Stateless Isn't the Same as Zero Retention
These two terms get used interchangeably, and for anyone evaluating a platform, that's a mistake worth catching early.
Stateless Processing means the application doesn't hold client state between requests.
Ephemeral Processing means client data is available only for the duration necessary to perform the computation, with controls preventing it from becoming persistent in storage, logs, caches, or diagnostic systems.
Zero-Retention Architecture means the entire chain is engineered and contractually configured so client payloads are not intentionally retained past the transaction that used them. That chain runs from the application and infrastructure through API gateways, logging, monitoring, backups, subprocessors, and the AI provider itself. Zero retention is an architectural and contractual property of that defined service boundary, not a claim that customer data is physically unrecoverable from every underlying infrastructure component at every instant.
A system can be architecturally stateless and still leak sensitive data into an error log or a caching layer without anyone intending it to. Zero retention isn't a property of the application alone. It's a property of the entire path the data travels, and it has to be verified as one.
What zero retention actually promises, and what it doesn't
In this paper, zero retention means zero persistent retention past the defined transaction. It applies to customer content specifically: uploaded documents, extracted text, prompts, responses, embeddings, intermediate parsing output. It does not automatically extend to operational metadata needed for security, billing, or abuse prevention, which deserves its own separate retention policy.
| Zero retention means... | Zero retention does not mean... |
|---|---|
| No persistent customer document store | No data is ever processed |
| No customer content in application logs | No security controls are required |
| No persistent vector index of customer data | No compliance obligations exist |
| Customer content excluded from model training | The AI provider can safely retain prompts for secondary uses |
| Ephemeral processing of payloads | No temporary data exists during processing |
| Derived outputs are subject to the same retention controls as inputs | Customer systems don't retain downloaded results |
The right column is where vendor claims usually go soft. The same discipline applies to reconciliation engines: a system that's precise about what it doesn't claim is more trustworthy than one that claims everything. Next: the actual architecture that makes the right column real.
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.