Back to Guides
Enterprise Software 15 min read

Decoding vendor data-handling promises: training, retention, and opt-outs

Decoding vendor data-handling promises: training, retention, and opt-outs

“We don’t train on your data” answers one question out of three. Whether your content improves a model, how long it sits in the vendor’s systems, and whether a human can read it are separate commitments, documented in separate places, and the same vendor often answers them differently.

AI vendor data handling has an unusual property: the facts are published, free, and ignored. Every major assistant keeps a trust page stating what it does with your prompts, and almost nobody at a 200-person company has read one — the reassurance arrives earlier in the sales cycle than the documentation, and it sounds complete.

The three promises people hear as one

Separate them before reading any vendor page; the page itself usually does not.

Training. Is your content used to improve a model? This is the promise vendors lead with, and the one most often honored on paid tiers.

Retention. How long is your content stored, and who sets the clock? A no-training commitment says nothing about this.

Human review. Can a person at the vendor, or a contractor it uses, read what your team typed? Staff care about this most; procurement asks about it least.

The gap between the three is documented. Stanford HAI published a review on October 15, 2025 of the privacy documentation of six developers — Amazon’s Nova, Anthropic’s Claude, Google’s Gemini, Meta AI, Microsoft Copilot and OpenAI’s ChatGPT. Lead author Jennifer King and colleagues read 28 documents and found that all six use consumer chat data by default to train their models, that some keep it indefinitely, and that opt-out availability varies. Their broader finding matters more: the policies routinely omit what a reader needs to answer basic questions.

What the four assistants document today

Four products cover most of what a US company of 50 to 500 people has deployed. Here is what each one’s own documentation said in October 2026 — a worked example rather than a lasting fact, since every row is one announcement away from changing.

Product Training on your content Retention default Human review
ChatGPT business, enterprise and API tiers Excluded by default. The platform documentation dates the API commitment to March 1, 2023, and the same exclusion covers the business workspace tiers API abuse-monitoring logs up to 30 days; the workspace window is set by the workspace owner, not by the vendor Abuse-monitoring logs are reviewable; zero data retention, where approved, removes content from them
Claude commercial tiers Excluded by the commercial terms. The September 2025 consumer change did not reach Claude for Work, Enterprise, Education or the API Indefinite by default on Claude Enterprise. An owner can set a custom period; the stated minimum is 30 days Content flagged by automated trust and safety systems sits outside the ordinary window
Gemini consumer tier (Workspace terms differ) Used unless activity is turned off Auto-delete at 18 months by default, changeable to 3 or 36 months or off. With activity off, chats are still kept 72 hours A subset goes to human reviewers; reviewed chats are kept up to three years, and deleting your activity does not remove them
Microsoft Copilot in Microsoft 365 “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs” Interaction content is stored with your other Microsoft 365 content and governed by your own Purview retention policies, not a vendor window Microsoft documents that Copilot services have opted out of Azure OpenAI abuse monitoring, which includes human review

Read side by side, the no-training column is the least interesting: on paid tiers the promise is now broadly real. The spread is in the other two.

The retention defaults point in opposite directions. Claude Enterprise keeps data indefinitely until an owner changes it; Gemini’s consumer tier deletes on an 18-month timer; Microsoft hands you the clock. A company running all three on the assumption of one behavior has three different exposures.

Human review is where the answers diverge most and the documentation is most specific. Google’s three-year track for reviewed conversations survives deletion of your activity — exactly the fact a staff member would want before pasting a client file. Microsoft’s documented opt-out from abuse-monitoring human review is a checkable distinction no questionnaire asks for.

What an opt-out cannot undo

An opt-out is a deadline, not a remedy. Anthropic’s own retention page states it plainly: turning the model-improvement setting off stops future use, but data already in training runs in progress, or in models already trained, remains included. Model weights do not unlearn.

The practical consequence is worth more than any clause: if you pilot on consumer accounts while procurement works through paperwork, the exposure accrues and cannot be withdrawn afterward. Set tiers and settings on day one.

The second thing an opt-out cannot touch is a legal hold. In the consolidated copyright litigation against OpenAI, a preservation order required the company to retain output log data it would otherwise have deleted. A stipulated order approved by Magistrate Judge Ona Wang on October 9, 2025 terminated that obligation as of September 26, 2025, while leaving already-preserved logs in place. A published retention window is an intention, not a guarantee, and no vendor claims otherwise.

