Illustrative scenario. The procurement committee at a mid-sized insurer is reviewing the AI vendor that will read claims files. The headline question lands at minute three: «Do you train on our data?» «No. Never, on any business plan.» Relief, minutes, next item. Nobody asks the other one: if they don't use it for training, what do they do with it in the meantime? Do they keep it? Where, for how long, and who can read it?
The thesis in one line: «we don't train on your data» is a promise about use; retention is a promise about existence. Data that trains nobody but lives thirty days in a log, two years in a flagged-content folder, or forever in a thread nobody deleted is still data that can leak, be subpoenaed or be read in a review. «We don't train» doesn't fix that.
AI vendor data retention at your company: two different promises
The market has trained buyers to ask a single question. After years of headlines about models learning from whatever you pasted in, the big vendors have made it clear they don't train on business and API customer data. Good. But training is only one place data can end up. The others are duller and far more common: a log to watch for abuse, the state the API itself keeps so the conversation can continue, a cache to save compute, a copy held because a classifier flagged it or because a judge ordered it.
That's why AI vendor data retention isn't answered with a number. It's answered with a map: which layers keep something, for how long, and what you control. Here's that map, built from the terms the vendors themselves publish as of November 2026. They change often: treat it as a list of questions, not a fixed answer.
What does an AI vendor keep even when it doesn't train on your data?
Five layers. They almost never show up on the same documentation page, which is why almost nobody sees them together.
| Layer | What it is | Typical published period | Who controls it |
|---|---|---|---|
| Abuse logs | Copy of inputs and outputs to detect misuse | Up to 30 days on the OpenAI and Anthropic APIs | The vendor (unless zero retention is agreed) |
| Application state | What the API stores so things work: threads, files, stored responses, batches | Until you delete it, or 30 days for OpenAI stored responses | You, if you know it exists |
| Caches | Reusable prefixes or context to save compute | Hours; up to 24 h for OpenAI and Google caches | Shared |
| Flagged content | Whatever a safety classifier flags | Up to 2 years for inputs and outputs at Anthropic | The vendor |
| Legal obligation | Retention imposed by law or a court | Indefinite, for as long as it lasts | Neither of you |
The figures come from official documentation: OpenAI's API data controls guide, Anthropic's organization data retention article (dated July 1, 2026) and Google Cloud's zero data retention page. These are global product terms, not country terms: the period doesn't change because you sign from Chicago or from Austin. Where the data sits can change, though, and that's a separate question.
Abuse logs: the thirty days nobody reads in the demo
OpenAI keeps its API abuse-monitoring logs for up to 30 days unless the law requires longer. Anthropic deletes API inputs and outputs within 30 days of receipt or generation, with four written exceptions: services with retention under your control (like its Files API), a different agreement (like zero retention), enforcing its usage policy, and complying with the law. Thirty days is nothing for a business model and a lot for a medical record.
Application state: what you asked it to keep
This is the treacherous layer, because your own team switches retention on without knowing it. On the OpenAI API, assistants, threads, vector stores, files, batches and fine-tuning jobs are kept until you delete them; deleted objects leave its servers 30 days later. At Anthropic, the Files API keeps data until deletion or expiry, batch processing keeps it 29 days and code execution containers up to 30. None of those features fall under either vendor's zero retention. The agent you shipped in an afternoon with file uploads has a retention policy: «forever, until somebody remembers».
Flagged content and the judge's order
If an automated classifier flags a conversation, Anthropic can keep inputs and outputs for up to 2 years and classification scores for up to 7. You don't have to do anything wrong: it's enough that your use case — a criminal defense firm, a mental health platform, a security team — looks like something the classifier watches. And the law sits above everything: in 2025, a US federal court ordered OpenAI to preserve logs that would otherwise have been deleted in its lawsuit with The New York Times; the blanket obligation was lifted effective September 26, 2025, with exceptions, according to Engadget. «We delete it» is a policy. A court order is something else.
«Zero data retention»: what it covers and what it doesn't
Zero data retention (ZDR) exists and it works. But it's not a toggle in the dashboard: it's negotiated. At OpenAI it needs prior approval; at Anthropic it's enabled per organization through sales, and every new organization needs its own activation. And it covers endpoints, not your project: if your workflow uses files, batches or tools that aren't eligible, that part stays on standard retention even when the contract says «ZDR».
On the big clouds the rule repeats with nuances. On Azure, Microsoft's human reviewers only access prompts and completions that have already been flagged, and modified abuse monitoring — no storage, no human review — requires meeting extra criteria and applying through a form, per its data and privacy page. On Google, grounding with Search keeps logs for up to 3 days and can't be switched off; grounding with Maps, 30 days. Turning on a «smart» feature can add a retention layer your zero-retention contract doesn't touch.
The retention questions to ask before you sign
If you already ask for the data processing agreement — and you should, as we explain in what your AI vendor won't tell you about GDPR — add these six. Answers should come in writing and point to a documentation page or a clause, not a slide.
- How many days do you keep inputs and outputs by default, and why? Abuse, debugging, billing: each purpose is a period.
- Which features we'll use store state, and until when? Files, threads, memory, batches, stored responses. Check the list against your design, not their catalog.
- Do you offer zero retention, on which endpoints and which models? And whether the model you're proposing excludes it.
- What happens to what your classifier flags? Period, who reviews it, and whether you tell us.
- Who can read retained data, and from where? Human reviewers, subprocessors, region.
- How do we delete, and how do we verify it? A deletion you can't verify is one more promise.
Notice that none of these questions mentions training. You already have that answer. What you're missing is everything else.
What you do control: your architecture decides retention
The contract sets the maximum the vendor keeps. Your design decides how much you hand over. Three decisions cut exposure without waiting on any negotiation:
- Don't store state at the vendor if you don't need to. Turn off response storage, move history and agent memory into your own system, and delete files when the job is done. What isn't uploaded isn't retained.
- Minimize before you send. Swap names, ID numbers or policy numbers for identifiers and rebuild on the way back. The vendor gets the problem, not the person.
- Decide by data type, not by vendor. Public data goes wherever is cheapest; sensitive data goes to a zero-retention endpoint or to a model running inside your perimeter, as covered in open source AI automation tools.
Then write it down where you already decide who approves what: governance and control of AI automation isn't complete until it says what leaves the building, to which vendor and for how long.
The other retention: yours
Here's the irony. Plenty of companies that squeeze the vendor down to zero days keep prompts and responses themselves in a log with no expiry date, personal data included, «just in case». Or the reverse: they rotate logs at thirty days and, when the complaint arrives, there's nothing to show. Your own retention is designed too: how long each kind of trace is kept, what gets masked and who can query it. That's exactly what we ship when we build AI decision traceability, for the same reason: a period nobody chose is still a decision.
When isn't all this worth it?
When what you send is public or identifies nobody: summarizing press articles, generating product descriptions from your catalog, translating your website. Thirty days of abuse logs don't matter there, and negotiating zero retention is wasted time. The retention question matters in proportion to what the data would say about someone if it got out. Health, financial and employee data, named customers, case files: there, yes, the full map, in writing.
And don't confuse the map with the badge. A certification tells you the vendor has controls; it doesn't tell you how many days it keeps your prompts, as we explain in what SOC 2 and ISO 42001 actually prove. That number lives in the product docs, and it changes faster than any certificate.
Note: this article is not legal advice. The periods cited are those published by each vendor in its official documentation, checked on November 12, 2026, and they apply globally; always verify current terms and your contract before deciding.