Customers likely to leave are spotted early and offered something that fits their needs, within offer rules a person approved. Every customer still gets their end-of-contract and best-tariff notices, and nobody pays more because a model thinks they won't leave.
At a glance
An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.
Most customers leave at the end of their minimum term. Today the provider finds out when they phone to cancel or when a switch request arrives. The retention team then has a few minutes and a wide discount matrix, and the size of the offer often depends on how firmly the customer pushes and which adviser they reach. Customers who never call stay on out-of-contract prices. Ofcom's 2026 pricing work found that customers who switch or renegotiate typically pay less than those who stay out of contract.
The market has changed around this. Since September 2024, One Touch Switch has let residential broadband and landline customers switch by contacting only their new provider, and millions have used it, so there is often no cancellation call to save. Since 17 January 2025, new contracts must set out any in-contract price rise in pounds and pence, not as an inflation-linked or percentage rise, and the major providers signed a voluntary Telecoms Consumer Charter in February 2026. Customers can compare more easily and leave more easily.
A churn model can help: contact people earlier, about the things that are actually bothering them, such as faults, a bill jump, or a plan that no longer fits. But the obvious way to use it is unfair. If offers go only to people predicted to leave, the customers least likely to shop around, who are often older, busier or in vulnerable circumstances, end up paying the most. That is the loyalty penalty Ofcom's rules and the Fairness for Customers commitments exist to stop, and the design has to rule it out from the start.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
The share of customers who leave in the three months around the end of their contract, against a holdout group handled as today.
ExampleDown by a fifth
How much more out-of-contract customers pay than the best available price for the same service.
ExampleFalling every quarter
End-of-contract notifications sent 10 to 40 days before the contract ends, and annual best-tariff notices to out-of-contract customers.
ExampleEvery customer, every time
Price paid and offers made for comparable services, by tenure, age band, channel and vulnerability flag.
ExampleNo group worse off without an approved reason
Customers flagged as vulnerable or eligible for a social tariff who were routed to a trained adviser and told about the right tariff.
ExampleEvery flagged customer
Complaints that the customer found it hard to leave, felt pressured, or was misled by an offer.
ExampleFalling, measured against last year
Cross-sector rules are on the UK regulation map.
Residential customers must be told 10 to 40 days before their contract ends, with their best tariffs, and out-of-contract customers must get a best-tariff notice every year. These are driven by contract dates in billing and run whether or not the AI is working. Since 17 January 2025, new contracts can't include inflation-linked or percentage price rises, and any in-contract rise must be stated upfront in pounds and pence, so every drafted offer shows prices that way and is checked by a rule before it goes out.
Leaving must be no harder than joining. The AI never sits in the switching or cancellation path. A switch request is never delayed so someone can make an offer, and advisers can't be prompted to keep pressing a customer who has said they want to leave. The commitments ask providers to give customers a fair deal that's right for their needs, and that is what the monitoring tests.
Signs of vulnerability, from the customer's records or what they tell an adviser, route them to a trained team. They are never targeted with pressure retention, and the model never uses vulnerability to decide who gets a worse offer. Where the provider has signed the voluntary Telecoms Consumer Charter (February 2026), eligible customers are told about social tariffs, and those in financial difficulty can move to a cheaper plan without penalty.
Churn scoring is profiling, so the privacy information says so plainly, and a legitimate interests assessment and DPIA cover it. Where scores are used for direct marketing, customers have an absolute right to object, and that objection is honoured everywhere. Offers are made within rules a person approved, but withholding a better price could be a significant decision, so since 5 February 2026 the safeguards apply: customers are told, can ask why, and can ask a person to look again.
Retention offers by email or text are direct marketing. They need consent, or the commercial soft opt-in: details collected during a sale, offers for the provider's own similar services, and a simple way to opt out at collection and in every message. The required end-of-contract notice isn't marketing, so keep it clear and separate from offers. Since 5 February 2026, PECR fines can reach UK GDPR levels, up to £17.5m or 4% of global turnover. Consent is checked at the moment of sending.
