Duty Desk Note: Privacy Checks Before Relay Traffic

A working request is not a completed privacy check

An API relay sits between your application and a model provider. Your application sends a request to the relay, which forwards it upstream and returns the response. This can simplify access to several model families, but it also adds another organization to your data path.

Before sending customer messages, internal documents, or source code through that path, establish who receives the data and what happens afterward. A successful test request proves connectivity. It does not establish privacy.

Ask which organizations receive the actual request

Start with a simple question: “Where does my prompt go after it reaches your service?” Ask for the relay operator, the upstream provider, and any other processors that can receive request content. These might include services used for monitoring or support.

• Ask whether the selected model always uses the same upstream route or whether routing can change.
• Ask whether fallback behavior can send a failed request to another provider or region.
• Ask whether attachments, tool results, and generated responses follow the same handling rules as prompt text.

For example, selecting Gemini 2.5 Flash identifies a model choice, but it does not by itself explain the relay’s routing, storage, or contractual arrangements. Request those details separately.

Separate request processing from stored copies

A service needs to process your request to answer it. Keeping a copy afterward is a different activity. Ask whether the relay stores full prompts and responses, selected excerpts, or only operational information such as timestamps, model names, and token counts.

Operational information can still be sensitive. Account identifiers, IP addresses, and request timing may reveal who used a system and when. Ask about those records too.

• Ask for retention periods for request content, operational logs, abuse-monitoring records, and backups.
• Ask whether logging can be disabled and exactly which records that setting excludes.
• Ask whether errors, retries, and support investigations create additional copies with different retention rules.

“We do not store prompts” needs a defined scope. It might describe routine application logs without covering temporary caches, upstream storage, or incident records.

Check training rules at both sides of the relay

Ask whether your inputs or outputs can be used to train or improve models, evaluate systems, or support human review. Get separate answers for the relay and its upstream providers.

Do not assume that a provider’s consumer-chat policy applies to API traffic, or that every upstream account has identical settings. Ask which terms cover the actual route your requests use. If an opt-out is available, confirm whether it applies automatically, requires configuration, or has exceptions.

Locate the people and systems that can read content

Encryption in transit protects data while it travels between systems. It does not necessarily prevent the relay from reading content while processing it. Ask where encryption ends, whether stored data is encrypted, and who can access readable requests.

• Ask whether staff access requires a documented reason, approval, and an audit record.
• Ask where processing and storage occur, including any cross-border transfers.
• Ask how API keys are protected, revoked, and kept out of logs.
• Ask how customers are notified if an incident affects their data.

If your organization needs a data processing agreement or a particular processing location, verify that requirement before deployment. A privacy policy alone may not provide the commitments your organization needs.

Close the shift with evidence and a low-risk trial

Ask what deletion covers, how to request it, and what remains in backups or legally required records. Keep the answers with the relevant policy versions and contract terms. Treat undocumented or ambiguous answers as unresolved rather than filling the gaps with assumptions.

Relay365 is worth considering when pay-as-you-go access to multiple model families fits your workflow. It works with Claude Code, Codex, Gemini CLI, and OpenAI SDK clients; that convenience should be evaluated separately from its privacy arrangements. If it fits your needs, create a Relay365 account and begin with synthetic, non-sensitive requests while you verify the answers above.

Before production traffic, remove unnecessary personal data, use dedicated credentials, and check your own application logs. Privacy controls at the relay cannot undo a sensitive prompt copied into your debugging output.