Telecommunications · Field Notes
June 2026 · 6-minute read · Asseblia Insights
Churn prediction has existed in telecoms for two decades. Most operators have a model somewhere — usually a quarterly batch job outputting a list of at-risk customers that lands in the CRM and gets ignored because the retention team is already at capacity.
What’s changed in the last 18 months isn’t the prediction. It’s the response. LLMs and agentic AI systems now make it possible to act on a churn signal automatically, at scale, without a human in the loop for every intervention. The gap between knowing a customer is about to leave and doing something about it has collapsed from weeks to hours.
Why previous approaches failed
The traditional churn stack at most operators looks like this: a data science team builds a model, it runs monthly (or quarterly if resources are stretched), outputs a CSV, and the CSV goes to a retention team that calls the top 500 customers. At 30% contact rate and 15% save rate, you’re keeping maybe 22 customers per cycle.
The problems are structural. Batch processing means you’re acting on data that’s already 30 days old. The save offer is the same for every customer regardless of their actual reason for churning. And the retention team is a bottleneck — there’s no way to scale human outreach to match the volume of at-risk customers in a large base.
What an AI-native churn system looks like
A modern churn intervention stack has three layers:
Real-time risk scoring
A model running continuously against live usage data, billing events, service ticket history, and network experience metrics. Not batch. Not monthly. Continuously.
Intent classification
A second layer that classifies the likely reason for churn from behavioral signals. Price sensitivity looks different from service quality dissatisfaction, which looks different from competitive switching intent.
Automated intervention
Based on risk score and intent, the system selects and executes the appropriate intervention: a personalized offer from a predefined library, proactive service outreach, or a human-agent flag at high-value thresholds.
The LLM’s role in this stack is primarily at the intervention layer — generating personalized message content and selecting the right offer configuration for each customer segment. It’s not writing novel content from scratch; it’s selecting and assembling from a structured offer library in a way that feels personalized.
The data requirements
The biggest barrier to deployment isn’t the AI — it’s the data. A functional churn prediction and intervention system needs:
→ Usage data at ≥ daily granularity (call records, data consumption, roaming events)
→ Billing history and payment behavior — late payments are a leading churn indicator
→ Service ticket and complaint history
→ Network quality-of-service metrics per customer, where available
→ Contract tenure and plan history
Most tier-1 operators have all of this. The problem is that it lives in four different systems with four different schemas and nobody has connected it yet. Data integration is typically 40% of the implementation effort on these projects.
What to expect
Based on published operator deployments, a well-implemented AI churn system typically delivers the figures shown at the top of this article: 25–35% reduction in voluntary churn, 15–20% improvement in retention offer acceptance, and 60–70% reduction in manual retention workload.
The implementation timeline from project kick-off to live system is typically 8–14 weeks, depending on data readiness. The ROI calculation is straightforward: take your current monthly churn rate, your average revenue per user, and your customer acquisition cost, and model the impact of a 30% churn reduction. For most operators, it’s a compelling number.
The one thing most operators get wrong
They build the prediction and forget the governance. An AI system that’s generating retention offers autonomously needs a clear set of rules about what it can and can’t do: offer value caps, customer segment exclusions, frequency limits, and a human review queue for edge cases. Without governance, you end up with an AI that’s offering six months of free service to customers who were going to churn anyway because their contract just ended.
“Governance isn’t a constraint on the system — it’s what makes it safe to run at scale.”
Next step
Want to explore an AI churn system for your operation?
Book a 45-minute discovery call — we’ll map one high-impact automation opportunity. No commitment required.