Conversational AI + CSM: How a Major Telecom Deflected 3 Million Calls Annually
A tier-1 telecom replaced legacy IVR with virtual agents on ServiceNow CSM. Result: 3M deflected calls, 4.8/5 CSAT in 90 days, and $30M annual operational savings.
May 9, 2026·2 min read·By admin
At a glance
Industry
Telecom
Topics
CSM · Now Intelligence / AI · Servicenow
Published
May 2026
The challenge
A major telecom provider with 100M+ subscribers had reached the limit of its legacy IVR + agent-routed contact center. Hold times were creeping past 9 minutes; first-call resolution had stagnated at 52%; and the operational cost-per-contact was rising every quarter as call volume grew faster than headcount budgets.
Leadership wanted a 30% deflection rate without sacrificing CSAT — a target most peers had struggled to hit even with significant AI investment.
The Mergen approach
We ran a 4-week discovery analyzing 6 months of call recordings, identifying the top 50 intents that accounted for 78% of volume. Of those, 32 were genuinely self-serviceable with the right authentication, billing-API access, and conversational design.
The architecture: ServiceNow CSM as the system of record, virtual agents (ServiceNow + custom NLU on Azure for telco-specific intents like “my data is throttled” or “WiFi calling won’t activate”), seamless escalation to human agents with full conversational context preserved.
The outcomes (first 12 months)
3 million calls deflected annually via self-service or virtual agent
4.8 / 5.0 CSAT on virtual-agent-resolved interactions (within 90 days)
32% deflection rate overall — beat target
$30M / year in reduced operational cost
+18 NPS points on the customer service experience metric
Average handle time on escalated calls: down 28% (because agents picked up with full context)
What the team got right
Two things distinguished this rollout from peer attempts. First, we treated authentication as a first-class problem — the virtual agent could verify identity to the same standard as a human agent, which unlocked 70% of the deflectable intents. Second, we ruthlessly limited the launch scope to the 32 highest-volume intents and didn’t add more until each was sustaining 90%+ task-completion rates.