Summary

A product-by-product comparison of AI training policies for companies choosing where to put business and client data. It separates ordinary training defaults from feedback, retention and hosting decisions, then shows what to check in the account actually used.

What you'll learn

  • Which business and individual products use conversations for model improvement
  • How feedback and preview features change the ordinary data rules
  • Why training, storage, human review and deletion have separate checks
  • How buying the same model directly or through a cloud platform changes the agreement
  • What to record before staff or agents send client data

What is the difference between training and retention?

Training changes a model; retention describes stored data. OpenAI treats abuse logs and saved application state separately.[2] Ask how long each is kept, whether people can review it and where processing occurs, even when the product promises no training.

Read the setup through four questions: model improvement, storage, human review and abuse monitoring. Those are separate processing purposes. Ask for the product-specific rule and any exception for each.

Four separate questions for training, retention, human review and abuse monitoring
A no-training statement answers one question; storage and review need their own checks.

Which AI providers train on business data?

As of October 2026, AI-provider training policies differ by product and account. Mistral’s DPA authorizes training unless the customer has opted out.[7] Its help page gives Team administrators an organization-wide opt-out, while Enterprise has a different default.[8]

Product / pathTraining defaultRetention or exception to check
OpenAI business (Business, Enterprise, Edu)No model improvement by default[1]Separate sharing agreement; response feedback alone does not permit conversation training
OpenAI APINo training unless explicitly opted in[2]Default abuse logs up to 30 days, with exceptions; stored application state is separate
OpenAI individual (Free, Plus, Pro)Content may be used; opt out in data controls[1]Feedback can include the related conversation after opt-out
Anthropic commercialNo training by default[3]Feedback may train and be retained up to 5 years; API and saved-chat retention differ[5]
Anthropic consumer (Free, Pro, Max)Permission-based, with safety-review/feedback exceptions[4]Training-pipeline retention up to 5 years when enabled[6]
Mistral Team (listed as Vibe Teams)Training until the organization opts out[7][8]Admin disables training; feedback and Labs/Preview exceptions remain[9]
Mistral EnterpriseOpted out by default[8]Admin-managed opt-in; do not extend this to every Mistral product
Google Gemini Apps (consumer)Keep Activity affects model improvement[10]Default activity auto-delete 18 months; reviewed material retained up to 3 years
Foundry Models sold by AzureNo foundation-model training without permission[11]Check the deployment type and processing region
Microsoft 365 commercial CopilotNo foundation-model training on prompts/responses[12]Feedback not used for foundation-model training; check EU-boundary exclusions and routing[13]
AWS BedrockModel providers cannot access prompts/completions[14]Check the selected feature’s storage and AWS terms separately

Download the AI-provider training-policy comparison (PDF)

The table reports public product boundaries, not the terms of a negotiated agreement. “No training” and “no provider access” are different statements. Check storage and connected services separately before relying on either.

Does ChatGPT use your data for training?

As of October 2026, OpenAI separates individual accounts from Business, Enterprise, Edu and API use. Business inputs and outputs are not used for model improvement by default. Individual users can disable improvement for new conversations, with a separate feedback exception.[1]

Check the account someone is using for the actual task. A work instruction does not change the contract on a personal account. OpenAI’s current page also states that business response feedback alone does not authorize training on the associated conversation.

For the processing agreement and the differences between plans, read ChatGPT and GDPR.

The same check for more than 300 AI and work tools, plan by plan, each answer dated: which AI tools train on your data.

Does Mistral Team train on chats by default?

Mistral Team requires an organization-level training opt-out. The DPA’s section 2.3, effective 27 July 2026, permits training unless the customer is opted out; the 3 September help page lists Vibe Teams admin controls.[7][8] Do not import Enterprise’s default into a Team account.

The documented admin control disables use of interactions for training across the organization. Uploaded documents count as inputs in that guidance. Verify the saved setting before sending company or client documents.

Mistral’s commercial terms separately preserve feedback and Labs/Preview exceptions.[9] The default row cannot stand in for those exceptions. European ChatGPT alternatives covers the wider product decision.

Can feedback change the training policy?

