Most people encounter the AI tools data privacy question in one of two ways: either they read something concerning in the news and feel vaguely worried, or they paste something sensitive into an AI tool and only wonder afterward whether they should have. I’ve been in that second situation. Early in my use of these tools I pasted client email threads into ChatGPT to help draft responses — without stopping to think clearly about what I was actually doing: sending a client’s private communications to a third-party company’s servers, to be processed by their systems, potentially stored and used in ways the client had never consented to. The output was useful. The data handling wasn’t appropriate. For the bigger picture, our Complete Guide to AI Tools pulls everything together.
This guide covers AI tools and data privacy in 2026 with the directness that the topic requires — what the major tools actually do with your data, which categories of information should never go into consumer AI tools, and what the enterprise options look like for professional use that involves sensitive content.
What the major tools actually do with your data
The foundational principle: when you use a consumer AI tool, the data you send to that service is subject to their privacy policy, their security practices, and their business decisions about how to use that data. The policies below reflect the state of these tools in mid-2026; check current privacy policies directly before relying on this information for compliance decisions.
OpenAI (ChatGPT): By default on the consumer free and Plus tiers, conversations may be used to improve OpenAI’s models — meaning your inputs and the model’s outputs become training data. OpenAI provides an opt-out: Settings → Data Controls → Improve the model for everyone → toggle Off. With this setting off, your conversations are not used for training. OpenAI retains conversation data for up to 30 days for safety and abuse monitoring even with training opt-out enabled. The API — used by businesses building on OpenAI’s models — has different terms: API inputs and outputs are not used for training by default.
Anthropic (Claude): Claude.ai’s privacy policy provides that conversations on the consumer tier may be reviewed by Anthropic employees and used to improve Claude, though Anthropic states this is done with privacy protections in place. Claude has a setting to turn off conversation history, which limits some but not all data retention. The Anthropic API and Claude for Enterprise have stronger data handling commitments — inputs are not used for training, and data is subject to enterprise data processing agreements.
Google (Gemini): Google’s data practices with Gemini are governed by their broader privacy policy, which has historically been oriented toward using interaction data to improve Google products. Human reviewers may read conversations to improve safety and quality. Google Workspace users have different terms — Workspace data is subject to the Workspace data processing agreement rather than consumer terms.
The consistent pattern across all three: consumer tiers have more permissive data handling; enterprise and API tiers have stronger protections. For anything beyond personal, non-sensitive use, the consumer tier is not the appropriate choice from a data privacy standpoint.
What should never go into consumer AI tools
Based on the data handling practices above, these categories of information should not go into consumer AI tools without explicit enterprise data agreements in place:
- Personal data about other people — names, contact details, financial information, health information, location data about identifiable individuals. Pasting a customer database or client contact list into a consumer AI tool creates a data breach risk regardless of intent.
- Medical or health information — patient records, health insurance information, medical history. This is regulated under HIPAA in the US and equivalent frameworks elsewhere. Sending it to an unvetted AI tool is a compliance violation regardless of how secure the transmission is.
- Financial account information — account numbers, financial records, tax documents. Beyond compliance risk, specific financial details create direct fraud risk if exposed.
- Confidential business information — unreleased product plans, merger discussions, proprietary processes, competitive intelligence, trade secrets. This information leaving your organisation’s systems without authorisation may violate confidentiality obligations and NDA commitments.
- Legal documents with privilege implications — attorney-client communications, litigation strategy documents, anything covered by legal professional privilege. Sharing these with a third party can destroy the privilege.
- Login credentials and API keys — obviously, but stated explicitly: never paste passwords, API keys, or authentication tokens into any AI tool interface.
Enterprise options for professional use with sensitive data
For professional and business use involving sensitive data, the appropriate options are enterprise AI tiers with formal data processing agreements.
ChatGPT Enterprise and OpenAI API: Data is not used for model training. Enterprise agreements include data processing commitments compliant with GDPR and equivalent regulations. Data is processed in encrypted form and deleted after a defined retention period. For businesses that need to use AI tools with customer or client data, the API or Enterprise tier is the minimum appropriate option.
Claude for Enterprise: Anthropic’s enterprise offering includes formal data processing agreements, no training on enterprise conversation data, SOC 2 compliance, and options for data residency requirements. The enterprise tier is appropriate for regulated industries and organisations with formal data protection obligations.
Microsoft 365 Copilot: Built on Azure OpenAI Service with Microsoft’s enterprise data protection commitments. Data stays within the organisation’s Microsoft 365 tenant and is subject to the existing Microsoft data processing agreement. For organisations already on Microsoft 365 with established data governance, Copilot provides AI capability within an already-approved data handling framework — the data governance decision has effectively already been made through the Microsoft 365 relationship.
| Data type | Consumer AI tools | Enterprise AI tools |
| Personal writing and ideas | Appropriate with training opt-out | Appropriate |
| Non-confidential work content | Appropriate with training opt-out | Appropriate |
| Client personal data | Not appropriate | Appropriate with DPA in place |
| Confidential business information | Not appropriate | Appropriate with enterprise agreement |
| Medical/health information | Not appropriate | Only with HIPAA BAA in place |
| Legal privileged communications | Not appropriate | Legal advice required before use |
Practical risk reduction steps right now
For users who need to use AI tools for work but don’t have enterprise agreements in place, these steps reduce risk meaningfully:
Anonymise before pasting. Replace real names with placeholders, remove identifying details, and replace specific numbers with representative figures. The AI can help with the task without the identifying information in most cases. A client email with “[Client Name]” and “[Company Name]” substituted for the actual names is usable for drafting assistance without exposing personal data.
