OpenAI spotlights Private Intelligence, with Private Inference still in API preview

OpenAI is putting enterprise data controls at the center of its latest business pitch. In an October 30 announcement, the company promoted Private Intelligence as a way to use its advanced models for sensitive work. The key deployment distinction is easy to miss: its Private Inference capability is still an API preview, while other protections have workflow-specific eligibility.

The announcement brings several privacy measures under one name. It should not be read as a new model release, an automatic upgrade for every account, or a guarantee that every connected application handles information in the same way.

What the preview covers

OpenAI’s Private Intelligence overview describes Private Inference as confidential computing combined with controls intended to protect prompts and responses during model processing. The page labels it “API Preview.” That is narrower than general availability across ChatGPT, Codex and every API feature.

The same overview presents Zero Data Retention for eligible API workflows, alongside data ownership, access restrictions and controls over agent permissions. Those are different layers of protection: retention concerns what remains after processing, while Private Inference addresses how information is protected during processing.

OpenAI’s interest-registration page separately says Zero Data Retention with Private Safety Processing supports automated safety reviews without personnel seeing the underlying content. It invites organizations to register interest; submitting that form is not evidence of immediate access.

Existing commitments still have boundaries

The company’s enterprise privacy documentation already states that business data is not used for model training by default. It also describes product-specific retention and human-access rules. For API use, the documentation notes that inputs and outputs may be retained for up to 30 days, subject to feature-specific exceptions, unless an eligible zero-retention arrangement applies.

That context matters when evaluating the new messaging. A business should verify its actual service, endpoint and agreement rather than assume a general privacy statement overrides every existing setting. The same documentation says workspace administrators control which connected apps are enabled and that users authenticate with those apps. A model provider’s protections alone do not settle what happens after an agent sends information to another service.

What teams should check before deployment

For a pilot, the useful questions are concrete: Which endpoints and features qualify? Which records or safety signals remain? Who controls access to them? How can the organization verify the promised processing protections? And which agent actions require human approval?

OpenAI’s security overview lists audits, certifications and administrative controls, but readers should check their scope for the particular product and workflow. A certification listing does not by itself establish that a proposed deployment meets every requirement.

The opportunity is a clearer separation between using capable AI and exposing business information. The immediate decision remains a preview evaluation: establish eligibility, inspect the relevant technical and contractual details, and test with appropriate non-sensitive data before expanding the workload. PromptsZone has not independently audited Private Inference or tested access to the preview.

Illustration: original AI-generated conceptual artwork for PromptsZone. It represents protected processing and access controls, not OpenAI’s actual infrastructure or interface.