What problem does it solve?
In mature mobile and broadband markets, growth often depends on keeping customers as much as on winning new ones. Customers leave for a better price, after repeated faults, after a bill shock or at the end of a contract, often without contacting the operator first. Traditional retention reacts only when the customer calls to cancel, when the decision is already made, and relies on a save desk that offers the same discount to everyone.
Blanket discounts are expensive and can teach customers that threatening to leave pays. What operators need is earlier warning, a reason for the risk, and an action that fits: fixing the fault that annoyed the customer can be worth more than any discount. At the same time, regulators fine operators who make leaving hard (Ofcom fined Virgin Media £28 million in July 2026), so retention has to help customers, not trap them.
How does it work?
- Score the risk. A model scores every customer regularly, and in real time on events such as a failed repair or a price rise, from usage, network quality, billing, contact history and contract end dates.
- Explain the reason. For each high risk customer it gives the main drivers (repeated faults, a better competitor price, a bill shock) so the action matches the cause.
- Choose the next best action. A decisioning engine picks the action with the best value for customer and operator within budget: resolve the fault, move to a better fitting plan, a loyalty benefit, or an offer, and sometimes no action at all.
- Deliver it in the right channel. The action appears in the app, in a message, as a prompt to the advisor on the next call, or in a conversation with an AI agent, subject to consent.
- Learn from the outcome. Accepted and ignored offers and actual churn feed back into the models, measured against a control group.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Mainstream
- Channels
- Agent desktop, Mobile app, SMS, Email, Phone and voice
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Churn reduction | Too few to pool | 15% to 20% | 2 | 2 vendor |
| Conversion uplift | Too few to pool | 75% | 1 | 1 vendor |
Value drivers: Revenue growth, Customer experience, Lower cost to serve.
Indicative value
An operator with 2 million postpaid mobile and broadband customers
USD 900,000 to USD 8.1 million
First year revenue retained, net of retention offers per year
How this is calculated
Formula: customers * churnRate * churnReduction * annualRevenue * retainedMarginShare. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Postpaid customers customers, customers | 2,000,000 | 2,000,000 | The reference operator. |
| Annual churn rate churnRate, fraction of customers per year | 0.12 | 0.18 | Editorial assumption, replace with your own annual postpaid churn. |
| Relative reduction in churn among customers the programme reaches churnReduction, fraction of churn | 0.03 | 0.08 | Conservative against the benchmarks on this page (Pega reports a 20% churn reduction at Telenet and 15% at Etisalat, a figure it first reported for Etisalat's SMB division), because those are vendor figures without a published control group. |
| Annual revenue per customer annualRevenue, USD per customer per year | 250 | 400 | Editorial assumption, replace with your own average revenue per postpaid account. |
| Share of retained revenue kept after the cost of retention offers retainedMarginShare, fraction of retained revenue | 0.5 | 0.7 | Editorial assumption covering discounts and benefits given to retained customers. |
What it leaves out: Counts only one year of retained revenue net of offer costs. It leaves out the lifetime value of retained customers, savings from fixing root causes, the cost of models and decisioning, and customers who would have stayed anyway, which only a control group can remove.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Virgin Media O2
United Kingdom · Telecommunications · 2025
Virgin Media O2 built its own tool, Lumi AI, that analyses a live conversation and prompts the advisor with resolutions that worked for similar customers and with the products and services most likely to interest this customer. In July 2025 it was in pilot with a cohort of advisors in care, telesales and retentions, with a wider rollout planned. Alongside it the operator uses an AI contact centre service from Amazon Web Services that routes callers by their stated reason, software that flags potentially vulnerable customers, and automatic call summaries. The retention effect of Lumi AI has not been published.
No outcome disclosed.
Telenet
Belgium · Telecommunications · 2023
Telenet, a provider of connectivity and entertainment services in Belgium, uses Pega's AI based Customer Decision Hub as a single decisioning system that responds to customer signals in real time, anticipates how behaviour may change and proposes the next best action, such as personalised upgrades and solutions, with the stated goals of reducing churn and raising offer acceptance. Pega reports a 20% reduction in churn, a 75% increase in offer acceptance and a 33% increase in cross sell. These figures cover the whole decisioning programme, including upgrades and cross sell, not retention alone.
- Churn reduction: 20%
"20% reduction in churn"
Claimed by: vendor - Conversion uplift: 75%
"75% increase in offer acceptance"
Claimed by: vendor
Vodafone UK
United Kingdom · Telecommunications · 2023
Vodafone UK's Always on Marketing programme runs on Pega Customer Decision Hub. Customer interactions and events are processed as they happen and mapped to common intents, so one central system presents the next best action for each customer at every touchpoint instead of pushing products. Pega titles the case study as improving customer retention and says the programme was halfway through its transformation, but publishes no figures.
No outcome disclosed.
Etisalat
United Arab Emirates · Telecommunications · 2021
Etisalat in the UAE uses Pega Customer Decision Hub as its central decisioning engine, with predictive and adaptive models that identify each customer's context and orchestrate personalised next best actions across inbound, outbound and agent assisted channels, moving from a focus on sales to a focus on incremental value. During the COVID-19 pandemic it used the system to identify at risk customers and offer practical help such as a free VPN and its online collaboration platform. Pega reports a 15% reduction in churn and a 20% increase in renewals, and its executive quote describes a 20% year on year increase in incremental value; the churn and renewal figures first appeared in Pega's earlier case study on Etisalat's small and medium business (SMB) division, which used next best action to prioritise outbound sales calls and offers, not in connection with the COVID-19 work. All figures cover the whole decisioning programme, including sales, not retention alone.
