How to Actually Prove Zero Retention
A zero-retention claim only means something once it's backed by controls someone can actually check. That's:
- Logging Controls: Structured logging with strict payload suppression; audit trails that record metadata, timestamps, usage, success or failure, never customer content.
- Infrastructure Controls: Ephemeral compute, workload isolation, encryption in transit and at rest where applicable, restricted IAM roles, confidential computing or memory protection where it's warranted.
- Provider Controls: Signed data processing agreements with every AI model provider that explicitly rule out training and cap retention at the transaction window, backed by real technical configuration, not just a policy page.
- Subprocessor Controls: Every service that touches customer content (OCR, extraction, inference, observability providers) must either meet the same retention bar or be technically prevented from receiving that content in the first place.
- Validation: Automated dependency, code, and container scanning in the deployment pipeline, plus penetration testing specifically aimed at accidental state persistence and side-channel leakage, not just the usual perimeter tests.
What a real audit actually asks for
Treat zero retention as a testable property of the system, not a line in a privacy policy. For a security or procurement team doing real diligence, that means being able to produce:
- Data-flow documentation and architecture diagrams that trace one payload's actual lifecycle.
- Documented logging, monitoring, and tracing configuration.
- Signed provider agreements alongside evidence the retention and training controls are actually configured, not just promised.
- Evidence from testing designed specifically to validate deletion behavior and probe failure-path leakage.
- A maintained inventory of every subprocessor and exactly what data-handling constraint applies to each one.
This is the same evidentiary bar behind the audit trails we build into reconciliation engines: a system that can't produce a real answer to "why did you do that" isn't actually explainable, no matter how accurate it is in aggregate. Last in this series: what all of this is actually worth.
Adapted from the full white paper on engineering zero-retention AI systems. Download the PDF.