OpenAI enterprise data privacy terms. The nine clauses your contract needs.

OpenAI's defaults are stronger than most procurement teams assume, because business data is not used for training. The gap is everything the defaults do not cover, from retention to exit deletion, and all of it is set by contract.

OpenAI does not train its models on ChatGPT Enterprise, ChatGPT Business or API data by default, and it says so in writing. For most privacy reviews that is the first question, and it is the easiest one to close.

The harder work is everything the default leaves open: how long content is kept, when you hear about a new subprocessor or a breach, where data is stored and what happens to it when you leave. Those points are set by the documents you sign. A policy page can change with a web edit, and an order form term cannot.

What happens to OpenAI enterprise data by default?

By default, OpenAI does not use enterprise or API data to train its models. The commitment appears in the published privacy commitments, and the business terms repeat it: OpenAI will not use customer content to develop or improve its services unless the customer explicitly agrees.

Training and retention rules differ by channel. Consumer tiers train on conversations unless the user opts out. On the API, OpenAI may keep inputs and outputs for up to 30 days for abuse monitoring, and zero data retention is available for eligible endpoints only after approval.

Zero data retention has limits that engineering teams often miss. Even with it switched on, the Assistants API, the Conversations API, vector stores, file uploads and fine tuning jobs keep what you give them until you delete it. A narrower option, Modified Abuse Monitoring, keeps your content out of the abuse logs; both need approval from OpenAI's sales team.

Why is a published policy weaker than a contract term?

OpenAI controls its own pages and can edit them. Even the business terms carry their own update clause, so the commitments you rely on belong in the order form.

• Material changes. At least 30 days notice, and continued use after the effective date counts as acceptance.
• Changes to comply with law. Only as much notice as OpenAI can reasonably give.
• Your remedy. Terminate if you reject the update, which is rarely practical halfway through a rollout.

The 2025 litigation hold is the clearest recent example. A court preservation order in the New York Times copyright case required OpenAI to keep content from consumer ChatGPT, ChatGPT Team and standard API use that it would otherwise have deleted, until the obligation ended on September 26, 2025.

• Outside the order. ChatGPT Enterprise, ChatGPT Edu and API traffic under zero data retention.
• What decided it. The channel each customer bought and the retention terms it had signed, since no contract overrides a court.

Which OpenAI documents govern enterprise data privacy?

Three documents govern, and a privacy review that reads only one of them is incomplete. The business terms set ownership and liability, the addendum sets processor obligations, and the trust portal holds the evidence that the controls exist.

• Business terms. The business terms govern API and enterprise use. Read data ownership and indemnity first: you keep ownership of your input, OpenAI assigns you its rights in the output, and each party's liability is capped at the fees paid in the 12 months before the event that gave rise to the liability.
• The addendum. The DPA carries the processor commitments. It must be executed; a link in the online terms is not the same as a signed copy in your contract file.
• Trust portal. Audit reports and subprocessor lists live on the trust portal. Pull them before signature, while what they say can still change what you sign.

What should you pull from the trust portal before signature?

Download the current SOC 2 Type 2 report and read two parts closely: the exceptions the auditor found, and the complementary user entity controls, which are the controls OpenAI expects you to run yourself. Then save a dated copy of the subprocessor list.

File both next to the signed DPA. Check the answers to your security questionnaire against the report, and every future subprocessor notice against the saved list.

Which nine privacy clauses should an OpenAI enterprise agreement carry?

Nine clauses separate a negotiated agreement from a signed marketing page. Each one has a concrete position you can write into a redline, and the table after the lists shows what OpenAI's own documents already say on each.

Clauses one to five: data handling

• Training exclusion in the order form. Restate the no training default as a contract term so it survives policy edits.
• Retention window. Name the abuse monitoring window and require written notice if it changes.
• Zero data retention. If your workload qualifies, get approval in writing and reference it in the agreement, with the endpoints it covers.
• Subprocessor notice. 30 days advance notice before a new subprocessor handles your data, with a right to object.
• Breach notice. 72 hours, in place of the open "without undue delay" wording in the published addendum.

Clauses six to nine: control and exit

• Audit rights. Annual delivery of the audit report plus a right to send a written security questionnaire.
• Data residency. If you need regional processing, name the region. Silence means provider discretion.
• Deletion on exit. Certified deletion within 30 days of termination, including backups on a defined cycle.
• Indemnity scope. Confirm the output indemnity covers your usage tier and is not voided by your own fine tuning data.

What wording should the order form carry?

Put your specific commitments in the order form and make it the controlling document. Precedence wording does more work than any single clause, because it stops a later web edit from overriding what you negotiated.

How do OpenAI privacy terms differ between ChatGPT and the API?

The product you buy decides which defaults apply before any negotiation starts. Spend then decides how much of the rest you can change, which the next section covers.

Which OpenAI channel are you actually buying?

• ChatGPT Enterprise. Admins set the retention period, and data residency is offered for new workspaces.
• ChatGPT Business. Sold on standard terms to smaller teams, with no data residency option. Use it for personal data only once the DPA is signed and you accept workspace level controls.
• API platform. Zero data retention and Modified Abuse Monitoring need approval. Regional storage needs approval for advanced data controls and a new project, and OpenAI charges extra for it on eligible models.
• Azure OpenAI. The same models under Microsoft's product terms and DPA, which changes the answer to every clause above. Our Azure OpenAI and direct OpenAI comparison sets the two side by side.

