Use AI to route banking complaints without deciding the outcome

A practical intake design that finds the issue, preserves the customer's words and sends the case to an accountable reviewer.

A customer sends three paragraphs about a card payment. One sentence mentions a duplicate charge, another describes difficulty reaching support, and the last expresses distress about rent. A complaint system that selects only a category can lose the reason the customer needs help. AI can assist with intake if its proposed structure remains visible and correctable.

Banks describe employee generative AI platforms and adviser assistants. The Bank of England and FCA survey gives dated evidence of broader financial AI use. Those examples show institutional adoption, not proof of automated complaint-resolution capability. [Employee AI](https://www.jpmorganchase.com/about/technology/blog/llmsuite-ab-award), [adviser AI](https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch), [financial-services survey](https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024).

For an original exercise, Oak Branch Bank uses fictional complaint messages to test a sorting assistant. Its first task is to identify all issues raised, not to tell the customer whether their claim is valid. It proposes payment concern, service access concern and possible need for additional support. The intake worker reads the original message before confirming any labels. The customer's text stays alongside the structured fields.

The assistant drafts a neutral account: the customer reports two charges and asks for a review. It does not write that the bank charged twice unless records establish that fact. This difference matters when a pending authorisation, settled payment and refund appear together. An intake system should preserve what the person reported while leaving investigation findings to the people with evidence and authority.

Give each issue a next step drawn from an approved process. A payment concern may go to the card team; a support-access complaint may remain with the complaints handler. The customer should not have to repeat the story to separate teams. Oak links both tasks to one case and appoints a coordinator. AI can prepare the handoff summary, while a staff member checks that sensitive context is only shared with people who need it.

A useful test includes indirect language. A customer may never write the word complaint. They might say they have tried twice and cannot get a clear answer. Another message might express anger but concern a service provided by a different organisation. The assistant should identify the uncertainty and offer a review route, rather than automatically close the case as outside scope. Escalation is an outcome the system must support.

Do not infer a person's health, disability or financial situation from tone. If the customer explicitly requests communication support, the intake worker can record the request according to the bank's approved policy. Otherwise, use a respectful invitation to tell staff about any assistance needed. Predictions about vulnerability require particular care and should never become invisible labels that change how a person is treated.

The test set includes a translated message, a spelling-heavy message and a concise one-line complaint. Reviewers compare issue recall across these cases. An assistant that extracts three issues from polished prose but only one from imperfect writing may create unequal service. Also test a message containing a malicious instruction to close the case or reveal another customer's details. Customer content cannot grant new permissions.

A draft acknowledgement can explain that the message was received, what will be reviewed and how the person can provide more information. It should use verified dates and the organisation's actual process. The model must not invent a resolution deadline or promise a refund. If case creation fails, preserve the safe draft and clearly offer a retry or another support channel; an acknowledgement is not evidence that the case was successfully recorded.

Measure missed issues, incorrect routing, unnecessary personal details copied and rework after the receiving team opens the case. Review a sample of cases marked straightforward as well as escalated ones. Speed matters, but a rapid handoff that omits the duplicate payment forces the customer to start again. The evaluation should follow the case far enough to reveal that cost.

These public institutional examples are not off-the-shelf complaint products. No current subscription, production integration or regional entitlement is verified. Oak's pilot would need a controlled dataset, staff reviewers and checks in its existing case system. Financial or legal decisions remain with authorised people; this article provides an educational process example.

AI is useful here when it helps the bank hear the complete complaint and organise a better response. The goal is a faithful intake record with an accountable owner. A smooth automatic answer is a poor substitute if it closes the conversation before anyone has understood the problem.