subinternal Talk to us

ResearchPrivate AI

Is zero data retention enough for regulated data?

Zero data retention is a promise about what a model provider keeps. Your data still goes to that provider. For some data, only a model inside your own walls will do.

QuestionZero data retentionPrivate AI
Who runs the modelWhose machines read the requestThe providerIn its cloud, under its termsYouOn hardware you control
What the provider keepsAfter it answersLess, once set upEach feature set one by oneNothingNo provider in the path
CachingOf recent requestsOn by defaultUp to 24 hours, unless turned offYours to decideOn your own machines
Search groundingAnswers checked against the webKept up to three daysGoogle says it can’t be turned offNot part of itIt can run with no internet at all
Abuse monitoringPrompts logged for safetyMay log promptsUnder the provider’s termsNoneEvery lookup logged for your team
What you rely onFor the promise to holdThe contract and settingsChecked feature by featureWhere the machine sitsA boundary you can inspect
Zero data retention as Google documents it for Vertex AI, read in November 2026; other providers’ terms differ in detail. Private AI as we build it, from our own page. Both are cited below.

What does zero data retention mean?

That the provider running the model agrees not to keep what you send it, once it has answered. Your request still travels to the provider and is read on its machines. What changes is what it keeps afterward, and that depends on the features you use.

On Google’s Vertex AI it is not one switch. Google says that “to achieve zero data retention, customers must take specific actions” in each of several areas, and that some retention can’t be avoided at all.1

What is still kept under it?

Google’s own list is the clearest we have found.1

  • Caching is on by default for supported Gemini models, in memory and at rest, for up to 24 hours. It can be turned off per project.
  • Abuse monitoring may log prompts to Google models, under section 4.3 of the Google Cloud terms.
  • Grounding with Google Search keeps queries derived from your prompts for up to three days. In Google’s words, “there is no way to disable the storage of this information.”
  • Grounding with Google Maps keeps prompts, context and output for 30 days, and the Deep Research agent keeps prompts and session data for 7 days. Neither can be turned off.
  • The Interactions API stores by default until you set it not to.

None of this is hidden; it is documented. But it means a promise of zero retention is only as good as the settings and features behind it, checked one by one.

What do vendors mean when they promise it?

Usually that their model provider keeps nothing, under the terms above. Growth Protocol’s security page is a typical example: it says “no client data retained by model providers”, and also that zero data retention patterns are “available by deployment design”.2 That is a fair promise, and it is a promise about one deployment. Ask which design yours would get, and which features it leaves out.

What does it mean for data never to leave?

That no outside model provider is in the path at all. In private AI, open-weight models run on hardware you control, and nothing goes to a third-party model provider. It can run air gapped, with no internet connection. Frontier models stay off until you switch them on, and even then they see no names.3 When a team wants a ready-to-use workspace without co-mingling records in a shared multi-tenant SaaS database, we also deploy an isolated single-tenant Secure Cloud environment, and Business Super Intelligence can run either in your own cloud account or on hardware inside your network.4

The promise then rests on where a machine sits and what it is connected to, which your own team can inspect, rather than on a contract and a list of settings.

Which one does regulated data need?

It depends on what your obligation forbids. If it is about storage, training and multi-tenant co-mingling, and it allows a cloud provider to process the data under contract, an isolated single-tenant Secure Cloud with zero data retention set up carefully can meet it. If it forbids the data from being processed outside your control at all, as controlled unclassified information, export-controlled data and some client agreements do, no retention promise helps, because the data still leaves. Then the model has to come to the data on your own hardware.

This is our judgment, not legal advice. Read the obligation with whoever owns it, and ask one question first: does it limit what is kept, or where the data may go?

What should you ask a vendor that promises it?

Five things, in writing. Which provider runs the model, and in which region. Whether caching is off for your project. Whether prompts can still be logged for abuse monitoring, and on what terms. Whether any feature you will use, such as web search grounding or a research agent, keeps data regardless. And where all of that is stated in your contract, not only on a web page.

What can’t this page tell you?

Whether your obligation allows a provider. Google’s terms here were read from its documentation in November 2026 and change; OpenAI, Anthropic and Microsoft offer their own versions, with their own exceptions. Check the current page for the provider you would use, and the contract you would sign.

References

  1. Google Cloud, “Vertex AI and zero data retention.” Read November 2026. docs.cloud.google.com
  2. Growth Protocol, “Security.” Read November 2026. growthprotocol.ai
  3. subinternal, “Private AI on your own hardware.” October 2026. subinternal.com
  4. subinternal, “Business Super Intelligence.” November 2026. subinternal.com