What about screenshots of a spreadsheet?

Same thing. Modern AI tools read text out of images, and an uploaded screenshot is processed and retained under the same terms as pasted text.

What Not to Paste Into ChatGPT or Copilot: A Checklist for People Who Handle Client Data

Q: Is it OK if I delete the chat afterwards?

Deleting a chat does not reliably undo anything. On personal ChatGPT tiers, a conversation may already have been used for training before you delete it. On free Gemini, conversations that human reviewers have looked at are stored separately from your account for up to three years and are not removed when you delete your activity.

Q: Is the paid version safe for client data?

Paying does not change the data terms by itself. ChatGPT Plus and Pro are still personal tiers where training is on by default. The tiers with contractual protections are business ones: ChatGPT Team/Enterprise, the API, Microsoft 365 Copilot on a work account, and Gemini for Workspace.

Q: Does this break GDPR or my client contract?

Pasting client personal data into a consumer AI tool sends it to a third party without a data processing agreement — which is exactly what most client contracts, NDAs, and data protection laws like GDPR prohibit. Whether it becomes a reportable breach depends on the data and jurisdiction; that it was a disclosure is not really in question. This page is not legal advice.

Q: Can I still use AI for spreadsheet work at all?

Yes — ask the question with the shape of your data instead of the data. Five invented rows with the same columns get you the same formula, query, or steps. And much of what people paste spreadsheets in for — merging, dedupe, date fixes — is deterministic work a local tool can do without any data leaving the machine.

Last reviewed July 2026 · policies change — verify before relying on this

Somebody on your team pasted a client spreadsheet into ChatGPT this week. Probably to write a formula, summarize it, or "just clean it up." It felt like pasting into any other tool.

It isn't. On the personal tiers of the big AI assistants, what you paste can be kept, read by human reviewers, and used to train future models — by default, under the terms everyone clicked through. This page is the checklist we wished existed: what never to paste, what actually happens to pasted data on each tier, and how to still get the AI help without the exposure.

The one-sentence rule

If you couldn't email it to a stranger, don't paste it into a personal-tier AI chat. Everything below is detail on that sentence.

Never paste (the checklist)

1. Anything that identifies a client's customers

Names, emails, phone numbers, home addresses
National ID / Social Security / tax numbers
Dates of birth, account IDs, policy numbers

This is the classic case: a customer export pasted in "to remove duplicates." That's personal data disclosed to a third party — the thing your client contract and GDPR-style laws exist to prevent.

2. Financial records

Bank account or card numbers, even partial
Payroll and salary data tied to names
A client's revenue, margins, or unpublished financials

3. Credentials and keys

Passwords, API keys, tokens, connection strings
Anything from a .env file or a config with secrets

Treat a pasted credential as leaked: rotate it. The chat log now contains it, on someone else's servers, for as long as their retention says.

4. Regulated categories

Health information of any kind
Data about minors
HR records: disciplinary notes, medical leave, references

5. Confidential business material

Contracts, and anything covered by an NDA
M&A, legal disputes, unreleased plans
A client list itself — the names of who you work with

What actually happens to pasted data

The terms differ sharply between the personal tier most people use and the business tiers most people assume they're on. As of July 2026:

Tool & tier	Used to train models?	Worth knowing
ChatGPT — Free / Plus / Pro (personal)	Yes, by default. Opt-out toggle in Data Controls.	Giving a thumbs-up/down on a reply can submit that whole conversation for training even if you opted out.
ChatGPT — Team / Enterprise / API	No — excluded contractually (DPA).	This is the tier your company should be on if AI touches work data.
Microsoft 365 Copilot (work account)	No — prompts and responses stay inside your tenant boundary.	The strong commitments apply to the work deployment, not the free consumer app.
Copilot (free, personal account)	Not covered by the enterprise commitments.	No tenant boundary, no DPA. Different product from the work one, same name.
Gemini — free (personal)	Conversations can be reviewed by humans and used to improve models.	Chats a human reviewer has seen are stored up to 3 years, separately from your account — deleting your activity does not remove them.
Gemini for Workspace	No — workspace data excluded.
Claude — Free / Pro / Max (personal)	Yes unless you opt out (changed August 2025).	Opted-in data can be retained up to 5 years; opting out returns you to a 30-day window.

Every provider changes these terms periodically — usually in the direction of more data use, announced quietly. Re-check before relying on any row of this table.

How to still use AI — safely

Ask with the shape, not the data. You almost never need real rows to get a formula or an approach. Five invented rows with your real column names get the same answer: "columns: OrderID, CustomerEmail, Amount, Date — how do I find duplicates by email?"
Strip identity before pasting. If you must use real data, delete or scramble the identifying columns first, and replace names with row numbers. If two columns together identify someone (postcode + birth date), that combination counts too.
Get on a business tier. If your team pastes work data into AI weekly, the fix isn't a policy memo — it's ChatGPT Team/Enterprise, Microsoft 365 Copilot on work accounts, or Gemini for Workspace, where training exclusion is contractual.
Turn the toggles off anyway. On personal tiers: ChatGPT Settings → Data Controls → "Improve the model for everyone" off; the equivalent in Claude and Gemini. Belt and suspenders, not a substitute for the rules above.
Keep the mechanical work local. A lot of spreadsheet pasting isn't really an AI question at all — merging files, removing duplicates, fixing dates. That's deterministic work; nothing about it needs a cloud model. Do it in Excel, a script, or a local tool, and the question of "what happens to the data" never arises.

FAQ

Is it OK if I delete the chat afterwards? No. On personal ChatGPT tiers the conversation may have been used before you deleted it. On free Gemini, human-reviewed chats are stored separately from your account for up to three years and survive deletion of your activity.

Is the paid version safe for client data? Paying alone changes nothing — ChatGPT Plus/Pro are still personal tiers with training on by default. Contractual protection comes with the business tiers: Team/Enterprise, the API, M365 Copilot on work accounts, Gemini for Workspace.

Does this break GDPR or my client contract? Pasting client personal data into a consumer AI tool discloses it to a third party with no data processing agreement — which is what most client contracts, NDAs, and data protection laws prohibit. Whether it's a reportable breach depends on specifics; that it was a disclosure isn't really in question. (Not legal advice.)

What about screenshots? Same rules. The models read text out of images; an uploaded screenshot is processed and retained like pasted text.

Can I still use AI for spreadsheet work at all? Yes — with the shape-not-data approach above. And for the deterministic part (merge, dedupe, date fixes), use something local.

This page is maintained by the team behind Kramata, a Windows app that does spreadsheet cleanup — merging, dedupe, date fixes — entirely on your own computer, which is why we ended up reading all these data policies in the first place. The checklist stands on its own; use whatever local tool you like.