What happened
OpenAI announced OpenAI Private Intelligence on October 30, 2026. Its product page is titled "Bring advanced intelligence to your most sensitive work" and groups four areas of protection. Access restrictions: OpenAI says it sets strict limits on its own access to business content, and provides Zero Data Retention (ZDR) and Private Inference "for eligible workflows." Data ownership: OpenAI says you own your business data, including inputs and outputs. Agent actions: OpenAI says you decide what agents can access, which actions they can take, and when they need your approval. Data security: OpenAI describes encryption, strict access controls, and ongoing threat detection. The page also lists Private Inference, which it marks as a preview. OpenAI describes it as confidential computing with strict, verifiable controls. For compliance, the page mentions SOC 2 Type 2 audits, ISO certifications, HIPAA support through Business Associate Agreements for eligible services, and FedRAMP support for designated government offerings. The page points to a form to register interest. It does not publish pricing or a general availability date for the new bundle.
Why it matters for business owners
Most owners want to know one thing before putting real customer, contract, or financial data into an AI tool: who can see it, and for how long. A vendor putting its answers in one place helps. It also tells you where the market is going. Data handling is becoming something vendors sell and compete on, not a footnote in the terms. That gives buyers room to ask better questions and to compare vendors on answers, not on logos. The page matters for a second reason. It names agent controls as part of privacy: what an agent may read, which actions it may take, and when it needs a person's approval. For a business, a lot of data risk comes from what an agent can do, not only from where a prompt is stored.
What owners should not misunderstand
"Private" is not one thing. Look at the qualifiers in OpenAI's own wording: ZDR and Private Inference are for "eligible workflows," HIPAA support is for "eligible services," FedRAMP support is for "designated government offerings," and Private Inference is a preview. Each of those is a condition you have to meet or confirm. Zero Data Retention is not automatic. OpenAI's data controls documentation says these controls are subject to prior approval by OpenAI and acceptance of additional requirements. It also says some endpoints and features still store application state even when ZDR is on. Whether your tool or workflow qualifies depends on how you use it. Zero retention can still involve a copy you hold. Under ZDR with Private Safety Processing, OpenAI's developer guide says selected prompts and responses are encrypted and written to storage the customer controls, with a 30 day time to live, so an automated safety review can examine them. OpenAI says it keeps an index and a storage reference, not a copy of the content, and that its personnel cannot read the records. But you must set up that storage, its permissions, and ideally your own encryption keys. That is real work, and it is yours. The documents and the white paper describe first-party OpenAI traffic. OpenAI's white paper says Amazon Bedrock and other third-party distribution arrangements are outside its scope. If your staff reach OpenAI models through another vendor's product, these terms may not be the ones that apply to you. A vendor's description is not an audit of your setup. Compliance items on the page describe the vendor. Whether your use of the tool is compliant depends on your data, your contract, and how people actually use it. Atlacis does not give legal advice, so involve counsel for regulated data. Finally, a protection that exists is not a protection your team uses. Many staff use consumer or personal accounts at work. Better enterprise terms do nothing for data pasted into a tool that is outside them.
The operational lesson
Treat a privacy label as a list of claims, and write down which ones you have confirmed for your own account. A useful split is between what the vendor does and what you must do. The vendor can limit its own access, encrypt records, and describe its audits. You still decide which data goes in, who has an approved account, which agent actions need a person's sign-off, and who owns the storage and keys when a feature requires them. This also fits what private AI means in practice. It is a range of choices about where data goes and who can reach it, from a vendor's stricter plan to a model run in your own environment. A stronger vendor plan can be enough for some data and not for other data. The decision should follow the data, not the product name.
What a serious business should do next
List the AI tools your team actually uses, including personal accounts, and note which plan and contract each sits under. Sort the data into tiers. Public and low-risk material can go almost anywhere. Customer records, contracts, financials, and regulated data need written answers first. Ask your vendor, in writing, for each protection you plan to rely on: does it apply to my account, my plan, and my specific workflow, is it generally available or in preview, what do I have to configure, and what happens if the conditions change? If a feature needs customer-controlled storage or keys, decide who owns that before you turn it on, and who checks it. A setup nobody maintains is worse than a simpler one that someone does. Set the agent rules separately. Write down what an agent may read, what it may change, and which actions always need a person, such as sending money, contacting customers, or editing records. Do not move your most sensitive workflow because of a launch page. Test the process on lower-risk data first, and revisit once the preview items are generally available and your own answers are in writing.
The Atlacis view
Atlacis helps owners slow down, understand the workflow, map the risk, and choose the right tool or system before spending. Vendor privacy programs are one input. Sometimes the right answer is a stricter plan with a major vendor. Sometimes it is a different setup for a small share of very sensitive work. Often the first fix is simply knowing which tools your people use. If you are deciding which data can go to which AI tool, a short conversation can help you sort that out before you sign anything.
The short version
- On October 30, 2026, OpenAI introduced Private Intelligence, a business page that bundles its data protections under one name.
- Zero Data Retention and Private Inference are described "for eligible workflows," and Private Inference is marked as a preview.
- OpenAI's documentation says ZDR is subject to prior approval and that some endpoints may still store application state.
- Under ZDR with Private Safety Processing, selected records sit in storage the customer controls for 30 days, so the customer has setup work to do.
- Confirm each protection for your own account, plan, and workflow in writing before relying on it.
- Staff using personal accounts sit outside any enterprise protection.