pricing rather than seat licenses, with voice priced per minute and chat priced per session. Twilio Flex, launched in 2018, ships routing as a product called TaskRouter with expression-based skill matching. These are company-reported public figures, used as context, not as requirements for our fictional system.

For capacity planning, this answer explicitly assumes a mature multi-tenant platform with 50,000 registered agents, 20,000 agents concurrently signed in at daily peak, 1.2 million contacts per day across voice, chat, and asynchronous channels, and a five-times event peak. Unless a number is tied to a public source, it is a stated design assumption, target, or budget.

The four architectural planes

1. Contact plane: telephony ingress via SIP trunks and an SBC, WebRTC ingress for browser callers, IVR flows, durable queues, priority and skill-based routing, and the contact state machine.
2. Agent plane: presence and concurrency management, the agent desktop (softphone plus chat console), heartbeat leases, capacity slots, and wrap-up workflow.
3. Orchestration plane: routing engine, queue configuration, contact flows, workforce scheduling, dashboards, and reporting.
4. Compliance and learning plane: recording, redaction, transcription, sentiment analytics, moderation for chat, audit, and retention.

A strong interview answer keeps these planes separate. The contact plane must keep queuing contacts even when the dashboard plane is down. The agent plane must degrade a flapping agent to unavailable instead of corrupting routing state. The compliance plane must never block the live media path, and the live media path must never be able to bypass the compliance plane when recording is mandatory.