- Churn reduction: 15%
"15% reduction in customer churn"
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- At least two years of customer history with churn outcomes and reasons
- Network quality, fault and repair data linked to customers
- Billing events, price changes and contract end dates
- Contact history and complaint records
- Marketing consent and contact preferences
Systems to integrate
- Data platform joining network, billing, CRM and contact data
- Decisioning or next best action engine with offer rules and budgets
- Agent desktop and store systems for advisor prompts
- App, messaging and email channels for digital actions
- Campaign management with consent and contact frequency rules
Complexity: Medium
Churn models are well understood; the hard parts are joining network, billing and contact data per customer in near real time, running a decisioning engine with offer budgets, and delivering actions in every channel with consent and a measurable control group.
- 1
Define churn and the outcomes you will measure
Agree on what counts as churn (port out, cancellation, downgrade) and set up a permanent control group before the first model goes live.
- 2
Build reasons, not only scores
Train the model to expose its main drivers per customer, so actions can address the cause. A fault driven risk needs a fix, not a discount.
- 3
Put actions under a decisioning engine with budgets
Define the available actions, eligibility and cost, and let the engine choose within budget, including the option to do nothing for customers who will stay anyway.
- 4
Bring it into the conversation
Show the reason and recommended action to advisors on every call, and give AI agents the same recommendation, so a customer who says they want to leave gets a relevant answer.
- 5
Keep cancellation easy
Design retention so a customer who still wants to leave can do so in the same conversation. Treat that as a hard requirement, tested like any other journey.
Guardrails
- A customer who confirms they want to leave is helped to leave in the same contact
- Offers only within approved budgets and eligibility rules
- Marketing consent and contact frequency limits enforced before any outbound action
- Protected characteristics and proxies excluded from features, with regular fairness checks
- Vulnerable customers routed to trained people, not to automated offers
KPIs to instrument
- Churn in treated customers versus the control group
- Offer acceptance and cost per retained customer
- Share of high risk customers whose root cause was fixed
- Time to complete a cancellation for customers who still want to leave
- Complaints about retention contacts or cancellation
Human in the loop
Retention advisors decide on offers above standard limits and handle vulnerable customers. The commercial team owns the action catalogue and budgets, and an analytics team reviews model performance, fairness and the control group results monthly.
Common failure modes
- Paying customers who would have stayed
- Offers go to customers with a high score who were never leaving. Use uplift models and a control group.
- Making it hard to leave
- Retention becomes obstruction and a regulatory breach. Measure and protect the cancellation path.
- Discount addiction
- Customers learn that threatening to leave gets a discount. Favour fixing root causes and limit repeat offers.
- Unfair outcomes
- Better offers go systematically to some groups. Exclude protected attributes and proxies and audit outcomes.
What are the risks and rules?
EU AI Act
Depends on design
Churn scoring and offer selection for marketing are not listed in Annex III, so a back office design that only scores customers and prompts human advisors is minimal risk, with no specific obligations. When an AI agent delivers the offer to the customer in chat, messaging or voice, the system is limited risk: Article 50 requires telling customers they are dealing with AI. A design that used manipulative techniques or exploited vulnerabilities to keep customers from leaving could fall under the Article 5 prohibitions. GDPR rules on profiling and the right to object to direct marketing (Article 21) apply in full.
Rules that apply
Guidance
- Ofcom fines Virgin Media £28m for repeatedly preventing customers from cancelling contracts (Ofcom, Europe). A £28 million fine (July 2026) for retention practices, including a two tier cancellation process and agents rewarded for deterring cancellations, that caused customers unreasonable effort when trying to leave; the line any AI retention design must not cross.
- Directive on privacy and electronic communications (Directive 2002/58/EC) (European Union, Europe). Article 13 sets the rules for unsolicited electronic marketing, which apply to proactive retention offers by messaging, email or automated calls. Automated calls need prior consent; under the Article 13(2) soft opt in, an operator may email or message its existing customers about its own similar products or services, provided they can object free of charge with every message.
Controls to put in place
- Model inventory entry with owner, features, validation and fairness results
- Permanent control group and monthly incrementality reporting
- Documented action catalogue with budgets and approval
- Consent, frequency and vulnerability checks enforced in the decisioning engine
- Audit of cancellation journeys for unreasonable barriers
Frequently asked questions
- How much can AI reduce telecom churn?
- The public figures are vendor reported: Pega reports a 20% reduction in churn at Telenet, and a 15% reduction at Etisalat that it first published for Etisalat's small and medium business division. Neither comes with a published baseline or control group, so plan conservatively and measure against your own holdout.
- Is it legal to use AI to persuade customers to stay?
- Yes, if it helps rather than obstructs. Offers must respect marketing consent and profiling rules, and a customer who wants to leave must be able to. In July 2026 Ofcom fined Virgin Media £28 million for retention practices that made cancelling unreasonably hard.
- Where should retention actions be delivered?
- Wherever the customer is: as a prompt to the advisor on a call (Virgin Media O2 piloted its Lumi AI advisor prompts with care, telesales and retentions teams in 2025), in the app, or in a conversation with an AI agent that has the same recommendation.
How to cite this page
Blits.ai AI Use Case Library, "AI for telecom churn prediction and retention offers", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/churn-prediction-and-retention-offers. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published