Four phrases and what each one leaves out

Four sentences do most of the reassuring in this market. Each is usually true, and none means what a non-specialist hears.

What it says What it does not cover
“We never sell your data” Training, retention, human review, subprocessors. Almost no assistant vendor sells data, so the sentence is free to say
“Your data is not used for training” Whether it is stored, for how long, and who can read it. Check whether the wording covers outputs and feedback as well as inputs
“We are SOC 2 compliant” Nothing about AI specifically. Ask for the Type II report with its date, and for ISO/IEC 42001 if an AI management system matters
“We align with the NIST AI Risk Management Framework” Nothing checkable. The framework is voluntary guidance and NIST certifies no one against it, so every vendor can say this

The useful question is not whether a vendor makes these statements but which of the three promises each one is about. Mostly, none of them. The companion list of questions worth putting to a vendor, each with the document that settles it covers the procurement side.

Zero data retention is narrower than the name

Zero data retention is the term most often misread here, usually as a stronger no-training promise. It is a separate control with three limits. It is not a default — OpenAI’s documentation describes it as requiring prior approval. It is an API arrangement rather than a workspace setting, so it says nothing about how a ChatGPT workspace behaves. And it does not cover everything: the platform documentation lists ineligible endpoints including conversations, assistants, threads, vector stores, files, fine-tuning jobs, evals and batches, which may still store application state with the control enabled.

This stays background until someone builds an internal tool on an API key, which happens earlier and more quietly than leadership teams expect.

The subprocessor chain inside tools you already pay for

The harder version of this problem is not the assistant. It is the AI feature switched on inside software you bought for something else: the note-taker, the CRM summarizer, the help-desk reply drafter. That vendor is usually routing your content to one of the providers above, so the promise you need is not the one you read.

Two things are worth checking rather than assuming. A vendor’s own no-training commitment does not automatically bind the model provider behind it. And the routing can carry constraints invisible from the product page — Microsoft’s documentation notes that Anthropic models offered as a subprocessor within Copilot experiences are currently excluded from the EU Data Boundary.

So for any tool with an AI feature, ask which model provider is behind it, under what terms, and to be told when that changes. A model swap can move your data to a new provider in a new jurisdiction with no visible change to the product. The settings that have to be live before anyone is trained start with knowing which tools have AI switched on at all.

What a 200-person company can actually change

Nearly all published guidance here is clause-drafting advice for a buyer with real negotiating power. On a standard business subscription the terms are the terms, and the levers that remain are narrower and more effective.

Tier. The largest change available to you is which account people log into, and it is a purchasing decision rather than a negotiation. The consumer-to-business gap is wider than anything a contract would win you, which is the security argument for paying for seats.

Retention settings. Indefinite retention on one product and an 18-month timer on another are somebody’s default, not your decision. Set each deliberately and write down the number you chose.

Scope. Decide which categories of material are allowed near an assistant at all. This is the only control that does not depend on a vendor keeping a promise.

Notice and evidence. Ask for written notice of changes to retention, subprocessors and model providers, and for deletion confirmation when a contract ends. Notice converts a silent change into a dated email, and a vendor who cannot produce a deletion record is telling you something.

Budget honestly: reading four products’ documentation and recording the answers is a half-day, and keeping it current is an hour a quarter. The expensive part is the training that makes any of it hold, where market rates for hands-on workshop programs run roughly $2,000 to $15,000 by scope and headcount.

Turning a trust page into something staff can use

A decoded vendor page governs nothing by itself. The exposure happens when someone decides whether to paste a document, and that person will not consult a table. Three translations survive a session:

  1. Name the sanctioned account, once. Most staff cannot say which tier they are on. They need one sentence: which login is covered and which is not.
  2. State the retention number out loud. “Our workspace keeps conversations for 30 days” changes behavior in a way “the vendor has strong privacy practices” never does.
  3. Say who could read it. Staff assume either that nobody can or that everybody can. The actual answer, including the flagged-content exception, is what makes a never-paste rule feel reasoned rather than arbitrary.

