Guide for small firms · Updated November 2026

What not to paste into ChatGPT: a one-page guide for small firms

Eight kinds of information to keep out of AI chatbots, a safer way to ask the same question, and what to do when someone has already pasted the wrong thing.

Short answer: Don't paste passwords, ID or account numbers, client matters, patient information, tax return data, staff records, contracts under NDA, or unpublished business information into ChatGPT or any AI tool your firm hasn't approved. Ask the same question with made-up details instead. Public information and general drafting are usually fine.

This is general information, not legal, HIPAA or tax advice. Your counsel or compliance lead should decide what applies to your firm, and vendor terms change, so check them for your exact plan.

The rule that fits on a sticky note

If it belongs to a client, a patient or an employee, or it could unlock something, it doesn't go in an AI tool the firm hasn't approved in writing. When in doubt, swap in made-up details or ask the owner of your AI policy.

The one-page list

Type	Don't paste	Do this instead
Logins and keys	Passwords, API keys, one-time codes, screenshots that show them	Nothing. Describe the problem with the secret left out.
ID and account numbers	Social Security numbers, driver's license numbers, bank and card numbers	Use a placeholder such as [SSN] and fill in the real value yourself, outside the tool.
Client matters	Client names and facts, privileged emails, settlement figures, the other party's details	Ask the general question with invented facts, in your approved tool.
Patient information	Chart notes, appointment lists, billing records, intake forms	Nothing, unless your compliance lead has approved the tool and the required agreement is signed.
Tax return data	W-2s, 1099s, organizers, draft returns, anything a client gave you to prepare a return	Ask about the rule or the process, with no taxpayer details. Get counsel's view before going further.
Staff records	Reviews, pay, medical leave, complaints, résumés	Draft from a blank template, then add the specifics yourself.
Contracts under NDA	Vendor or client agreements with a confidentiality clause, anything marked confidential	Check the clause first. Ask about a clause type using your own template.
Unpublished business information	Financials, pricing, client lists, acquisition talks, unreleased plans	Use made-up numbers to get the structure, then enter the real ones in your spreadsheet.

Why this matters: what the vendors say

You don't have to take my word for it. Vendors say much of this themselves, and the answer depends on the plan.

Google Gemini, personal accounts. Google says some chats are read by human reviewers and warns: "Please don't enter confidential information that you wouldn't want a reviewer to see." Reviewed chats can be kept for up to three years and aren't removed when you delete your activity. With Keep Activity off, Google says future chats aren't used for training, but they are still saved for 72 hours.
Microsoft Copilot, personal accounts. Microsoft's privacy FAQ says it uses Copilot conversations to train models unless the user opts out or falls in an excluded group, such as people signed in with a work account. It also says an opt-out of human review is not available. Microsoft updated the app in August 2026 and notes that this section covers the older version, so check the current page.
ChatGPT. OpenAI publishes its own data-use policy and retention rules. Read them for the plan your people actually use, not the plan you assume they use.
Business plans are different, but not a free pass. Google says Workspace content isn't human reviewed or used for model training outside your domain without permission. Microsoft says prompts and responses in Copilot for organizations aren't used to train foundation models. Microsoft also notes that its HIPAA and agreement coverage doesn't extend to web search queries, so even a good plan has edges.

The practical point: training is only one of the questions. Ask who can read it, how long it's kept, and whether you have a contract with the vendor. A personal account gives the firm no say in any of those.

Why deleting the name isn't enough

People tend to treat redaction as a cure. It usually isn't. Dates, street addresses, job titles and unusual facts can point to one person, especially in a small community.

HIPAA shows how high the bar is. Its safe-harbor method for de-identifying health information (45 CFR 164.514) lists 18 kinds of identifiers to remove, including dates more specific than a year, most ZIP code detail, phone numbers, account numbers and full-face photos. Removing a name is one of eighteen.

A safer habit is to write the question fresh rather than editing the real document.

Instead of pasting the file (made-up example): "Here is Maria's termination letter and her last two reviews. Make the tone kinder."

Ask this: "Write a respectful two-paragraph letter ending employment for repeated attendance problems after two written warnings. Use [NAME] and [DATE] as placeholders." You get the same help, and nothing about Maria left the building.