Opt out of training data use. In ChatGPT: Settings → Data Controls → turn off model training use. In Claude: check current settings for conversation history and data use options. This doesn’t eliminate all data retention but reduces the most significant ongoing risk.
Check your organisation’s AI use policy before using work data. Many organisations have established policies on which AI tools are approved for which types of data. Using unapproved tools with sensitive data creates personal liability beyond the data privacy risk itself. If no policy exists, the absence of prohibition doesn’t mean approval — apply professional judgment about what your employer would consider appropriate.
Apply the same standard you apply to other cloud services. You wouldn’t paste customer data into a random web tool without checking its data handling practices. AI tools warrant the same scrutiny — arguably more, given that many people paste more into AI tools than into any other single service.
The regulatory context — GDPR, HIPAA, and industry-specific obligations
Data privacy obligations around AI tool use aren’t just about avoiding embarrassment — they have legal force in many contexts, and the penalties for violations are significant.
GDPR (European Union): If you process personal data of EU residents, GDPR applies regardless of where your business is located. Using a consumer AI tool to process EU residents’ personal data without a lawful basis for processing and appropriate safeguards is a GDPR violation. Under GDPR, AI tool providers you use to process personal data are data processors, and you need a Data Processing Agreement (DPA) with them. Consumer-tier AI tools typically don’t provide DPAs; enterprise tiers do.
HIPAA (United States healthcare): Processing Protected Health Information (PHI) with AI tools requires the AI tool provider to sign a Business Associate Agreement (BAA). Most consumer AI tools do not sign BAAs. Microsoft 365 Copilot, Nuance (Microsoft) for healthcare, and some specialised healthcare AI vendors do. Any healthcare provider or healthcare adjacent business using AI tools to process patient information needs to verify BAA availability before use.
Financial services regulation: Financial services firms are subject to various regulations (FINRA, SEC, FCA, and others depending on jurisdiction) that govern how client data is handled and shared. Consumer AI tool use with client financial information may violate these regulations regardless of whether the AI tool itself is secure.
The regulatory question is worth explicit legal advice for organisations in regulated industries — the analysis of what is and isn’t compliant depends on specific facts, specific regulations, and specific AI tool contracts that this guide cannot substitute for. Our guide on using AI tools safely at work covers the organisational and policy dimensions of AI data privacy. Our guide on ethical use of AI tools covers the consent and transparency dimensions beyond strict legal compliance — including the question of whether data subjects would consider your AI tool use appropriate even if it’s technically legal.
What good AI data hygiene looks like in practice
Good data hygiene around AI tools isn’t complicated. It mostly comes down to applying the same professional standards to AI tools that you’d apply to any other third-party service, and being honest with yourself about whether what you’re putting into these tools is appropriate to share with a third party.
The mental check that I now apply before pasting anything work-related into a consumer AI tool: “Would I be comfortable if my client/employer/regulator could see exactly what I just put into this tool and why?” If yes, proceed. If the honest answer is anything other than yes, either anonymise the content or use an enterprise tool with appropriate data agreements — or find a different approach to the task entirely.
Most AI tool data privacy incidents in professional contexts aren’t the result of malicious intent. They’re the result of convenience overriding judgment at a moment when a task is urgent, the AI tool is open, and the specific data handling implication wasn’t top of mind. Building the habit of the pre-paste check — before convenience can override judgment — is the most practical data privacy protection available to individual AI tool users.
AI data privacy for specific professional roles
The general framework above applies across all professional contexts. A few specific roles where the implications deserve more explicit attention:
Lawyers and legal professionals: Attorney-client privilege is a foundational professional obligation, and sharing privileged communications with a third party — including an AI tool — can destroy the privilege. Legal professionals using AI tools for work involving client matters need enterprise agreements that include explicit treatment of privilege and confidentiality, and should treat any consumer AI tool use with client information as off-limits regardless of how useful the AI assistance might be. Bar association guidance on AI tool use is evolving quickly; check your relevant bar’s current guidance before relying on AI tools for client work.
Healthcare professionals: The HIPAA analysis above applies, but the practical implications deserve stating plainly: transcribing patient information into an AI tool without a BAA in place is a HIPAA violation. This includes de-identified information that has been incompletely de-identified — if there’s any realistic possibility of re-identification, HIPAA’s safe harbour standards apply. Healthcare professionals should assume consumer AI tools are not appropriate for any patient-related information.
HR professionals: Employee data is personal data. Using AI tools to draft performance reviews that include specific employee situations, analysing HR data using AI tools, or processing recruitment information in AI tools creates data privacy implications for employees — who typically have not consented to their information being processed by third-party AI services. HR AI tool use requires specific attention to the employee data protection obligations of the organisation’s jurisdiction.
Financial advisors and accountants: Client financial information is subject to professional confidentiality obligations and, in many jurisdictions, specific regulatory requirements about data handling. The convenient shortcut of pasting client financial details into an AI tool for analysis or reporting creates obligations that may not be easily satisfied by consumer-tier AI tools.
The common thread across all of these professional roles: the professional confidentiality obligation attaches to the information, not to the medium through which it’s processed. Information that cannot be shared with unauthorised parties cannot be pasted into an AI tool that constitutes an unauthorised party for that purpose. Enterprise AI tiers with appropriate contracts are the professional-grade solution; consumer tiers are a personal productivity tool, not a professional services infrastructure. Our guide on AI Tools Limitations and Risks Every User Should Understand covers an adjacent issue.
