An AI interaction can create more records than the text a user sees: prompts, attached context, retrieved documents, system instructions, outputs, feedback, safety events and technical logs may be stored in different places. A bank needs explicit retention rules because keeping everything can increase privacy and security exposure, while deleting too quickly can remove evidence needed for customer service, compliance, investigation or model oversight.
The bank first maps what the AI service actually records
An inventory identifies prompt text, uploaded files, retrieved data, model responses, user edits, ratings, identifiers and operational logs. It also shows where each record is created, whether it enters an analytics or training process, and which internal teams or outside providers can access a copy.
A single screen can hide several data paths. A prompt may appear transient in the user interface while copies remain in monitoring tools, backups, security records or a provider environment, so the retention decision must follow the complete data flow rather than the most visible application setting.
Purpose and obligations determine what evidence is needed
The approved use case establishes why a record might be retained: completing a customer request, supporting quality review, investigating an incident, demonstrating control performance or meeting an applicable recordkeeping obligation. The bank involves records management, legal, compliance, privacy, security, business and AI governance functions according to the use and data involved.
Different records can require different periods. A customer communication, an internal brainstorming prompt and evidence supporting a regulated decision should not inherit one default simply because the same model processed them. Litigation holds, investigations and other preservation requirements can also pause ordinary disposal for identified material.
Minimization and classification narrow the retained data
The bank limits collection before setting a clock. Restricted inputs can be blocked or redacted, retrieved context can be scoped to the task, and logs can preserve necessary control evidence without copying every piece of sensitive content into every monitoring system.
Classification considers customer information, confidential business data, credentials, source code and other sensitive material in both prompts and outputs. Retention periods then reflect the stated purpose, data sensitivity, risk, operational need and applicable requirements, with owners and approved exceptions documented.
Retained records need controls throughout their life cycle
Access is limited by role and environment, and encryption, logging, monitoring and segregation are matched to the information’s sensitivity. Search, export and administrator functions receive particular attention because a protected archive can still create risk if broad tools let users retrieve or combine records without a valid purpose.
Retention also affects model evaluation and incident response. The bank keeps enough trustworthy metadata to connect an event with the approved model, configuration, data sources and user action, while avoiding the assumption that a full prompt history is always the only way to produce an auditable record.
Disposal must reach providers, replicas and derived copies
Contracts and technical settings address whether a provider stores interaction data, uses it to improve services, places it in backups or subprocessors and can delete it at the end of the agreed period. The bank monitors those commitments because outsourcing the platform does not outsource accountability for the bank’s data risk.
A controlled disposal process covers primary stores, indexes, exports and other reachable copies, records completion and tests whether records are no longer ordinarily available. Where a backup cannot be changed immediately, access and eventual expiration are defined rather than calling the data deleted while an active copy remains recoverable for routine use.
Read the primary material
Banking Explained prioritizes regulators, official publications and first-party announcements.