Rules that already apply to your firm

Most of the rules here were written long before chatbots. They still apply, and how they apply to AI is something to settle with your own advisors.

Healthcare practices. HIPAA lets a covered entity share protected health information with a business associate only with satisfactory assurances, documented in a written contract (45 CFR 164.502(e)). If no agreement is signed with the vendor, talk to your compliance lead before any patient information goes near the tool.
Tax preparers. Section 7216 makes it a crime for preparers to knowingly or recklessly disclose information furnished for a return, or use it for other purposes. The regulations define disclosure broadly and don't mention AI, so how they treat AI tools is unsettled. Get counsel's view before return data goes into any of them.
Law firms. Confidentiality duties apply to whatever goes into a tool. My guide to AI policies for Tennessee law firms covers this, and your ethics counsel should have the final word.
Everyone. Read your client contracts and NDAs. Some restrict sharing with third parties or say something specific about AI, and those promises are yours to keep whichever law applies.

What's usually fine

A list of bans with no green light just sends people to their phones. Be just as clear about what's allowed.

Anything already public: your website, published rules, press releases.
First drafts of marketing copy, job posts and general emails with no client details.
Rewriting your own general text for tone or length.
Spreadsheet formulas and how-to questions, using made-up sample data.
Brainstorming, outlines, meeting agendas and training ideas.

Everything AI writes still needs a person to check it before it reaches a client.

If someone already pasted something

It will happen, and how you respond decides whether you hear about the next one.

Stop and report it the same day to the person who owns your AI policy.
Write down the facts: which tool, which account, what was entered, and when.
Don't count on deleting the chat. It may help, but as Google's terms show, deletion doesn't always remove everything. Check the vendor's terms for that plan.
Let counsel or your compliance lead decide whether a client, patient or regulator needs to be told.
Fix the cause. If people paste because they have no approved tool, give them one.

Make it stick in one meeting

Print the list, pick one approved tool and one owner, and walk the team through it in 30 minutes. Then put it in writing with the free one-page AI policy template, which has a "never enter" section ready to edit.

Frequently asked questions

Is a paid AI plan safe for client information?

A paid plan is not enough on its own. Personal paid plans and business plans can have different terms, so read the data terms for the exact plan. Business terms from Microsoft and Google say prompts are not used to train models outside your organization, but your own confidentiality, contract and regulatory duties still apply. Decide in writing what may go into the approved tool.

Does turning off model training make it safe to paste anything?

No. Training is one question. Retention and human review are separate ones. Google says Gemini chats are still saved for 72 hours with Keep Activity off, and Microsoft's Copilot FAQ says an opt-out of human review is not available. A setting on one person's personal account is also something the firm cannot see or enforce.

Can I paste a document if I delete the client's name?

Often not. Dates, addresses, account numbers, job titles and unusual facts can still identify someone. For health information, HIPAA's safe-harbor method lists 18 kinds of identifiers to remove, and a name is only one of them. Rewrite the question with made-up or generic details instead of editing a real document.

Can staff use their personal ChatGPT accounts for work?

My advice is no, at least not for anything about clients, patients, employees or the business's finances. The firm can't see what went in, can't set the controls, and has no contract with the vendor. Give people one approved tool on a business account so they don't need the workaround.

Does this apply to Gemini, Copilot and Claude too?

Yes. ChatGPT is in the title because it's what people search for. The same list applies to any AI tool, including AI features built into email, notes and browser extensions. Check each vendor's terms for the plan you actually use.

What if someone at my firm already pasted something sensitive?

Tell the person who owns your AI policy right away and write down the tool, the account, what was entered and when. Don't assume deleting the chat removes it. Have your counsel or compliance lead decide whether anyone needs to be notified. No penalty for prompt reporting, or people will stop telling you.

Want this done for your firm? Take the free AI readiness check to see where you stand, or start with the free policy template. The AI Opportunity Audit is a fixed-price, two-week look at where AI helps and what to put in writing, from $3,500. Or book a free 30-minute call.

Sources

Vendor pages and legal texts read November 1, 2026. I am not a lawyer, and terms change.

By Dane Hale, Franklin, Tennessee. Published and last updated November 2026. General information, not legal, HIPAA or tax advice.