Feedback can have different data rules from ordinary use. Anthropic’s commercial policy dated 18 August 2026 allows submitted feedback to be used for training and retains the related conversation for up to five years.[3] Microsoft’s commercial Copilot page says feedback is not used for foundation-model training.[12]

Anthropic excludes raw connected-source content from this feedback description unless it is copied into the conversation. Team and Enterprise owners can disable response ratings. Decide whether that control fits your client-data commitments before users start submitting ratings.

Mistral’s DPA also covers associated input and output submitted as feedback.[7] Review the feedback mechanism separately from the ordinary training toggle. Explain that difference in onboarding; do not claim every provider follows the same rule.

A business account’s ordinary training rule does not answer what happens when someone submits feedback.

Does zero data retention cover every feature?

Zero Data Retention has a defined feature scope. As of October 2026, OpenAI requires prior approval and lists feature-specific eligibility.[2] Mistral’s commercial terms effective 5 August 2026 exclude Labs and Preview models from opt-out preferences, including zero data retention.[9]

For OpenAI, stored application state is a separate category from abuse logs. A setting for eligible calls does not delete every saved object or third-party record. Ask the operator to identify the features actually used and their storage behaviour.

For Mistral, avoid Labs/Preview models when the published training exception conflicts with the intended data boundary. Check the model actually selected, including fallback models. A workflow can use a preview model even when the account’s ordinary setting looks restrictive.

Does deleting a chat remove every copy?

Chat deletion and review retention are separate in Google Gemini Apps. As of October 2026, the activity default is 18 months, while human-reviewed material can remain for up to three years after activity deletion. Temporary and activity-off chats are retained for 72 hours.[10]

Anthropic’s July 2026 retention pages also distinguish saved conversations, deletion processing and training pipelines.[5][6] The API’s ordinary 30-day window is not an automatic deletion rule for all Claude for Work chats.

For Anthropic API use, the standard deletion window for inputs and outputs is within 30 days, with service, agreement, safety and legal exceptions. Saved commercial chats remain until deleted, then backend deletion takes up to 30 days.[5] That distinction is why the chart labels the short Claude bar as API use.

Illustrative retention periods: OpenAI abuse logs and Claude API standard data up to 30 days, Gemini activity default 18 months, reviewed Gemini material up to three years, and Anthropic feedback up to five years
These clocks cover different data and triggers. Saved commercial chats, feature storage and exceptions need separate checks.

Keep retention questions on their own checklist: what is stored, which event starts the clock, whether review or feedback creates an exception and what deletion actually removes. Data erasure owns the detailed removal process.

Does the hosting path change the training policy?

The provider contract depends on the service path. For models sold by Azure, Microsoft keeps the data separate from OpenAI and other model providers.[11] AWS says Bedrock model providers cannot access deployment accounts or customer prompts and completions.[14]

Identify who operates the actual service and read the terms for that offering. Then check the storage used by its optional features. A cloud platform can sell several types of service; use the agreement for the one your workflow actually calls.

Processing geography is another question. Microsoft distinguishes deployment types; its commercial Copilot documentation excludes Anthropic models from the EU Data Boundary.[12] Flex routing is on by default for eligible tenants created after 25 March 2026.[13] Training restrictions do not settle those location questions.

Read Copilot and GDPR for the settings and data residency for the separate location owner.

One agent routes to a direct provider, Azure or Bedrock, with different contractual paths
Check the named service and agreement; check the product and route your workflow actually uses.

How should you check client-data use?

Before sharing client data, match the product, account and settings to the commitments you have made to that client. Mistral’s DPA distinguishes processing for the customer from its own controller role for training.[7] Read the training clause alongside the processing agreement.

Record the account, plan, selected model, hosting path and settings evidence. Compare the documented processing with the commitments already made to clients. Escalate a mismatch to the person responsible for the agreement before sending data.

Verify the configured account and product. A company instruction cannot turn a personal account into a protected commercial workspace.

Keep the data flow narrow enough to inspect. Connected applications can introduce additional recipients and storage. Ask what leaves the company, who receives it and which agreement covers it.

For an agency or adviser, make the review concrete before connecting the inbox:

  • Confirm the tier. Record the account used by the running workflow, not only the account used during the demo. Ask who can change the selected product or model.
  • Confirm the region and deployment type. If the client expects EU processing, have the operator identify the exact service route and any fallback. “We use Azure” leaves the deployment choice open.
  • Read the client agreement. Compare the described use, recipients and locations with what you have promised. If they differ, let the person responsible for that agreement resolve it before the connection goes live.

