Energy debt in GB has not peaked. It has plateaued at a level that twelve months ago would have looked extreme. The Debt Relief Scheme, launching in early 2026 with around £500m of support, is engineered to take the worst of the back-book pressure off. It will work. It will also expose, for the first time at scale, how AI is making affordability and prioritisation decisions inside suppliers.
Where AI is already in the chain.
Eligibility scoring. Outbound contact prioritisation. Voice handlers on hardship lines. Affordability classifications inside collections journeys. Vulnerable-customer identification flags. None of these are exotic. All of them are in production at most large suppliers today. Few of them have a current AIIA. Even fewer have a documented mitigation for the most predictable failure mode: that the AI quietly disadvantages exactly the customers the Debt Relief Scheme is designed to help.
What goes wrong without guardrails.
Customers in temporary hardship classified as long-term high-risk because the model has not seen recovery data. Customers in language minorities deprioritised because the contact model favours channels they do not use. Self-disconnection signals missed because the smart-meter feed is treated as a billing input, not a vulnerability input. Each of these is a regulator-visible outcome, and each is preventable with an AIIA that is younger than the model.
What "with guardrails" looks like.
An AIIA per use case, refreshed when the model or the data changes, naming the failure modes and the mitigations. Human review at a defined threshold for any decision that materially affects access to support. A self-disconnection signals review on the last twelve months of smart-meter data, with the gap documented. Named accountability for each AI system inside collections and hardship. A weekly frontline-led signal review. These are not aspirational. They are the operating norm in a small number of suppliers and absent in most.
Why this matters now.
The Debt Relief Scheme will be reviewed in operation, and the review will ask how customers were identified, how decisions were evidenced, and how vulnerable cohorts were protected. Suppliers without assurance will not lose the right to participate; they will lose the right to be trusted with the next iteration. That is a slower, more permanent loss.