Skip to content

OpenAI background mode, ZDR, store=true, and retention limits

OpenAI background mode, ZDR, store=true, and retention limits

Section titled “OpenAI background mode, ZDR, store=true, and retention limits”

Background mode is attractive because it lets long-running work continue after the initial request. For enterprise teams, the harder question is data control. A background response needs retrievable state. That makes it different from a stateless synchronous call, and it changes which workloads should be allowed onto that lane.

This is not only a compliance footnote. It is an architecture boundary.

Current official signals checked April 24, 2026

Section titled “Current official signals checked April 24, 2026”
Official sourceCurrent signalWhy it matters
OpenAI background mode guideBackground mode stores response data for roughly 10 minutes to enable polling and is not Zero Data Retention compatibleTeams with strict ZDR expectations should not treat background mode as a drop-in replacement for stateless requests
OpenAI background mode guideBackground sampling requires store=true; stateless requests are rejectedThe long-running job model depends on stored response state
OpenAI data controls documentationBackground mode data storage for polling is called out separately from normal response behaviorData-control review should happen before moving sensitive jobs into background mode

Use background mode only for workloads where stored response state is acceptable under your data-control policy.

If a workload requires strict Zero Data Retention behavior, background mode may be the wrong runtime lane even if the task is technically long-running.

That creates a three-part decision:

  1. Does the job need long-running execution?
  2. Does the product need polling or later retrieval?
  3. Is stored response state acceptable for this data class?

All three need to be true.

A background job must be retrievable after the initial request. The product may need to poll status, reconnect a UI, recover from a dropped stream, cancel a response, or fetch the final output. Those behaviors require provider-side state for the duration of the background workflow.

That is the tradeoff:

  • background mode improves reliability for long-running tasks;
  • stored state enables polling and recovery;
  • strict statelessness becomes harder or unavailable for that lane.

The architecture should make that tradeoff visible.

Background mode can be a good fit for:

  • long research jobs using non-sensitive sources;
  • document transformation where retention is allowed;
  • internal report drafting;
  • coding-agent analysis where repository data is approved for this lane;
  • review-gated generation where the output is not immediately published.

The common pattern is that the organization accepts temporary stored state in exchange for a healthier long-running workflow.

Be careful with:

  • regulated customer data;
  • sensitive internal investigations;
  • confidential legal or financial material;
  • highly restricted support tickets;
  • workloads contractually requiring ZDR or equivalent handling;
  • data that should not be retrievable outside a live request.

For these workloads, teams may need a different architecture: synchronous constrained calls, internal processing, a separate provider configuration, or a product-owned workflow that avoids background storage assumptions.

Do not make data-control decisions one request at a time. Attach them to job types.

Job typeBackground mode default
Public-source researchusually allowed
Internal low-sensitivity draftingoften allowed
Customer-support analysisdepends on data class
Legal or regulated document reviewusually needs security review
High-sensitivity customer dataavoid unless policy explicitly permits

This keeps engineers from discovering the data-control problem during incident review.

Before enabling background mode for a workflow, answer:

  1. What data class does the job process?
  2. Is temporary stored response state acceptable?
  3. Does the workflow require polling, cancellation, or later retrieval?
  4. Is the job compatible with the organization’s ZDR expectations?
  5. Should this job type be blocked from background mode by policy?
  6. Does the audit trail record why background execution was allowed?
  7. Is there a fallback lane for sensitive work?