Give the team a usable rule for client documents. Specify which workspace to use, which feedback controls to avoid or disable, and whom to ask about an unfamiliar tool. Include a realistic example, such as a contract pasted into a personal chat or attached to a feedback report. People need to know the permitted route at the moment they are doing the work.

Ask how the agent limits what it sends. A reply draft may need the relevant contract clause, not the complete client archive. Review the proposed information and recipients during the supervised trial, including what is copied into logs or summaries. Apply the same account and route review when another connected application is added.

How do you verify your own settings?

Save the product name, plan, account and the settings used by the workflow. Check the training control and feedback mechanism separately. For Mistral Team, the training control belongs to the organization’s administrator.[8] OpenAI also distinguishes business feedback from personal-account feedback.[1]

Use this review sequence: identify the product and agreement; confirm training and feedback controls; check preview-model exceptions; record retention and processing location. Re-check when a plan, model, connector or contract changes.

Keep the review with the workflow documentation so a colleague can check what was agreed. Record who owns the decision and when a change should trigger another review. If the selected product or model changes, update the record before sending the next batch of client data.

What can public policies tell you about your own setup?

Public policies explain product defaults and named exceptions. To make a decision about your workflow, read the agreement that actually applies to your account, check the saved settings and confirm the features in use. A negotiated term may change the public default, so keep it alongside the configuration record.

The table does not show whether your team followed the rule. An employee can select a personal account, enable an optional feature or submit feedback containing client material. Confirm the actual route with the operator and give users clear instructions for those cases.

Ask separately who can access retained records, what starts each deletion clock and how data moves through connected tools. Keep legal and client-contract questions with the person responsible for those commitments; give them the product name, service path and data flow so they can assess the concrete use.

Frequently asked questions

Does ChatGPT train on your data? Individual ChatGPT accounts can have content used for model improvement unless you opt out in data controls. The opt-out applies to new conversations; submitted feedback can include the associated conversation. Business, Enterprise, Edu and API use have separate defaults.

Does ChatGPT Business train on your data? OpenAI does not use Business, Enterprise or Edu inputs and outputs for model improvement by default. The current policy says response feedback alone does not authorize training on the associated business conversation.

Does Mistral Team train on chats? Yes, unless the organization opts out. Mistral’s DPA permits training unless opted out, and its help page gives Team administrators an organization-wide control. Enterprise is opted out by default; feedback and preview exceptions need separate review.

Does Claude train on business data? Anthropic commercial products do not train by default. Submitted feedback may be used for training, with related conversations retained for up to five years and specified connected-source exclusions.

Does no training mean no retention? No. Training, retention, human review and abuse monitoring are separate questions. Check the selected product and feature rather than treating a no-training commitment as zero storage.

Is my data safe if I use AI through Microsoft or AWS instead? Check the exact service, agreement and deployment. For models sold by Azure, Microsoft separates customer data from the model providers; AWS says Bedrock model providers cannot access customer prompts and completions. Still check storage, processing region, optional features and connected services for your actual setup.

Notes

  1. Model improvement controls — OpenAI, living page, checked 14 October 2026.
  2. API data controls — OpenAI, living documentation, checked 14 October 2026.
  3. Commercial model training — Anthropic, 18 August 2026.
  4. Consumer model training — Anthropic, 16 March 2026.
  5. Commercial retention — Anthropic, 1 July 2026.
  6. Consumer retention — Anthropic, 1 July 2026.
  7. Data Processing Addendum — Mistral AI, effective 27 July 2026.
  8. Training opt-out controls — Mistral AI, 3 September 2026.
  9. Commercial Terms of Service — Mistral AI, effective 5 August 2026.
  10. Gemini Apps Privacy Hub — Google, living page, checked 14 October 2026.
  11. Foundry model privacy — Microsoft, living page, checked 14 October 2026.
  12. Microsoft Copilot privacy — Microsoft, living page, checked 14 October 2026.
  13. Flex routing — Microsoft, 29 September 2026.
  14. Bedrock data protection — AWS, living page, checked 14 October 2026.