Those three facts fit on half a page and belong in the AI security training session that teaches how staff recognize what must never go into a chatbot. The documentation is what makes that judgment defensible; without it, the rule is a list people half-remember. Closing the distance between an AI use policy and actual behavior is the subject of the safe AI use playbook.

This is the part of AI enablement no license includes. Terms, settings and clauses are buyable inside a quarter; the judgment to apply them is built, which is the argument in why buying licenses is not the same as building capability.

When to read it all again

Everything above has a shelf life, and Anthropic’s consumer change is the clearest evidence. Announced on August 28, 2025 with a decision deadline of September 28, 2025, it flipped the training default on for Free, Pro and Max users and extended retention to five years for anyone who allowed it. Commercial customers were excluded. It was handled properly, and it still moved the ground under every company whose policy quoted the old default. A cadence that works without a compliance function:

  • Quarterly — re-read the trust pages for the products you actually use, and record the date and the three answers.
  • At renewal — ask what changed in the terms, and whether an AI feature was switched on by default in software you already owned.
  • On any vendor notice — check whether it touches training, retention or human review. Tell staff only if one of those three moved; the rest is noise.
  • Annually — confirm the retention numbers you set are still the numbers in the admin console. Settings drift when administrators change.

Dating what you read is the step people skip and the one that makes the record useful. An undated claim that a vendor does not train on your data is how a policy ends up describing a world that has moved on.

For a second read on which of the three promises your stack actually makes, book a free AI-readiness call.

Frequently asked questions

Does “we don’t train on your data” mean our data is not stored?

No. They are separate commitments. Content can be excluded from every training run and still be retained for years: Claude Enterprise documents indefinite retention by default until an owner sets a shorter window.

Can we get our data removed from a model it was already used to train?

No. Opting out stops future use, and Anthropic’s documentation states that data already in a training run in progress, or in a model already trained, stays included. Model weights do not unlearn, which is why tier and settings decisions belong at the start of a pilot.

Do the major assistants train on business data by default?

Generally not on paid business tiers. OpenAI dates its API exclusion to March 1, 2023 and applies the same default to its business workspaces; Anthropic’s commercial terms exclude customer content; Microsoft states that Copilot prompts and responses are not used to train foundation models. Consumer tiers run the opposite default, so verify each against the current page.

How long does a vendor actually keep our conversations?

It depends on the product and on who set the clock. Published defaults range from an 18-month auto-delete on Gemini’s consumer tier to indefinite on Claude Enterprise, with ChatGPT workspaces and Copilot handing the decision to you. Flagged content and feedback sit on longer, separate tracks.

Can anyone at the vendor read what our team typed?

Sometimes, and the documentation is specific. Google states that a subset of Gemini conversations goes to human reviewers and that reviewed chats are kept up to three years, surviving deletion of your activity. Microsoft documents that Copilot services have opted out of abuse-monitoring human review. Flagged content is the common exception.

Is a settings toggle as good as a contract term?

No. A toggle can be moved and a help page rewritten; signed terms cannot be changed unilaterally. Where terms are not negotiable, the next best thing is a written notice commitment, so a change arrives as a dated email.

What about AI features inside software we already bought?

Treat each as a new vendor review. The tool is usually routing your content to one of the major model providers, and its own promise does not bind the provider behind it. An AI feature enabled by default at renewal is the most common way an unreviewed model provider enters a company.

Does any of this satisfy an auditor or an insurer?

Only if it is written down and dated. A record of which products you use, what each one’s terms said when you read them, what retention you set, and when AI governance training was delivered is the artifact they ask for.

Key takeaways

  • Training, retention and human review are three separate promises, answered differently by the same vendor. On paid tiers, the no-training commitment is the least informative.
  • Retention defaults diverge sharply and none is neutral. Verify each product’s terms, record the date, and set the number deliberately.
  • An opt-out is a deadline, not a remedy: data already used in a completed training run stays in the model, and a legal hold overrides any published deletion window.
  • “No training”, “SOC 2 compliant” and “NIST aligned” are all compatible with indefinite retention and human review.
  • At 50 to 500 people the real levers are tier, retention settings, scope and written notice, not contract redlines.
  • None of it governs anything until the people pasting material can state which account is sanctioned, how long it is kept, and who could read it.

Last Updated: Nov 16, 2026

DJ

Dirk Jan van Veen, PhD

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles