The industry talks about AI. Customers leave quietly.

Why AI-driven customer retention at energy suppliers depends on processes and ownership, not on the technology

Many suppliers lose existing customers and only notice once the cancellation is already in the system. This is exactly where AI can make a difference. Whether it does depends less on the model than on whether processes, data and responsibilities across sales, service and IT are organised well enough to turn a prediction into action.

Customer retention used to be an afterthought

For decades, keeping customers was barely a task for energy suppliers. Once someone signed up for electricity or gas, they usually stayed. That’s over. Switching providers within 24 hours, expiring sign-up bonuses, digitally native competitors and comparison portals that make every price change visible instantly have pushed the switching barrier to almost zero.

That has shifted the weight of customer retention. Every lost existing customer takes their contribution margin with them and has to be replaced through expensive new customer acquisition. Common estimates put win-back costs at several times what retention would have cost. This is especially painful for customers won through bonuses. Once the bonus expires, the price gap becomes visible, the acquisition cost is lost, and replacing that customer costs money all over again.

The real problem is invisibility

Churn is not a new phenomenon for suppliers. In 2024, around 7.1 million electricity customers and 2.3 million gas customers switched providers, more than at any point since records began, and 2025 came in only slightly lower.¹ What’s new is the speed. At most suppliers, a customer only shows up on the radar once the cancellation arrives, and with a 24-hour switch, there is no room left to react.

On top of that, the customer base is unevenly split. One part switches regularly, driven by price. Another stays out of habit, proximity or convenience, often on the standard supply tariff. Anyone who targets both groups with the same campaigns spends retention budget on customers who would have stayed anyway, and still reaches the switch-prone ones too late.

Price is the most common reason customers switch, but a weak lever for retention. A customer who stays only for the price leaves for the next cheaper offer. A regional supplier rarely wins a price war against digital competitors. What sets it apart is the experience customers have with it: how fast a query gets answered, how understandable the bill is, how easy it is to register a wallbox or switch a tariff, whether an offer arrives at the right moment. That experience is hard to copy from outside. So far, though, it is also barely managed systematically, because most suppliers have only limited visibility into which customer is having which experience, and when they are at risk of leaving.

Four points where AI makes a difference

Across the customer relationship, there are four phases where AI has a measurable effect on retention. They differ significantly in what they require.

Onboarding. Sign-up, data transfer and the first communication decide whether trust is built. Automated checking and processing of incoming data reduces queries and errors at a stage when the customer is paying particularly close attention. The prerequisite is a cleanly defined process that AI can carry out.

Day-to-day operations. This is where most contact happens. AI in email, voice and chat shortens response times and resolves recurring requests, making billing and consumption easier to understand. The benefit lands quickly and is easy to measure. It also lays the foundation for the next two phases, because this is where the customer data is generated.

Critical moments. Price changes, bonus expiry, contract end and house moves are the points in time where cancellations originate. A model that flags an individual customer’s churn risk ahead of these moments buys back the lead time that a 24-hour switch otherwise removes. That requires data history, segmentation and, above all, a defined response.

Targeting and offers. Consumption, contract status and life situation reveal which offer fits which customer at which moment, from a tariff switch to a heat pump or a wallbox. This broadens the customer relationship and raises the switching cost on the customer’s side. Here too, the quality of the recommendation depends on the quality of the master data.

The first two are achievable with manageable effort. The last two need a foundation that is still missing at many suppliers.

Why many AI initiatives fail to deliver

All four applications share the same prerequisite, and it is the one most often overlooked. Without standing processes, reliable data and clear ownership, AI scales the problems that were already there. A churn score stays just a number until someone has defined who calls at which threshold, which offer sales is allowed to make, and who measures whether the action worked. AI-powered service running on messy master data mostly automates errors.

From our projects, we estimate that around 70% of friction comes from people and processes, about 20% from technology and infrastructure, and around 10% from the models themselves. The lever, then, is in how the deployment is organised.

This is mainly about how the business side and IT work together. At many suppliers, it is still unclear who actually owns a churn model. IT runs it, sales is expected to act on it, service sees the customers, and data quality belongs to everyone and no one. AI shifts these roles. The business side takes on responsibility for the content: what the model should do, what response follows, and whether it worked. IT provides the platform, ensures data quality and traceability, and checks whether a use case meets the transparency and human-oversight requirements the EU AI Act demands anyway. Where this division is missing, AI stays an IT tool that no one in sales or customer service actually uses.

What one example from practice shows

A regional energy supplier with around 200,000 customers handled roughly 82,000 emails and 200,000 calls a year, largely manually. Restructuring service processes, targeted AI use in email, voice and chat, and cleaned-up data management brought the email backlog down from over 4,000 to around 400 open cases. Response time fell from up to four weeks to two to three days, and phone accessibility rose from 55% to over 80%.

Those are service metrics. On their own, they say little about retention. The connection is direct nonetheless. Only a service that actually resolves issues creates the customer experience that offsets a price comparison. And only with the clean data that results from it can churn risk be identified at all. The sequence follows from that. Day-to-day operations first, then the critical moments.

What suppliers can take from this

Customer retention has turned from an administrative task into a revenue lever for suppliers, and AI is the first tool that makes churn visible before it happens. Suppliers that put the prerequisites in place gain in four areas.

The ability to act before the cancellation. A backward-looking cancellation statistic becomes an early-warning system. Churn risk becomes visible at the critical moments, while sales and service can still react.

Less wasted spend on retention and win-back. Retention budget flows to the customers actually at risk, not to loyal customers who would have stayed anyway. Every relationship kept this way saves the far more expensive cost of winning a new customer.

A service that keeps customers. Shorter response times, resolved requests and an understandable bill work at exactly the touchpoints that offset a price comparison. At the same time, this is where the data that churn detection and tailored offers build on gets generated.

Protected margin. Retained existing customers preserve the margin that would otherwise be lost to a price war or an expensive replacement. Well-matched add-on products broaden the relationship and raise the switching cost on the customer’s side.

Our project experience points to five steps that build on each other in this order:

1. Name where customers are lost.
   Where in sales and service are customers actually lost, at which triggers, and in which segments? This can be worked out from existing cancellation and contact data within a few weeks, and shows where service and churn detection should focus first.
2. Stabilise service.
   Automate recurring requests in email, voice and chat, and clean up master data. The benefit is quickly measurable, and the data foundation for churn detection is built as a side effect.
3. Define the response before the model.
   Work with sales, service and IT to define who responds at which risk level, with which action, which offers are allowed, and how success is measured. Only then does building the churn model pay off.
4. Clarify ownership.
   A business owner for the use case, a technical owner for the platform, data quality and the check against the AI Act’s requirements. This pairing decides whether AI gets used in sales or stays an IT project.
5. Start small, scale deliberately.
   One segment, one trigger, one feedback loop of prediction, action and measurement. What works there can be carried over to other moments in the customer journey.

Suppliers that go through these steps usually find that the model is the smallest part of the task. The question to start with, then, is this: if a model flags 500 customers with high cancellation risk tomorrow, who calls them, with what offer, and how will you know in three months whether it worked?

¹ Bundesnetzagentur and Bundeskartellamt, Monitoring Report 2025 (https://www.bundesnetzagentur.de/1081882); Bundesnetzagentur, press release on supplier switches 2025, March 2026 (https://bundesnetzagentur.de/1097294).