The short version

If you are a solo founder or a small team piping customer information through AI tools, three knobs matter most: how long the provider keeps your prompts and outputs, whether anything you send can be used to train future models, and whether you can submit a data deletion request that actually wipes the record. Each major provider handles those three knobs a little differently, and the labels in the dashboards do not always match what is happening on the backend.

This guide walks through what the settings actually do, where to find them, and how to think about trade-offs without getting lost in policy language.

Why this matters when you handle customer data

Even a tiny product touches personal data. A support reply that includes a customer’s email, a transcript fed into a summarizer, a doc uploaded for analysis — all of it becomes a prompt. Once that prompt lives on someone else’s server, your customer’s privacy posture is partly your vendor’s privacy posture.

Three questions usually come up:

  • Will my prompts be used to train the next model version?
  • How long does the provider keep my prompts and outputs?
  • Can I ask them to delete what I sent?

Most providers now answer those questions in plain settings panels, but the answers are buried under marketing-friendly labels. Reading them carefully saves you from a nasty surprise during a customer security review or a contract renewal.

What “data retention” actually means

Retention is the time window between sending a prompt and the provider permanently deleting the request and its response from active systems. It is not the same as training opt-out.

  • Training opt-out controls whether your conversations may be used to improve future models. Turning this off does not erase anything; it just stops one specific use.
  • Temporary retention covers the days or weeks a provider holds data to monitor abuse, troubleshoot outages, or meet legal holds.
  • Permanent retention covers anything you agree to keep for training, fine-tuning, or product improvement under an opt-in.

A common mistake is treating “I turned off training” as “nothing is stored.” In practice, most providers still hold a copy for a short period even when training is disabled.

How OpenAI handles it

OpenAI splits the experience across consumer ChatGPT, workspace plans (Team, Business, Enterprise, Edu), and the API Platform. The behavior differs in each.

Consumer ChatGPT (Free, Plus, Pro). Training is on by default. You can switch it off in Settings under Data Controls, by toggling “Improve the model for everyone” to off. Once you do that, new conversations are not used for training. You can also export or delete your history from the same area. Temporary Chat is a separate mode that skips history entirely.

Workspace and Enterprise plans. Training is off by default. Admins control retention and connected sources, and SAML SSO is available. Because training is already disabled, the more practical question becomes retention length and audit access.

API Platform. The API has its own Data Controls documentation describing what is stored, how abuse monitoring works, and which endpoints support stricter guarantees. Eligible API customers can also apply for Zero Data Retention (ZDR), which is a stricter commitment: prompts and responses are not retained by OpenAI after the request is processed, and customer content is not available for human review. ZDR is the right setting if your prompts contain regulated or customer-identifying information and you can route them through the API rather than a chat UI.

A useful habit: decide which interface you use for which kind of data. Casual product brainstorming can live in a consumer account with training off; anything with customer content should go through the API under ZDR or through a workspace plan with admin-controlled retention.

How Anthropic handles it

Anthropic’s commercial terms for Claude generally cover the main points. The relevant policy details shift over time, so treat any specific figure you read as a snapshot and verify against the current policy page before relying on it.

Key habits still apply:

  • Review the commercial terms for Claude before sending customer data.
  • Check whether the API surface you use offers stricter retention than the consumer chat surface.
  • Decide whether your prompts fall under a category the provider flags as opt-in for training.
  • Document your choice internally so future you, or a future hire, knows why a particular setting is on.

For most indie workflows, the practical question is not the exact number of days — it is whether the API tier you pay for has a stricter retention profile than the free consumer chat, and whether your prompts contain anything that would matter if a human reviewed them.

Practical settings checklist for founders

A short routine beats memorizing policies. Walk through this whenever you onboard a new AI tool:

  • Find the data controls page. Search the provider’s help center for “data controls,” “privacy,” or “training opt-out.” Skim the FAQ once.
  • Disable training by default. Unless you have a specific reason to share data for model improvement, switch the training toggle off.
  • Choose the right surface. Route customer data through the API or a workspace plan with stricter commitments, not a personal chat account.
  • Ask about retention length. Note the default window for abuse monitoring and incident response. Some providers publish this; others describe it in policy language.
  • Confirm deletion paths. Check whether you can export and delete your history, and whether enterprise admins can do it across the team.
  • Write it down. One paragraph in your internal security doc, with screenshots and the date you checked, beats six months of guessing.

When a zero-retention promise matters

Zero Data Retention is not a marketing line at OpenAI — it is a contractual commitment for eligible API customers that the provider does not retain prompts or responses after processing. For solo founders, the question is whether your workflow can live on the API instead of a chat UI.

That trade-off usually looks like this: ZDR costs you the convenience of a chat interface and a chat history you can scroll back through, and it asks you to build a thin layer that records what you need locally. In return, you get a much cleaner answer to the question “where does my customer’s email live after I paste it into an AI tool?”

If you are only occasionally summarizing a doc, a workspace plan with training disabled is probably enough. If you are processing customer support tickets, contracts, or anything covered by a regulator, ZDR or an equivalent commitment is worth the small engineering cost.

Common misconceptions worth retiring

  • “I turned off training, so nothing is stored.” Storage and training are separate. Most providers still keep a short-window copy for abuse monitoring.
  • “An enterprise account deletes everything.” Enterprise accounts usually disable training, but they do not necessarily erase everything immediately. Admin-controlled retention windows still apply.
  • “Deleting my chat deletes it everywhere.” Deletion usually applies to your view and future training. Anything already used in a snapshot or fine-tuning before you opted out is typically not retroactively removed.

A short founder FAQ

Do I need zero retention to use AI safely? Not always. For internal drafting and research, training-off plus reasonable retention is fine. For customer data under a contract or a regulator, stricter settings are worth it.

Where do I find the opt-out toggle? For ChatGPT, it is Settings, then Data Controls, then “Improve the model for everyone.” For the API, check the Data Controls guide for your provider. Other vendors usually label it “Improve the model,” “Share data for training,” or similar.

Can I delete a conversation after the fact? On most consumer plans, yes, from the same Data Controls area. On workspace and enterprise plans, admins control deletion policy. Deletion typically applies to your history and future use, not to data already absorbed into model weights.

Is the API always safer than chat? Usually, yes, because the API gives you direct control over which endpoints you call and which guarantees you request. Chat UIs tend to prioritize convenience over configurability.

Closing thought

Treat AI data settings the way you treat a password manager: configure once, write it down, and revisit when the provider ships a change. The biggest risk for a small team is not that someone reads one prompt — it is that nobody on the team knows which surface is being used for which kind of data, and a default setting quietly trains the next model on a customer’s support ticket.

Sources