Customer remediation for AI errors

What it does

It puts right the harm an AI error has caused. Because a model applies the same fault to every similar case, one discovered error implies others. The control establishes the scope, finds the customers affected, corrects the decision or record, compensates for loss and tells them what happened. The business owner of the system is accountable; operations and compliance run the exercise. It operates after an error is identified, by incident, complaint, appeal or monitoring.

How to implement it

• Define what triggers a remediation review, including a single upheld complaint that points to a systematic fault.
• Use decision records to identify every customer affected, over the whole period.
• Decide the remedy by reference to where the customer would otherwise have been.
• Contact customers proactively, in plain language.
• Fix the cause before reprocessing cases through the same system.
• Record the exercise and report it to governance, and to the regulator where required.

Evidence it operates

Design: a remediation procedure with triggers and governance. Operation: records of exercises with scope, method, numbers and amounts; customer communications; evidence that the root cause was fixed; reporting to the board or regulator.

Where it fits

After detection, alongside AI incident response and rollback. It is a primary control for inadequate contestability and redress, and supports wrongful automated refusal and poor customer outcomes from AI-driven journeys. It depends on AI record-keeping and audit trail to find who was affected.

Limits

It acts after harm. Some losses cannot be made good, such as a missed house purchase. Finding affected customers depends on records that may not have been kept. Remediation is expensive, which creates pressure to define the scope narrowly; the decision on scope therefore needs independent challenge.