It's a Saturday afternoon in early April, and a staff accountant is working from her kitchen table. She's stuck on a footnote in a client's K-1. She opens the AI app on her phone, the free one she uses for meal planning, pastes in the footnote along with the client's name and the partnership's details, and asks what it means. The explanation helps. Nobody at the firm will ever know it happened.
Was that safe? The honest answer is that it depends on things she didn't check: which account she used, what its terms say about training and retention, and whether her firm's obligations allowed that disclosure at all. A paralegal pasting a deposition excerpt or a front desk pasting a patient's message faces the same three questions.
Consumer accounts and business accounts are different products
The biggest difference isn't the model. It's the contract and the settings around it. Here's what the major providers publish about using your data for training, as of November 2026. Terms change often, so check the current version for the exact plan you use.
| Provider | Personal accounts | Business plans and API |
|---|---|---|
| OpenAI (ChatGPT) | Free, Plus and Pro conversations may be used to train models unless the user turns off the setting in Data Controls | Business, Enterprise and API data not used for training by default |
| Anthropic (Claude) | Each Free, Pro or Max user decides in settings whether their chats help train models; saying yes extends retention to five years | Commercial products are not covered by the consumer terms |
| Google (Gemini in Workspace) | Not covered here; check the consumer app's privacy settings before any work use | Workspace content not used to train generative models without the customer's permission |
| Microsoft (Copilot with a work account) | Not covered here; check the consumer app's privacy settings before any work use | Prompts, responses and Microsoft 365 data not used to train foundation models |
Training is only one question. Retention (how long prompts and files are kept), human review (who at the vendor can read them, and when), subcontractors and storage location matter as much. A business plan that doesn't train on your data but keeps it indefinitely, with no admin controls, is still a weak home for client files.
Your professional duties don't change because the tool is new
Each profession already has rules about who may see client information, and AI tools fall under them. Consider this general information, not legal advice; state rules differ.
Law firms
ABA Formal Opinion 512, issued in July 2024, works through how existing ethics duties apply to generative AI. On confidentiality, it says a lawyer needs the client's informed consent before entering information about the representation into a tool that could disclose it, such as one that learns from what users type, and that boilerplate in an engagement letter isn't enough. State bars have added opinions of their own, so check yours too.
Accounting and tax firms
Section 7216 of the Internal Revenue Code limits how tax preparers use and disclose return information, and uses beyond preparing the return generally require the client's written consent in the form the regulations set out. The regulations do allow some disclosures, without separate consent, to service providers that help prepare the return, under specific conditions, so check with your adviser before assuming either way. The FTC Safeguards Rule, which lists tax preparation firms among the businesses it covers, requires a written information security program and oversight of service providers, including contracts that spell out security expectations. An AI vendor is a service provider.
Medical and dental practices
Under HIPAA, a vendor that creates, receives, maintains or transmits protected health information for your practice is a business associate and needs a Business Associate Agreement. Confirm the agreement covers the specific feature you'll use: Google, for example, says Gemini in Workspace can support HIPAA workloads, but that Gemini Notebook (formerly NotebookLM) is not covered by its BAA. HIPAA's minimum necessary standard applies too, so share only what the task needs.
Can this go into this AI tool? A decision path
Flowchart
Before pasting client information into an AI tool
- StartA staff member wants AI help with a task that involves client information
- DecisionIs the tool on the firm's approved list, under a business account?NoA personStop: use an approved tool, or strip every client detail firstYes
- DecisionAre the terms this data needs in place, or are none needed?No or unsureA personStop and ask the privacy lead: a BAA, 7216 consent or client consent may be required firstYes
- StepShare only the fields the task needs; leave out SSNs, account numbers and unrelated history
- StepRun the task in the firm account, where it's logged under the staff member's name
- DoneStaff member reviews and edits the output before it goes anywhere
Before and after
Personal accounts vs an approved firm workspace
Personal AI accounts
- 01Staff use whatever app is on their phone
- 02Training setting left at its default
- 03Full client details pasted in to get a better answer
- 04No record the firm can see, export or delete
- 05No contract covering client data
ResultExposure the firm can't measure or explain
Approved firm workspace
- 01One business account, with no training on firm data by contract
- 02Retention and admin access set by the firm
- 03Staff share only what the task needs
- 04Activity logged and tied to each user
- 05Vendor listed in the security plan, with a BAA where needed
ResultClient data used under terms the firm has read
The vendor checklist
Before any client data goes into a tool, get written answers to these twelve questions. A vendor that can't answer them in writing has answered them.
- 01Is our data used to train or improve any model, and is that excluded by contract rather than only by a setting?
- 02How long are prompts, files and outputs kept, and can we shorten that or delete on request?
- 03Who at your company or your subcontractors can read our data, and in what circumstances?
- 04Which subprocessors handle our data, and in which countries is it stored?
- 05For patient data: will you sign a BAA, and which features does it cover?
- 06Can we control access by user and role, and remove someone the day they leave?
- 07Do you support single sign-on and multi-factor authentication?
- 08Is there an audit log of who asked what, and can we export it?
- 09How and how quickly will you tell us about a security incident?
- 10Is data encrypted in transit and at rest?
- 11What happens to our data when we cancel?
- 12If the tool connects to our drive or email, does it respect existing permissions?
Rules for staff, starting Monday
Even before you pick a tool, a short written policy closes most of the gap. Five lines are enough to start:
- Approved tools only for anything involving a client, patient or matter. Personal accounts are for personal use.
- Minimum necessary. Remove names, numbers and identifiers whenever the task works without them.
- No whole-file uploads of returns, charts or case files unless the tool is approved for that kind of file.
- A person reviews every output before it reaches a client, a court, a payer or the IRS.
- Report slips the same day. If something went into the wrong tool, telling the privacy lead quickly is how you limit the damage.
If client data has already gone into a personal account
- 01Write down what was shared, when, in which app and under which account.
- 02Delete the conversation and turn off the training setting if the app offers one; note that deletion may not undo use that already happened.
- 03Tell the person responsible for privacy at your firm, who decides with counsel whether any notification duty applies.
- 04Use it as the example in your next staff reminder, without naming anyone.
Our AI policy template turns these lines into a one-page document. If you want an assistant built on approved business accounts from the start, see our custom AI chatbots and assistants.
Sources
- OpenAI Help Center: How your data is used to improve model performance
- OpenAI: Business data privacy, security, and compliance
- Anthropic: Updates to Consumer Terms and Privacy Policy
- Google Workspace Help: Generative AI in Google Workspace Privacy Hub
- Microsoft Learn: Data, Privacy, and Security for Microsoft Copilot
- American Bar Association: Formal Opinion 512, Generative Artificial Intelligence Tools (July 29, 2024)
- IRS: Section 7216 information center
- eCFR: 26 CFR 301.7216-2, permissible disclosures or uses without consent of the taxpayer
- FTC: FTC Safeguards Rule, What Your Business Needs to Know
- HHS: Business Associate Contracts (sample provisions)
- eCFR: 45 CFR 164.502, uses and disclosures of PHI, including the minimum necessary standard