What not to paste into AI: the checklist and safer alternatives
A password, an IBAN, a sick note, a figure not yet published: some data has no business in a consumer AI tool. Here is the list, and ways of working without it.
End of the month, accounts department. Overdue clients need chasing, and the spreadsheet refuses to calculate the days overdue correctly. The reflex: export the two hundred rows from the invoicing software, paste them into a consumer AI tool and ask "find me the right formula". The export contains the clients' names, their addresses, their bank details, the amounts owed, and a comments column where someone has written "on sick leave, call back in March".
To write a formula, the AI needed three things: the column names, the date format and the expected result. Everything else went out for nothing.
That is the central idea of this guide: AI almost always needs the structure of a problem, very rarely the data itself. Here is what should never go out, followed by how to get the same help without sending it.
The checklist: seven types of data to keep to yourself
1. Access credentials
Passwords, login codes, API keys, authentication tokens, two-factor codes. People paste them without thinking when asking for help with a configuration file or an error message. No tool needs them to help you, not even a confidential one: replace them with [PASSWORD] or [KEY]. And if one of them has already gone out, change it.
2. Official and bank identifiers
Social security or national insurance number, passport or identity card number, tax number, IBAN, card number. They identify a person with certainty and are used for identity theft. An AI never needs them to write, calculate or rephrase.
3. Health data and other sensitive data
Sick notes, diagnoses, disabilities, but also political or religious opinions, trade union membership, sex life, criminal convictions. The GDPR and the Swiss Federal Act on Data Protection (FADP) both treat these as sensitive data, with stricter rules. They often hide where you least expect them: a comments column, a line in meeting minutes, a forwarded email.
4. Other people's personal data
Names, addresses, phone numbers and email addresses of clients, candidates, colleagues or members of the public. These people have not authorised you to hand them over to a third-party service. A CV, an annual review or a customer complaint is full of them.
5. Anything covered by secrecy
Professional secrecy, medical confidentiality, official secrecy: breaching them is a criminal offence in Switzerland (Swiss Criminal Code Art. 320 and 321), as it is in many other countries. For the professions concerned, the question is not even one of risk, but of what the law allows. See our guide for lawyers or for public sector employees.
6. Non-public company information
Results before they are announced, a planned acquisition, a reorganisation not yet communicated, a commercial offer under negotiation, source code, an internal audit report. Trade secret law and the confidentiality clauses in your contracts often cover these documents, including those a client has entrusted to you.
7. Screenshots and files attached "for context"
A screenshot taken to show a bug also shows your inbox, the name of the client open in the next tab, sometimes an entire table. A document attached "for context" usually contains far more than the useful passage. Before attaching anything, look at everything that is visible, not just the part you care about.
How to work anyway: three techniques
Describe instead of copying
For the spreadsheet formula, the useful request fits in a few lines, without a single piece of real data:
I have a table with the columns "Client", "Invoice date" (DD/MM/YYYY format), "Due in days" (number) and "Amount". Write a formula that calculates the number of days overdue as of today, shows 0 if the invoice is not yet due, and explain how to extend it to the whole column. Also give me a conditional formatting rule that highlights in red anything more than 30 days overdue.
If the AI needs examples, make up three fictitious rows in the same format. The result is the same.
Replace with consistent markers
When the wording matters, as in an email or a memo, replace each identifying element with a stable marker: [CLIENT A], [COMPANY B], [AMOUNT]. Keep it consistent (the same person keeps the same marker throughout) and keep what matters for the reasoning: "an industrial SME with forty employees" tells the AI more than a proper name. For figures, fictitious values in the same proportions are enough to get the right analysis.
Before: the reminder email pasted as it is, with the client's name, her contract number, her bank details and the mention of her sick leave.
After: "Friendly reminder for [CLIENT A], two invoices 45 days overdue, a loyal client for six years who has told us about personal difficulties. Kind tone, offer to spread what is owed over several instalments." The AI writes the same letter; you put the real details back in your email client.
Switch tools when the content itself is sensitive
Some tasks cannot be described: reviewing a reorganisation memo, summarising a file, rewording minutes where every detail counts. For those, you need a tool where the text does not go to a third party.
In IA Confidential, questions are handled by default by open models installed on servers in Switzerland, under the Swiss Federal Act on Data Protection, outside the reach of the CLOUD Act. These confidential models pass nothing on to third parties. The tool offers three families of models: local, specialised and external; the choice is automatic or manual, and external models are only used if you choose them.
If you choose an external model, the Confidentiality filter examines your message first. If it detects sensitive data, it offers to answer with a confidential model, your message untouched, or to anonymise it for the external model, with a version you can review and correct before it is sent. Sending it as it is remains possible, but that is your decision, made with full knowledge of the facts.
Attachments (text, PDF, Word, images) are read in your browser. Only pseudonymised extracts are sent to an external model. And if a document is still too identifying even once masked, the assistant tells you: it is then handled by a confidential model, or only your question is sent.
Finally, in the Privacy panel, you can write your own anonymisation instructions: your company's name, a project code name, how to handle amounts. These are precisely the things no automatic detector can guess on its own.
The limits to be aware of
A filter detects; it does not understand your line of work. An internal project name, a client's nickname, a reference that looks like nothing in particular can slip through: that is what your anonymisation instructions are for, along with your review of the anonymised version.
AI can also get things wrong or make them up: a formula that works on three rows but not on two hundred, a figure miscalculated. Test, check, reread. Whatever goes out under your name remains your responsibility.
And a confidential tool does not replace your employer's rules. If your IT policy forbids any tool that has not been approved, raise the need rather than work around it. For the general habits, read our guide to using AI at work safely.
What stays confidential
With a confidential model, your question, your attachments and the answer are processed on servers located in Switzerland, and nothing is kept on our side: your conversation history stays in your browser. We could not hand it over if asked, because we do not have it.
With an external model, nothing goes out without passing through the Confidentiality filter, and you choose between the anonymised version, pseudonymised extracts or sending the message as it is.
One precaution remains in your hands: since the history is in your browser, anyone who uses the same session on the same computer can read it. On a shared computer, log out when you leave. This is also why the rule on access credentials applies everywhere: a password has no business in a conversation, even a confidential one.
To try the "describe instead of copying" technique on your own spreadsheet: open the chat.
