The Hardest Part of Zero Retention: Failure Paths
Zero retention has to apply to what comes out, not just what goes in. Extract one liability clause from a fifty-page contract and that clause is often just as sensitive as the document it came from. The generated report, the summary, the reconciliation exception, all of it gets the same disposal treatment as the source material, immediately after it's returned.
Where this actually gets hard
The real test of this architecture is what happens when something breaks halfway through. Most enterprise monitoring tools are built to grab as much context as possible the moment something fails, which is exactly the wrong instinct for a system that's promised not to keep anything.
A system that deletes customer data on success and captures the same data in an exception trace on failure hasn't actually built zero retention. It's built zero retention for the happy path only, and the happy path was never the concern. To hold the boundary under failure, the architecture has to enforce:
- Failed workloads must not leave customer content behind in temp files, writable container layers, attached volumes, or node storage.
- APM and tracing systems must not capture prompts or document text in telemetry.
- API gateways must not log request bodies on 5xx errors.
- Third-party OCR or extraction vendors must not write failed jobs to a dead-letter queue for someone to review by hand later.
None of this is exotic engineering. It's the unglamorous work of checking every place data could accidentally end up, the same discipline behind catching accidental persistence anywhere in a pipeline, just applied specifically to the failure path instead of the normal one. Next: how any of this gets proven to someone who wasn't in the room.
Adapted from the full white paper on engineering zero-retention AI systems. Download the PDF.
Rosemarie Withee has spent thirteen years helping operations teams get real work out of their software, first Microsoft 365, now AI. She’s written six books for Wiley and builds AI products at Portal Integrators.