Since 6 April 2025, the CMA can enforce consumer law directly and fine up to 10% of global turnover. AI-drafted offers must not mislead, invent urgency, hide costs or pressure people. The CMA's guidance on AI agents (March 2026, updated October 2026) is clear that a business is responsible for what its AI says, as it is for its staff. Every template, and every draft built from it, is approved and checked.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Customer eventsFaults, bill changes, usage, contacts and contract dates, as they happen. | Amazon Kinesis Data Streams | Azure Event Hubs | Apache Kafka |
| Customer analytics storePseudonymised. Keep the link to customer identities in your own systems. | Amazon Redshift, with data in S3 | Microsoft Fabric | PostgreSQL, or your existing data warehouse |
| Feature storeThe same features for training and scoring, so the model sees in production what it learned from. | Amazon SageMaker AI Feature Store | Azure Machine Learning managed feature store | Feast on Kubernetes |
| Churn model and reasonsReasons matter more than the score: an adviser needs to know it's the faults, not the price. | Amazon SageMaker AI, with open-source SHAP | Azure Machine Learning, with open-source SHAP | MLflow and Kubeflow on Kubernetes, with open-source SHAP |
| Fairness monitoringSageMaker Clarify and Model Monitor closed to new customers on 30 July 2026, so on AWS the same tests run as your own scheduled jobs. | SageMaker AI scheduled jobs and pipelines, with open-source fairness libraries such as Fairlearn | Azure Machine Learning (model monitoring and the Responsible AI dashboard) | MLflow and Kubeflow on Kubernetes, with open-source fairness libraries such as Fairlearn |
| Offer rulesPlain, tested rules owned by pricing and approved by the offers committee. The model can choose among eligible offers; it can't create one or remove the floor. | AWS Lambda, or your existing offer or rules engine | Azure Functions, or your existing offer or rules engine | Containers on Kubernetes, or your existing offer or rules engine |
| Drafting messages and adviser notesDrafts from approved templates only. Prices, dates and terms are filled in from billing, never written by the model. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Agent runtimeThe original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore. | Amazon Bedrock AgentCore | Foundry Agent Service | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| Workflow | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| MessagingAmazon Pinpoint reached end of support on 30 October 2026. Check consent from the CRM at send time, not from a copy. | AWS End User Messaging (SMS), Amazon SES (email) | Azure Communication Services (SMS and email) | Your existing SMS and email gateway |
| Link to billing and CRMScores and drafts go to the CRM. Only billing applies a price, and only when the customer accepts. | API Gateway, plus your existing integration platform if you have one | API Management, plus your existing integration platform if you have one | Kong or your existing API gateway, plus your existing integration platform if you have one |
| Human reviewAdvisers see the summary and eligible offers where they already work. | In advisers' existing CRM and contact-centre desktop | In advisers' existing CRM and contact-centre desktop | In advisers' existing CRM and contact-centre desktop |
| GuardrailsBlocks urgency language, prices not in the approved offer, and anything that discourages leaving. | Bedrock Guardrails, plus rule checks on every draft | Foundry guardrails (Azure AI Content Safety), plus rule checks on every draft | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard), plus rule checks on every draft |
| Identity, keys, data protection | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog every score, offer shown, offer accepted, message sent and the consent basis used. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.
Stays on-prem
Billing, the CRM, the contact-centre platform and network data stay where they are, often on your own servers. The end-of-contract and best-tariff notices keep running from billing, so they never depend on the AI.
Uses cloud AI
Model training, scoring, fairness monitoring and drafting can run in AWS or Azure UK regions on pseudonymised data, or on your own GPUs with open-weight models for the sovereign option.
How they connect
Customer events stream over a private link (AWS Direct Connect or Azure ExpressRoute). Scores, reasons and drafts come back to the CRM through the API layer, and customer identities stay on your side.
If the link drops
Notifications, switching and cancellations carry on exactly as today. Advisers use the standard approved offer list, campaigns pause, and scoring catches up when the link returns.
What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.