If you run OpenAI and Anthropic side by side, the same clause list applies to both. Our notes on negotiating with Anthropic and our GenAI licensing practice cover the differences, and episode three of the negotiation series walks through the signature stage.

How do you negotiate OpenAI privacy clauses without stalling the deal?

Tie every ask to something OpenAI has already published, because the fastest redlines are the ones the vendor has conceded elsewhere. The privacy page, the addendum and the trust portal give you a published basis for seven of the nine clauses.

Sequence matters as much as wording. Send the addendum and your security questionnaire in week one, in parallel with the commercial terms. Teams that treat privacy as a closing formality lose two to four weeks at signature.

What will the OpenAI account team push back on?

• Zero retention eligibility. It is granted through a use case review, and negotiating pressure does not change the outcome, so submit the use case early.
• Custom breach windows. Smaller deals get the standard terms, and below roughly $250,000 a year you should expect limited movement.
• Audit beyond the standard report. On site audits are reserved for the largest commitments, so the report plus questionnaire route is the realistic position.

What will OpenAI say, and how should you answer?

• "Our privacy commitments already cover training." Agree, and ask for that sentence in the order form. If it is already policy, writing it into the contract costs OpenAI nothing.
• "We do not redline the DPA." Leave the DPA as published and put your specific terms in the order form, under a precedence clause that makes the order form win.
• "Without undue delay is the market standard for breach notice." Article 33 of the GDPR gives you, as controller, 72 hours from awareness to notify your regulator. Open wording on the processor side makes that clock hard to plan around, so ask for a fixed number of hours.
• "Zero data retention is a product decision." Accept that, request the review now, and write the approval into the agreement once it is granted, with the endpoints it covers.

What have we seen in recent OpenAI contract reviews?

We reviewed roughly 25 to 35 GenAI vendor contracts between 2024 and 2025. OpenAI paper was the one most often signed without a single privacy redline, more than any other vendor in that set.

Which patterns came up again and again?

• Training defaults misread. Teams assumed API traffic trains models by default, then paid for assurances OpenAI had already given in writing.
• Retention left unmanaged. In 60 to 70 percent of agreements, the buyer kept the standard abuse monitoring window where zero data retention was available for its use case.
• Addendum never executed. Roughly 1 in 4 enterprise buyers had no signed DPA, despite processing personal data through the API.

The first and third patterns point in opposite directions. One is paying for protection you already have; the other is operating without protection you assumed you had.

Why we disagree that OpenAI's terms are fixed below seven figures

The usual advice is that OpenAI's terms cannot be negotiated below seven figures, so you should sign the standard paper and get on with deployment. Our files say otherwise. In roughly 12 of the 30 reviews we ran in 2024 to 2025, named retention, subprocessor notice and deletion clauses were accepted at mid six figure spend.

They were accepted because each ask restated a published commitment as a contract term instead of inventing a new obligation. A restatement adds no risk for the vendor's legal team to price, so it clears review quickly, while a new obligation goes into a queue.

Most privacy clauses are conceded fastest when the redline quotes the published commitment it restates.

Converting policy into contract is cheaper than any assurance you could buy, and we would make that change to almost every OpenAI agreement we see.

How do you check your current OpenAI privacy position?

Start with the documents and settings you already have. Work through this list with your security lead before any renewal or expansion talk.

1. Find the signed DPA. You want the executed version naming your legal entity, filed with the order form.
2. Check the API data sharing settings. Confirm in the organization settings of the API dashboard that the organization has not opted in to share inputs and outputs with OpenAI.
3. List projects and endpoints. Flag every project that uses stateful endpoints such as assistants, vector stores or files, whatever your retention approval says.
4. Look for written approvals. If zero data retention or Modified Abuse Monitoring was approved only in an email thread, get the approval referenced in the agreement.
5. Review ChatGPT Enterprise retention. Check the retention setting in the admin console and who has the rights to change it.
6. Compare subprocessor lists. Set the current list on the trust portal against the last copy in your file.

Which mistakes cost buyers the most?

• Treating the privacy page as the contract. It explains OpenAI's intent, yet it is not a signed term and it can change.
• Signing for one entity only. If several subsidiaries use the same OpenAI organization or workspace, the DPA has to cover each of them as a controller.
• Choosing a region too late. ChatGPT workspaces and API projects take their region at creation, so a residency clause signed after rollout can mean moving to a new workspace or project.
• Leaving deletion open. "Return or delete at your instruction" gives you no date to hold OpenAI to after termination.

For the commercial clauses that sit alongside these, see the seven OpenAI clauses to push back on and our guide to data governance terms in AI agreements.

What to do next

1. Execute the DPA first. Get the processing addendum signed before anything else, because it is the gap that exposes you most once personal data runs through the API.
2. Restate the training default. Put the no training commitment in the order form, so it survives a policy page edit.
3. Submit the retention review early. Check whether your use case qualifies for zero data retention and file the request now, since approval depends on the use case.
4. Name the numbers. Write the retention window, the subprocessor notice period, the breach notice hours and the deletion timeline into the contract, under a precedence clause.
5. Start privacy in week one. Send the addendum and security questionnaire alongside the commercial terms, so privacy review runs in parallel with pricing.
6. Keep dated copies. File the business terms, DPA and subprocessor list as signed, and compare them with each update notice. Our OpenAI contract risk review covers this if you want a second reader.

Frequently asked questions