Does Your SaaS Vendor Train AI on Your Data? 25 Major Vendors' Terms Compared

A slim majority of vendors keep business content out of training by default, but 5 of 25 train unless someone flips a setting. If nobody on your team owns that setting, assume the default applies to you.

Only 29% of organizations say they feed unedited customer data into generative AI, but 40% use anonymized or pseudonymized customer data. Some vendors in our set, including Intercom and Otter.ai, describe their training data the same way, so ask how they de-identify it.

Combined, 79% of organizations say their generative AI providers agreed to adjust terms or settings to limit data use. Just 13% disagreed and 3% strongly disagreed, which suggests asking is usually worth the email.

In Pew's May 2023 survey, 81% of U.S. adults who had heard of AI expected companies to use collected data in ways people aren't comfortable with. Your own customers likely share that view about their data in your tools.

Does the vendor train AI on your data? 25 SaaS vendors' terms compared (business or enterprise tier)

Source: Adobe, Adobe's approach to generative AI (Firefly); Anthropic, Is my data used for model training? (commercial products); Asana, Asana Product-Specific Terms; Atlassian, Atlassian Intelligence and Rovo trust page; Amazon Web Services, Amazon Bedrock FAQs; Box, Box AI Principles; Cursor (Anysphere), Data Use & Privacy Overview; Docusign, AI Attachment for Docusign Services; Figma, Manage AI settings and content training for your team or organization; GitHub, Updates to GitHub Copilot interaction data usage policy; Google, Generative AI in Google Workspace Privacy Hub; Grammarly, Product Improvement and Training Control; HubSpot, Understand HubSpot's AI model training; Intercom, AI model training and fine-tuning; LinkedIn, LinkedIn and generative AI (GAI) FAQs; Miro, Miro AI trust and ISO 42001; Microsoft, Data, Privacy, and Security for Microsoft Copilot (Microsoft 365 Copilot); Notion, Notion AI security practices; OpenAI, Enterprise privacy at OpenAI; Otter.ai, Otter.ai Privacy Policy; Salesforce, Trusted AI: Einstein Trust Layer; Slack, Privacy principles: search, learning and artificial intelligence; Workday, Concept: Workday AI for Universal Main Subscription Agreement Customers; Zendesk, Zendesk AI data use and privacy; Zoom, Zoom Terms of Service, Section 10. Download as image

Same vendor, different answer: individual plans vs business plans

Source: Microsoft, Privacy FAQ for Microsoft Copilot (consumer); Microsoft, Data, Privacy, and Security for Microsoft Copilot (Microsoft 365 Copilot); OpenAI, How your data is used to improve model performance; OpenAI, Enterprise privacy at OpenAI; Anthropic, Updates to Consumer Terms and Privacy Policy; Anthropic, Is my data used for model training? (commercial products); GitHub, Updates to GitHub Copilot interaction data usage policy; Google, Gemini API Additional Terms of Service; Grammarly, Product Improvement and Training Control; Figma, Manage AI settings and content training for your team or organization. Download as image

Retention and deletion periods vendors state for AI-related data

Source: Anthropic, Updates to Consumer Terms and Privacy Policy; Microsoft, Privacy FAQ for Microsoft Copilot (consumer); OpenAI, Data controls in the OpenAI platform (Your data); Google, Gemini API Additional Terms of Service; Workday, Manage Workday AI Data Contributions for UMSA Customers; Intercom, AI model training and fine-tuning. Download as image

Does Microsoft Copilot train on your data?

It depends on which Copilot you are using, and that confusion is where most people go wrong. Microsoft 365 Copilot, the version tied to a work account, does not use your prompts, responses or Microsoft Graph data to train foundation LLMs. Admins control optional feedback and can apply Purview retention policies.

Consumer Copilot, the one you sign into with a personal Microsoft account, works differently. Conversations may be used to train generative models, there is an opt-out, and Microsoft keeps conversation activity for 18 months by default.

So if an employee pastes a client contract into Copilot on a personal login, your company's enterprise promise doesn't cover it. The same split shows up at OpenAI, Anthropic, GitHub, Google, Grammarly and Figma. Check the plan, not the brand.

