Churn prediction is one of the oldest applications of machine learning in telecoms, and one of the most over-promised. Almost every operator has a churn score. Far fewer can show that the score changes what happens to a customer, or that the retention budget it directs earns a return. The gap is not the model; it is everything between the model and the customer.
Why churn economics differ by segment
The starting point is that churn is not one number. In a large US operator's mid-2025 results, postpaid phone churn was 0.90% a month while prepaid churn was 2.65%2; another large US operator reported postpaid phone churn of 0.86% in the second quarter of 20261. A Canadian operator reported 1.22% monthly postpaid phone churn against 4.02% for prepaid in early 20263. In a Latin American group's largely prepaid markets, monthly churn in the second quarter of 2026 ranged from 2.6% in Mexico and 3.7% in Peru to 4.3% in Colombia and 5.5% in another, predominantly prepaid, market4.
Churn definitions also vary. Some operators count only voluntary disconnections, others include involuntary churn from non-payment; prepaid definitions depend on inactivity windows that differ by market. Before any model is built, the business needs one agreed definition per segment, or model performance and financial impact cannot be compared.
Prepaid customers churn roughly three times as fast as postpaid
Monthly churn, %, selected North American operators (%)
Note: From operators' published quarterly results. Sources: [2], [3].
Source: SEC EDGAR, “Second-quarter 2025 results of a large US operator (SEC Form 8-K exhibit)” (2025)
Those differences change the problem. In postpaid, churn is rare, contract-bound and predictable from events such as contract end, bill shock, network experience and competitor offers; the task is targeted retention of high-value customers. In prepaid, churn is frequent and often silent: customers simply stop topping up or move to another SIM. There the task is engagement, top-up behaviour and multi-SIM share. A single model and a single retention playbook serve neither well.
The economics also argue for prevention. A widely cited heuristic holds that acquiring a new customer is five to 25 times more expensive than retaining an existing one5. It is cross-industry and dated, but in markets where handset subsidies and acquisition commissions are material, operators usually find it directionally right.
What the evidence says about results
McKinsey estimated that companies implementing a comprehensive, analytics-based approach to base management can reduce churn by as much as 15%6. Its later work on customer value management (CVM) found that operators can raise revenues by up to 10% and customer satisfaction and engagement by 20% to 30% by harnessing analytics-driven CVM, yet only around 5% of telcos were unlocking that potential, although roughly 70% had established successful pilots7. In one case, an operator tripled revenue from CVM from around 2% to around 6% of revenue within two years7.
Reported and estimated value from AI-driven customer value management
Selected results and estimates, telecom operators
| Lever | Result or estimate | Type |
|---|---|---|
| Analytics-based base management | Churn reduced by as much as 15% | Estimate |
| Full-potential CVM | Revenue up by as much as 10%; satisfaction and engagement up 20–30% | Estimate |
| Bill-shock intervention using propensity and CLTV models (Asia-Pacific telco) | Churn down 5%; ROI almost four times higher than before | Reported case |
| GenAI-personalised marketing content (European telco) | Campaign conversion up 40% | Reported case |
| Service-aware next best experience | Early-life churn reduced by as much as 30% | Estimate |
Note: Sources: [6], [7], [8], [9], [10]. Case results are self-reported and anonymised by the source.
Source: McKinsey & Company, “Next best experience: how AI can power every customer interaction” (2025)
The most instructive case is an Asia-Pacific telecom company that chained three capabilities. A propensity model identified customers most likely to call about billing issues; a decision engine used customer lifetime value (CLTV) to prioritise interventions; and proactive outreach addressed the bill before the customer called. The improvement in bill-shock management reduced churn by 5% and drove an ROI almost four times higher than before8. GenAI is extending the content side: one European telco increased marketing-campaign conversion by 40% while reducing costs by using GenAI to personalise content9, and customers receiving another European telco's GenAI-personalised messages engaged and acted 10% more often11.
Service is part of retention. McKinsey estimates that service-aware next-best-experience approaches can improve customer satisfaction by 10% to 20% and reduce early-life churn by as much as 30%10. Many churn events start as a network or billing problem that the retention team never sees.
Prepaid needs its own playbook. Without a contract end date or a bill, the signals are behavioural: declining top-up frequency, shrinking recharge values, falling data use, a second SIM appearing in the same device, or reduced activity on the operator's wallet and loyalty apps. The window to act is short, and the right action is often small and immediate, such as a personalised bundle at the next recharge or a loyalty reward delivered through the channel the customer already uses. Distribution partners matter too: much prepaid engagement happens through retailers and agents, whose incentives and data need to be part of the loop.
Governance applies here as well. Retention offers are, in effect, personalised prices, so operators need rules on fairness, consistency and explainability, and a clear line between retention and credit decisions. Where AI is used to evaluate the creditworthiness of individuals, for example in postpaid vetting or device financing, the EU AI Act treats it as high-risk, with obligations now applying from December 2027 (see our article on telecom AI governance). Keeping those decisions in a distinct, well-documented model reduces regulatory exposure for the rest of the customer decision stack.
The decision layer is the product
- Uplift, not propensity. Target customers whose behaviour an offer will change, not those most likely to leave regardless; a high churn score on a customer who will leave anyway is wasted budget.
- Value-weighted. Prioritise by expected lifetime value saved net of offer cost, so retention spend follows margin.
- Arbitrated. A single next-best-action engine chooses between retention, upsell, service fix and doing nothing, so channels stop competing for the same customer.
- Closed loop. Hold-out groups measure incremental effect for every action, and results retrain the models.
- Governed. Offers respect consent, fairness and pricing rules, and front-line staff can see why an action was recommended.
Customers expect this. McKinsey research found 71% of consumers expected companies to deliver personalised interactions and 76% were frustrated when that did not happen11. Operators hold more signal about their customers than almost any other consumer business; the limiting factor is the ability to act on it consistently.