Is it safe to put client information into AI?
Putting client information into AI on a personal account is not safe: OpenAI, Anthropic, Google, and Microsoft all advise against entering confidential or sensitive data in their consumer assistants. On a business plan with a data processing agreement and training off by default, the risk is lower but not zero, and UK regulators of lawyers and accountants say that paying for a tool does not, on its own, make it safe.
Is it safe to put client information into an AI assistant?
Not on a personal account, and only with care on a business one. That is the short answer in keeping client data safe with AI, and the companies themselves give it: each warns, in its help pages, against entering confidential data in its consumer assistant. UK professional regulators go further. The Solicitors Regulation Authority has reported cases of confidential client information entered into public AI tools, and says that both free and paid tools may pose risks.
What the companies say about their own consumer tools
- OpenAI asks users not to enter sensitive information they would not want reviewed or used.
- Anthropic lists Social Security, card, and account numbers, medical records, passwords, and confidential business or personal documents as things to be careful with.
- Google asks Gemini users not to enter confidential information they would not want a human reviewer to see.
- Microsoft says users of its consumer Copilot should not provide confidential or sensitive personal data they would not want Microsoft to use.
These warnings apply to personal accounts, where training is usually on by default and human review is possible.
What a business plan changes
| Personal plans | Business plans | |
|---|---|---|
| Training on your content | Usually on by default, can be turned off | Off by default at OpenAI, Anthropic, Google, Microsoft, and Perplexity |
| Contract | Consumer terms | A data processing agreement is available or built into the terms |
| Security audits | Not the selling point | SOC 2 reports, and ISO 27001 at Anthropic, Google, and Microsoft |
| Who can read your chats | The company, in limited cases | The company in limited cases, and your administrator |
A business plan turns the AI company into your service provider under a contract, which is what privacy laws expect when you hand personal data to a third party. It does not remove the need to decide what goes in, and it gives your employer or administrator access in most cases.
What regulators of lawyers and accountants say
- US lawyers: the American Bar Association’s Formal Opinion 512 of July 29, 2024 applies the rules on competence, informed consent, confidentiality, and fees to generative AI. An ABA summary adds that boilerplate consent in an engagement letter is not enough.
- UK solicitors: the SRA says client data should only go into an AI system with appropriate contractual, technical, and organizational safeguards, in a secure environment, without being used for training unless explicitly authorized, and without being kept longer than necessary. It cites a 2026 tribunal decision that uploading letters to a tool like ChatGPT places them in the public domain, with a risk of losing legal privilege.
- UK accountants: the ICAEW says its confidentiality rules apply to AI, that entering data into a tool outside a secure environment can be like posting it publicly, and that it expects firms to have AI policies and training. Tax practitioners’ guidance adds that using a public tool without the client’s consent is likely a breach.
- US health data: a cloud provider that handles protected health information for a covered entity must sign a business associate agreement under HIPAA; that rule predates AI tools and applies to them too.
Anonymizing is not enough on its own
Removing a client’s name helps, but UK tax practitioners’ guidance warns that an AI tool may still recognize a client from details, such as an unusual service or a combination of facts. Remove identifying details, not just names, and keep the rest to what the task needs.
A traffic-light rule for your team
| Light | What it covers | Where it can go |
|---|---|---|
| Green | Public information, your own drafts without client details, general questions | Any approved assistant |
| Amber | Client documents, contracts, internal figures, personal data about clients | Only a business plan with a data processing agreement and training off, if your professional rules and your client allow it |
| Red | Health records, passwords, bank and card details, ID numbers, privileged material | Nowhere, unless a specific tool has been approved for it with a legal basis |
This rule is our recommendation, not a legal requirement. Professionals with confidentiality duties should check their own regulator’s guidance, and this page is not legal advice.
If you already pasted something sensitive
Delete the conversation, which OpenAI and Anthropic say removes it from their systems within 30 days, and do not rate it. If the data was personal data about clients, check whether your country’s breach rules apply.
Training settings, assistant by assistant
Turning off training is the first step on any personal account. See how to stop the main assistants from training on your chats.
The law in your country
What privacy law requires before you share client data with an AI provider differs by country. See AI and privacy law in the UK, EU, Canada, Australia, and the US.
Writing it down
The traffic-light rule belongs in a one-page team policy. Our AI policy template for small teams includes it.