Which SaaS vendors train AI on customer data by default?

A smaller group in our set trains on business data by default or ties it to a setting nobody may have touched. HubSpot may train its own models on customer data, with a Super Admin opt-out. Intercom trains its Fin models on de-identified conversation data and takes up to 30 days to process an opt-out request. Workday turns data contributions on automatically for purchased AI features.

Docusign's terms grant it a training licence, and its opt-out only works going forward, so data used before you switch it off stays used. Otter.ai trains on de-identified audio and transcripts with no opt-out described in its privacy policy. Cursor trains when Privacy Mode is off, and its page doesn't say what the default is. LinkedIn uses member data unless you turn off a specific setting.

None of these is hidden. Each one is in the vendor's own published terms. The problem is that the person who signs up for the tool rarely reads them, and the admin who could change the setting often doesn't know it exists.

Here is what that looks like in practice. A sales team signs up for a CRM, connects its inbox and starts logging calls. Months later, nobody remembers that the AI settings page had a training switch. Every deal note, pricing discussion and customer complaint in that account may now be eligible for the vendor's own models. The fix took one click. Finding out it was needed took a policy page most people never open.

What does "we don't train on your data" leave out?

Read the promise word for word. Five vendors in our set made commitments that sound complete but cover only part of the picture. Here are four of them.

Slack won't use your data for generative AI unless you opt in, but it does use it for global non-generative machine learning models, and you have to contact Slack to opt out. Zendesk trains its proprietary non-generative models on sanitized customer data. Salesforce and Atlassian bar outside model providers from keeping or training on your data, but the pages we read don't address their own in-house models.

Other gaps to watch for: feedback buttons that count as opting in, retention periods for prompts and logs, and whether de-identified or aggregated data counts as "your data" at all. A thumbs-up on an answer can be all it takes to share that conversation.

This is why a vendor's marketing line is a starting point, not an answer. The line usually says the most comforting true thing. The detail sits in the product terms, an AI addendum or a trust center page, and those are the documents worth reading before you connect your customer data.

Why do free and paid plans get different answers?

Business plans get the stronger promises. One likely reason is that business buyers negotiate and individual users usually don't. GitHub began training on Copilot Free, Pro and Pro+ data on April 24, 2026 unless users opt out, while Business and Enterprise plans were not affected.

Anthropic lets Claude Free, Pro and Max users choose. If they allow training, data is kept for 5 years. If they don't, the 30-day retention period stays. Grammarly and Figma both switch training on by default for smaller plans and off for their organization tiers.

For a freelancer or a small agency, this matters a lot. You may be paying for an individual plan that trains on your client files while a larger competitor's enterprise plan of the same product doesn't. Either move client work onto a business tier or find the opt-out and screenshot it.

What should freelancers and small businesses do first?

Start with the tools that touch other people's data. Your clients' contracts, your customers' support tickets and your meeting recordings carry the most risk, because you may have promised confidentiality to someone else. If a vendor trains on that material, you could be the one breaching an NDA.

Meeting tools deserve special attention. A transcription app that joins a client call records everyone on it, not just you. Otter.ai's privacy policy describes training on de-identified audio and transcripts with no AI-training opt-out, so think twice before inviting it into a confidential meeting.

Then look at your own consent. Pew found that 81% of U.S. adults who had heard of AI expected companies to use collected data in uncomfortable ways. Your clients probably feel the same, and a short line in your own contracts saying which tools you use can head off a hard conversation later.

Can you negotiate AI training terms with a vendor?

Often, yes. In Cisco's 2026 benchmark, 79% of organizations said their generative AI providers were willing to adjust contract terms or tool settings to limit data exposure. Yet only 55% require contracts that spell out data ownership, usage rights and IP for AI vendors, even though 70% acknowledge risk from using proprietary or customer data in AI training.

Think about who actually holds the cards. Asking a vendor to confirm its existing policy in writing is the easy request, since it costs the vendor nothing. It is harder to get a vendor to change a default that its product depends on, like a support platform that trains on conversations. In that case, ask for a narrower clause: training only on de-identified data, no use of attachments, and deletion within a fixed period after you opt out.