Adviser and messaging assistant
Where a person signs offThe offers committee approves every offer rule and the floor. Advisers apply offers only when a customer accepts. Customers in vulnerable circumstances are handled by the trained team. The assistant's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
Billing stays the source of truth for contracts, prices and dates. The end-of-contract and annual best-tariff notices are generated from billing through your existing notification process, outside the AI, so they go out on time even if every model is switched off.
Scores and reasons are written to the CRM and the analytics store, never into billing. An offer only reaches billing when a customer accepts it, through the same API every other channel uses. Each offer carries an ID, so a retry never applies a discount twice.
The AI never sits in the switching path. One Touch Switch requests, and the switching information the customer must receive, go through your existing process unchanged.
Consent and objections live in the CRM's preference record and are checked at the moment of sending, not copied into a campaign list that goes stale. An objection to profiling for marketing stops scoring for marketing everywhere.
Map today's notices, retention offers and adviser discretion, measure the out-of-contract price gap, agree the floor and fairness measures, and check the consent records.
Gate · must be true to passBaseline agreed, notice process confirmed independent of the AI, DPIA and legitimate interests assessment started, legal and compliance involved.
Broadband customers within six months of contract end. Scores and adviser summaries in shadow first, then live for half the group, with a holdout handled as today.
Gate · must be true to passChurn and price gap better than the holdout, no customer group worse off, notices 100% on time, offers committee sign-off.
Live for broadband, with fairness monitoring, support and a rollback to the standard offer list.
Gate · must be true to passMonitoring, support model, rollback plan and adviser training in place.
Add mobile and bundles, then in-contract signals such as repeat faults, and later the business base under its own rules.
Gate · must be true to passEach new product and segment passes the same pilot gate.
| Activity | Customer director | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Pricing & offers committee | Legal, compliance & DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Consulted | Informed |
| Target architecture | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Informed | Consulted | Consulted |
| Offer rules, floor and fairness thresholds | Consulted | Consulted | Consulted | Informed | Informed | Accountable and Responsible | Consulted | Informed |
| DPIA, consent and marketing basis | Consulted | Consulted | Consulted | Informed | Informed | Consulted | Accountable and Responsible | Consulted |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Consulted | Consulted |
| Pilot with holdout group | Informed | Accountable | Responsible | Responsible | Responsible | Consulted | Consulted | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted | Consulted |
| Run, monitor and improve | Informed | Accountable | Consulted | Consulted | Responsible | Responsible | Consulted | Consulted |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: the event flow from billing and the CRM, the model and offer-rules platform, the link to your contact centre and messaging, and the gates. I'd report to your programme board and work closely with your pricing team and compliance.
Option B
If your partner or your own team already has an architect, I'd take on the data pipeline, the offer rules and the CRM and billing integration under their lead, and give them a second senior opinion on fairness monitoring, consent handling and keeping the regulated notices independent of the AI.
Goes wrong
Offers go only to people the model says will leave, so loyal customers pay more each year.
How the design avoids it
A floor of standard offers for everyone, offer rules approved by people, and monthly fairness monitoring by customer group.
Goes wrong
The end-of-contract notice is folded into a marketing campaign, and customers without consent never get it.
How the design avoids it
Notices run from billing on contract dates, separate from offers, and are measured at 100%.
Goes wrong
Retention becomes a barrier to leaving.
How the design avoids it
The AI is kept out of the switching and cancellation path, advisers accept 'I want to leave' first time, and guardrails block pressure language.
Goes wrong
A generated message gets a price or term wrong.
How the design avoids it
Templates approved in advance, prices and dates filled from billing, and a rule check on every draft before it goes.
Goes wrong
Marketing goes to people who opted out, from a stale campaign list.
How the design avoids it
Consent and objections checked from the CRM at the moment of sending.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Map which parts of each bundle fall under which regime, and make sure renewal and cancellation messages for add-ons meet both.
Keep pricing, notice and offer data structured by customer group, so you can show the effect of your offers on out-of-contract prices.
If you've signed it, build these into the offer rules and the vulnerability route now, and measure them like any other outcome.
Keep the offer rules, the reasons, the fairness results and the route to a person in a form you could show the ICO.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.
