AI Voice Assistant for Self-Service Operations

The problem

Telecom call centres handle enormous volumes of routine calls — billing questions, plan changes, connectivity troubleshooting. The volume is high enough that small deflection percentages translate into large operational numbers, but voice self-service has a reputation problem: callers have learned that phone menus waste their time.

How it works

Dialogflow handles intent recognition so callers state their problem in their own words rather than navigating a menu tree. Twilio Voice carries the telephony. The workload is bursty and unforgiving — call volume spikes with outages and billing cycles — so the system runs on Kubernetes, where capacity can follow demand rather than being provisioned for the worst case permanently.

What shaped it

At subscriber scale, the tail matters. An assistant that handles the common path well but strands callers on anything unusual generates escalations angrier than the ones it avoided, so routing to a human early was treated as a correct outcome rather than a failure.

Satisfaction was tracked alongside deflection deliberately. Deflection alone is easy to improve by making it harder to reach a person, which is not the same as solving anything.

Outcome

A significant share of call volume was handled without an agent while satisfaction scores held up.