AI tools and client data

Is this AI tool safe for client data?

The honest answer for every tool below is “it depends which plan and which account”, so this page says what each vendor publishes about its own product, quoted and dated. A free or personal plan and a business plan are often opposites. Where we did not find a statement in the pages we read, the table says that rather than guessing.

By Kfir Yair, CISSP. Updated .

See which of these your staff use

The three answers that decide it

Training on a personal plan, training on a business plan, and whether the vendor publishes a Business Associate Agreement position. Each tool’s own page has the retention, admin access, single sign-on and data-location answers, with the quotes.

ToolFree or personal plan: trains on inputs?Business plan: trains on inputs?BAA offered?
ChatGPTYes. On a personal account, OpenAI says data sharing for model improvement is on until someone turns it off.No, not by default. OpenAI says business data is excluded unless the customer opts in.OpenAI says it can sign one; the sentence sits in its API Platform FAQ, and ChatGPT for Healthcare is its HIPAA workspace.
Microsoft CopilotNot found in the pages we read.No. Microsoft says prompts, responses and the Graph data behind them are not used to train the foundation models.Microsoft lists HIPAA among Copilot's compliance offerings. The page states no BAA itself; the agreement lives in your Microsoft terms.
Google GeminiSome conversations are read by people, and Google says that is part of how it improves the service.No. Google says Workspace content is not reviewed by people or used for training outside your own domain without permission.Google signs a BAA for Workspace, but only for the services on its HIPAA Included Functionality list, which this page does not reproduce.
ClaudeOnly if the person says yes. Anthropic frames training on consumer chats as a choice, and a safety review as the exception.No, under commercial terms, unless the organisation chooses to hand data over through a named programme.Enterprise plans only, switched on from organisation settings; Anthropic says the offering includes the agreement itself.
PerplexityYes. Improving and creating its AI models is one of the purposes Perplexity lists for the content people submit.No. Perplexity says an organisation's data is never used to train or fine-tune its models.Not found in the pages we read.
DeepSeekYes. DeepSeek lists training its models among the purposes for the content people type in.Not found in the pages we read.Not found in the pages we read.
Zoom AI CompanionNo. Zoom says it does not use customer content to train its own or third-party AI models.The same statement, and Zoom extends it to the model providers it uses.Zoom signs BAAs, and ties the AI features to having one in place.
Otter.aiYes. Otter.ai lists training its own AI on de-identified recordings and on transcriptions among its purposes.Otter.ai offers Enterprise a way out of model training through an account manager, which means data is in scope until somebody arranges it.Enterprise plan only, arranged with an account manager. Otter.ai says it is a business associate only once the agreement is signed.
Fireflies.aiNo. Fireflies.ai says it does not train on personal information and that it holds its vendors to the same rule.The same statement: Fireflies.ai's policy does not distinguish plans on training.Fireflies.ai says an agreement is available. Ask which plan it needs and what else has to be switched on before any patient data goes in.
Read AIRead AI's support page says the default is opt out for product improvement. Its privacy policy still lists training among its purposes.The same default, and on Enterprise+ an admin can add a retention policy on top of it.Annual Enterprise+ only, and Read AI will not sign until SAML and domain capture are set up.
GrammarlyYes. The policy served at Grammarly's privacy URL lists training AI models among the uses of collected information.Not found in the pages we read.Grammarly says it will sign one on request, and does not say which plans qualify.
Notion AINo, by default. Notion says neither it nor its AI providers train on customer data, and that its contracts forbid it.The same no-training statement, which Notion says flows through to every provider in the chain.Enterprise plan only. Notion also puts its beta services outside the agreement.
GitHub CopilotYes, since 24 April 2026, unless the person opts out. GitHub names inputs, outputs, code snippets and context.No. GitHub says business and enterprise customer data is not used to train models, under its data protection agreement.Not found in the pages we read.
CursorIt turns on one setting. Cursor says Privacy Mode is available on every plan and that it then does not train on your data.An admin can set Privacy Mode for the team, so it is not left to each developer.Not found in the pages we read.
Claude CodeOnly with the setting on, and Anthropic says that covers Claude Code used from those accounts.No, under commercial terms, unless the organisation opts in through a named programme.The agreement sits at organisation level on an Enterprise plan, which is the same organisation the Claude Code seats belong to.

“Not found in the pages we read” means what it says: the vendor pages listed on each tool’s page, read on the dates shown there, do not answer it. It is not a criticism and it is not a reassurance — it is a question to put to the vendor in writing before client information goes near the tool.

Every tool, with what the vendor says

You cannot answer this from a list

Every answer above turns on which plan and which account a person signed in with, and that is the one thing a list cannot tell you. Obsidian Ridge’s free 30-day AI security assessment takes a quiet inventory of the computers for 30 days, then a CISSP goes through the findings with you.

Start my free assessment

Or work through the ones you know about first: the AI tool risk checker turns what you already know into a list of decisions, and the staff AI policy builder turns that into a one-page policy, both in your browser. See a sample set of findings, or read how to find the AI tools employees use.