Strong for business products
OpenAI — ChatGPT Business, Enterprise, and API
OpenAI distinguishes business products from personal ChatGPT use. For business products, OpenAI states that customer inputs and outputs are not used for model training by default.
Relevant wording
"By default, we do not train on any inputs or outputs from our products for business users, including ChatGPT Business, ChatGPT Enterprise, and the API."
Practical application
Suitable for normal business workflows when the account is explicitly ChatGPT Business, ChatGPT Enterprise, or API. Do not treat personal ChatGPT accounts as equivalent.
Strong for commercial offerings
Anthropic — Claude Team, Enterprise, and API
Anthropic's commercial terms and privacy center provide a clear business posture: Claude for Work and API content is not used for training unless the customer opts in or submits feedback.
Relevant wording
"We will not use your chats or coding sessions to train our models, unless you choose to participate..." Anthropic's commercial terms also state that Customer Content is Customer's Confidential Information.
Practical application
Appropriate for confidential workflows only when using Claude Team, Claude Enterprise, or API under the commercial terms, with retention and access controls reviewed.
Client-controlled AWS path
AWS — Amazon Bedrock with Claude
Amazon Bedrock is the relevant AWS path for organizations that want model access inside their cloud governance perimeter. Claude models can be invoked through Bedrock and reviewed under AWS data protection, IAM, logging, and regional controls.
Relevant wording
The key point is deployment architecture rather than a consumer privacy promise: client data can remain in the client's AWS environment while the application invokes approved foundation models through Bedrock.
Practical application
Best fit for regulated or sensitive workflows where real KYC, suitability, complaint, or client-document data should stay inside a client-approved cloud environment.
Client-controlled Google Cloud path
Google Cloud — Vertex AI with Claude or Gemini
Google Cloud provides enterprise model access through Vertex AI and related Gemini services. This should be treated separately from consumer Gemini use, which has different privacy and review implications.
Relevant wording
Google Cloud's Gemini documentation states that prompts and responses are not used as data to train Gemini models. Vertex AI also supports partner-model access, including Claude models.
Practical application
Viable when the client already operates under Google Cloud or Workspace controls. The exact model, region, data residency, and contract terms should be confirmed before using real client data.
Microsoft tenant / Azure path
Microsoft — Microsoft 365 Copilot and Azure AI Foundry
For Microsoft-heavy organizations, the strongest posture is Microsoft 365 Copilot inside the tenant, or approved Azure-hosted models through Azure AI Foundry. Microsoft states that M365 Copilot prompts, responses, and Microsoft Graph data are not used to train foundation models.
Relevant wording
"Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs."
Practical application
Useful for workflows that depend on SharePoint, Teams, Outlook, or Microsoft Graph permissions. For Azure AI Foundry, confirm the exact model, seller, region, and contract before representing availability or privacy terms.
Connector boundary
Connected systems — Gmail, SharePoint, Dropbox, Slack, Teams, Gamma, and MCP tools
The LLM is only one part of the confidentiality boundary. If a workflow connects to external systems, each system must be reviewed for training use, retention, subprocessors, data residency, audit logs, admin controls, and whether data leaves the approved environment.
Relevant issue
A business AI plan does not automatically make every connected application compliant. The connected services can become the weak point in the privacy analysis.
Practical application
For regulated workflows, prefer systems already covered by the client's enterprise agreements, such as Microsoft 365, SharePoint, Google Workspace, AWS, Azure, or GCP. Avoid consumer SaaS connectors for real client records unless their enterprise terms satisfy the review.