Where Your Client's Data Goes: The Switches to Set on OpenAI, Claude, Gemini and Mistral Before the First Paste, and the Data Note to Give the Client

Before a client's spreadsheet, transcript or customer list goes into a model, there is an hour of setup, and a page to write. The setup is a handful of switches, different at each provider, that decide where the material goes, how long it stays and who may read it. The page is the data note you attach to the client contract, in the providers' own words. This guide does both, provider by provider, reading each one's terms and documentation, and it ends with the two files you should have when the hour is over.

What you hand over

A configured account per provider, on the right tier, with the right model and the right storage settings, and screenshots of each setting with the date. And the data note: one page, one paragraph per provider, quoting the training sentence, the retention period, the deletion rule and what zero retention would take. The note is what the client's lawyer reads; the screenshots are what you read when a page changes.

Step one: the training sentence, in each provider's words

Every provider answers the first question in one sentence, and the note quotes it rather than paraphrasing it. OpenAI's data page: as of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models unless you explicitly opt in to share data. Anthropic's retention page: retained data is never used for model training without your express permission; its commercial terms add that Anthropic may not train models on Customer Content. Gemini's pricing page, under every model: on the Free tier your content is used to improve Google's products, on the Paid tier it is not. Mistral's commercial terms: Mistral AI will not use Customer Data or Outputs to train its models except when you opted in on a product set to opt-out by default or did not opt out on a product set to opt-in by default, when you provide Feedback, as provided in an Order Form, or when you use Labs or Preview Models.

Four sentences, and the whole rest of the page is the switches that keep each one true.

Switch one: OpenAI

OpenAI's data page has a table by endpoint, and it is the setup in one place. For /v1/chat/completions and /v1/responses: no training, abuse-monitoring retention of 30 days, no application state except the listed exceptions, and eligible for Zero Data Retention. For /v1/conversations, /v1/conversations/items and /v1/chatkit/threads: retained until deleted, and not eligible. For /v1/agents, /v1/assistants and /v1/threads: application state until deleted, not eligible. So the first switch is a choice of endpoint: client work goes through responses or chat completions, and anything built on conversations or threads holds the client's text until you delete it yourself.

The second is the log that no setting removes on a standard account. Abuse-monitoring logs may contain prompts and responses and are generated for all API usage, retained for up to 30 days by default unless the law or protection of the service requires longer. Eligible customers can have their content excluded from those logs through Zero Data Retention or Modified Abuse Monitoring, subject to prior approval by OpenAI and acceptance of additional requirements; once approved, a Data Retention tab appears under Settings, Organization, Data controls, with a choice at the organization level and, per project, default, one of the two controls, or none. Under Zero Data Retention the store parameter on responses and chat completions is always treated as false, whatever the request says. Write the 30 days into the note.

Switch two: Anthropic

Anthropic's retention page starts from the other end: conversation content, your prompts and Claude's outputs, is not retained by default, and only what a feature technically needs is kept. Then it names the exception, and the exception is a model choice. Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5 and Claude Mythos 5 are Covered Models that require 30-day data retention and are not available under zero data retention unless Anthropic expressly authorizes it. So the switch is the model name in your request: for a client who needs nothing retained, use a model that is not on that list. The privacy centre's general rule still stands, that API inputs and outputs are deleted within 30 days of receipt except under services with longer retention that you control, such as the Files API.

Zero data retention itself is requested from the sales team and enabled per organization, and its exclusions are the working rules for a freelancer. It does not cover any usage in the Claude Console, including the playground, so a client's document pasted into the Workbench to test a prompt is outside the arrangement; it does not cover Claude Managed Agents, whose session transcripts persist until you delete them; it does not cover the consumer plans; and CORS is not supported under it, so a browser-based tool must call through your own backend. Organizations that handle health data can enable HIPAA readiness from Console, Settings, Privacy, with a standard agreement.

Switch three: Gemini

Gemini's first switch is the tier, and the rate-limits page says setting up billing in AI Studio moves the project to Tier 1, typically instantly. That is the row under every model turning from yes to no. The tier compare reads that row in full.

The second and third switches are on the paid tier only, in the Logs and Datasets page of AI Studio. Logging of Gemini API calls is opt-in and produces logs that are your own data, separate from Google's abuse-monitoring logs. If you switch it on, logs are retained for a default maximum of 55 days, and the Settings panel lets you set 7, 14, 28 or 55; logs you save into datasets have no retention period at all. The two APIs behave differently by default: the Generate Content API does not store requests (store false) while the Interactions API does (store true), and both can be set per request or toggled per project in that panel. The line to underline for the note is about sharing: a dataset shared with Google is processed under the terms for Unpaid Services and may be used to develop and improve Google's products, including training its models, and Google's own page says not to include personal, sensitive or confidential information in one. For client work, the answer is to never share a dataset.

Switch four: Mistral

Mistral's terms describe two kinds of product, those set to opt-out by default and those set to opt-in by default, and the training exception depends on which kind you are using and what you chose. Find the preference for each product you use in your organization's settings and record it in the note; the terms name the preference but do not name the page. The terms also grant Mistral a licence to use Customer Data and Outputs to provide, maintain and optimize the products, which the same clause says excludes model training. And they carve out Labs and Preview Models: by using them you acknowledge that Mistral AI may use Customer Data and Outputs from them to train its models, and that opt-out preferences selected for other products, including through zero data retention, do not apply to them. For client work, Labs and Preview are off limits, whatever the toggle says elsewhere.

Voices are data too

If the client's material is a recording, the note gains a paragraph. ElevenLabs says a Professional Voice Clone stays only in your account unless you share it, and that generated content may be used commercially as long as you hold the necessary intellectual property rights to it. A cloned voice is personal data with a face on it; the consent to clone and the promise not to share belong next to the training sentences, and the voiceover page reads the licence side.

Asking for zero retention

Both providers that offer it make it a conversation, not a checkbox. OpenAI's requires approval and acceptance of additional requirements, and its table lists endpoints that stay ineligible even when it is on. Anthropic's is an agreement per organization, requested from sales, with the exclusions above. If a client's contract says zero retention, the request goes in writing before the first token, and the note records the date it was granted and the endpoints and models it covers.

What is deleted when you leave

OpenAI's services agreement deletes all Customer Content from its systems within thirty days of termination and hands nothing back. Anthropic's privacy centre deletes API inputs and outputs within 30 days of receipt, the Files API aside. The copies you kept for the client are the only copies; the drafting and transcription pages say the same about their deliverables.

Suspension is a data event

OpenAI may limit or suspend access if required by law or if the customer breaches the agreement or its policies, without prior notice where necessary. Anthropic may suspend all access over a suspected Usage Policy breach and disclaims liability for any loss of data or profits that suspension causes. ElevenLabs suspends or terminates for conduct such as fraudulent transactions or submitting misleading or unauthorized content. A suspended account is one whose stored data you cannot reach or delete on the client's timetable, which is why nothing on this page relies on going back later to clean up.

What to do today

Open the four pages linked in the sources and copy the four training sentences into a document. Under each, write the switch you set and the date: the endpoint and the 30-day log at OpenAI, the model name at Anthropic, the tier and the Logs and Datasets settings at Gemini, the product preference at Mistral. Screenshot each setting into a folder with the same date. That document is the data note, and the next client who asks where their data goes gets it before they ask.