That gap is the opportunity. You don't need enterprise spend to ask for a short AI addendum or a written confirmation that training is off for your account. Start with these steps:

• List every tool your team uses that has an AI feature, including free sign-ups on personal emails.
• For each one, find the vendor's AI or data use page and note which plan it covers.
• Switch off training settings at the admin level, then save a dated screenshot.
• Ask the vendor to confirm in writing that your content won't train any model, generative or not.
• Add a clause to renewals that requires notice before the vendor changes its AI data terms.

Sources and method

Every number on this page comes from the sources below, each opened and read on or before October 18, 2026. A second, independent review checked every figure against its source and corrected anything that did not match. Figures that could not be confirmed were left out.

1. Cisco, Cisco 2026 Data and Privacy Benchmark Study (2026 edition (survey of 5,200+ professionals in 12 markets, fielded September 2025))
2. Pew Research Center, Key findings about Americans and data privacy (October 18, 2023 (survey of 5,101 U.S. adults, May 15-21, 2023))
3. Microsoft, Data, Privacy, and Security for Microsoft Copilot (Microsoft 365 Copilot) (Last updated August 18, 2026)
4. Microsoft, Privacy FAQ for Microsoft Copilot (consumer) (Accessed October 18, 2026)
5. GitHub, Updates to GitHub Copilot interaction data usage policy (March 25, 2026)
6. Google, Generative AI in Google Workspace Privacy Hub (Last updated October 6, 2026)
7. Google, Gemini API Additional Terms of Service (Last updated April 28, 2026)
8. OpenAI, Enterprise privacy at OpenAI (Page as archived October 2, 2026)
9. OpenAI, How your data is used to improve model performance (Page as archived October 18, 2026)
10. OpenAI, Data controls in the OpenAI platform (Your data) (Accessed October 18, 2026)
11. Anthropic, Updates to Consumer Terms and Privacy Policy (August 28, 2025)
12. Anthropic, Is my data used for model training? (commercial products) (August 18, 2026)
13. Amazon Web Services, Amazon Bedrock FAQs (Accessed October 18, 2026)
14. Salesforce, Trusted AI: Einstein Trust Layer (Accessed October 18, 2026)
15. Slack, Privacy principles: search, learning and artificial intelligence (Accessed October 18, 2026)
16. Zoom, Zoom Terms of Service, Section 10 (Effective date shown on page: August 11, 2023 (accessed October 18, 2026))
17. Atlassian, Atlassian Intelligence and Rovo trust page (Accessed October 18, 2026)
18. Notion, Notion AI security practices (Accessed October 18, 2026)
19. Figma, Manage AI settings and content training for your team or organization (Accessed October 18, 2026)
20. Grammarly, Product Improvement and Training Control (Accessed October 18, 2026)
21. Zendesk, Zendesk AI data use and privacy (Updated May 27, 2026)
22. Intercom, AI model training and fine-tuning (September 8, 2026)
23. HubSpot, Understand HubSpot's AI model training (Last updated September 8, 2026)
24. Workday, Concept: Workday AI for Universal Main Subscription Agreement Customers (Last updated June 26, 2026)
25. Workday, Manage Workday AI Data Contributions for UMSA Customers (Last updated June 26, 2026)
26. Docusign, AI Attachment for Docusign Services (Version date July 8, 2026)
27. Adobe, Adobe's approach to generative AI (Firefly) (Undated page (accessed October 18, 2026))
28. Box, Box AI Principles (Effective November 4, 2025)
29. Asana, Asana Product-Specific Terms (Last updated October 1, 2026; effective October 5, 2026)
30. Miro, Miro AI trust and ISO 42001 (Accessed October 18, 2026)
31. Otter.ai, Otter.ai Privacy Policy (Effective June 16, 2026)
32. Cursor (Anysphere), Data Use & Privacy Overview (Last updated September 3, 2026)
33. LinkedIn, LinkedIn and generative AI (GAI) FAQs (Accessed October 18, 2026)