When a compliance tool sends a prompt to a large language model, the prompt can contain customer names, transactions and case notes. Zero data retention addresses what the model provider keeps afterwards. Under ZDR, the provider processes the request and doesn't store the inputs or outputs, which limits exposure if the provider is breached and keeps the data out of later training.

How it works

ZDR is a contract and configuration setting, not a property of the model. Cloud AI services publish default data handling and the exceptions to it. For example, Azure's data privacy documentation for OpenAI models states that prompts and completions aren't used to train foundation models and describes abuse monitoring that approved customers can modify. Amazon Bedrock states that model providers don't have access to customer prompts and completions. When evaluating a vendor, ask three separate questions:

• Training. Can any model, the vendor's or a provider's, be trained or fine-tuned on your data?

• Provider retention. Does the model provider keep prompts or outputs, and for how long?

• Platform records. What does the compliance platform itself keep for audit, and under whose retention policy?

The last point matters for regulated teams. Examiners expect to see how an AI recommendation was produced, so a compliance platform should keep an audit record of AI actions under the institution's own retention rules even when the model provider keeps nothing.

Why banks and regulators care

Data minimization and storage limitation are core principles of the GDPR, and AI providers are third parties that banks must assess under US interagency third-party guidance. The NIST AI Risk Management Framework treats data privacy as a characteristic of trustworthy AI. Retention and training terms are usually among the first questions in a bank's AI vendor review, alongside where the model runs and whether prompts can reach shared, multi-tenant endpoints.

How Corsa supports it

Corsa can run AI on privately hosted models that don't share infrastructure with public AI services. Customer data is never used to train, fine-tune or improve any model, which is contractually guaranteed, and sub-processor agreements prohibit training, evaluation or improvement on customer data. Inside the platform, AI interaction logs follow the customer's configured retention for audit, guardrail activations are logged separately, and deleted data is purged from AI logs and caches. HYOK-encrypted fields never reach AI at all. See Private AI for compliance.