"It's only a rewording, there's nothing confidential in it." That is often where professional secrecy gives way. The five most common mistakes before handing your work to AI, and how to avoid them.
Tuesday, 6 pm. At an accounting firm, an associate has to reply to a worried client: the client's company has just received notice of a tax audit. The email is delicate, so she pastes it into a consumer AI tool and simply asks it to fix the tone. "It's only a rewording, there's nothing confidential in it."
On the contrary, it's all there: the name of the company, the fact that it is a client of the firm, the audit under way, the owner's anxiety. It is the most widespread mistake among professionals bound by secrecy, and it is not the only one.
Professional secrecy does not only protect serious confidences. It covers what professionals learn in the course of their work: the facts, the documents, and often the very existence of the relationship with the client, patient or member of the public. In Switzerland, Article 321 of the Swiss Criminal Code makes its breach a criminal offence for the professions it lists, and Article 320 punishes breaches of official secrecy by public servants. Most other countries have comparable rules, set out in criminal law or in professional codes of conduct.
Lawyers, doctors and nurses, notaries, psychologists, social workers, public servants: each profession has its own rules and often its own professional body, which publishes its own recommendations on AI. They all share one point: the duty of secrecy rests on you, not on the tool. If the text ends up with a third party that keeps it, you are the one who answers for it.
This is the mistake from the introduction. You ask the AI to fix the tone, the spelling or the phrasing, yet the text to be fixed is the case file itself.
The consequence: the whole message goes to a provider you did not choose, and is often kept in the service's history. The disclosure happens the moment you send it, whatever the provider does with it afterwards.
How to avoid it: for this kind of task, work with a confidential model. In IA Confidential, the confidential models, Standard and Specialised, are open models installed on servers in Switzerland and Europe, outside the reach of the CLOUD Act, and they keep nothing of your conversations. They are the default choice. And when only the tone matters, the situation can be described in one sentence:
Rewrite this email so that it is reassuring without promising anything: a client has learned that an audit is going to take place and is asking me what they risk. Keep it short, free of jargon, and end with an offer to meet.
"A notary in a village of three thousand people, an estate that includes a vineyard": there is no name any more, but there is no doubt either.
The consequence: the text is still personal data and the case is still recognisable. Secrecy covers the content, not just the identity. The trap is explained in detail in our guide to anonymisation vs pseudonymisation.
How to avoid it: if the context alone identifies someone, stay on a confidential model, with the message untouched. If you choose an anonymised model for a general question, the Confidentiality filter first detects the sensitive data and offers to answer with a confidential model, or to send an anonymised version that you review and correct. In the Privacy panel, your anonymisation instructions add what no detector can guess: the name of your practice, a case's nickname, an internal reference.
"I turned off the history", "the service says it keeps nothing": you are handing your duty of secrecy over to a ticked box on someone else's system.
The consequence: you cannot check what is actually kept, the terms change, and a provider subject to foreign law can be compelled to hand over what it holds.
How to avoid it: choose a setting where there is nothing to hand over. In IA Confidential, your conversations are kept in your browser, never on our servers: if anyone asked us for them, we would have nothing to give. Confidential models process your message without keeping it, and an anonymised model is never chosen on your behalf.
A PDF, a scan or a photo carries more than the passage you are interested in: appendices, numbers, names in the letterhead, sometimes the GPS location of an image.
The consequence: for a question about one clause, the entire contract goes out, with the parties and the amounts.
How to avoid it: attachments (text, PDF, Word, images) are read in your browser, and no file is stored on the server's disk. When an anonymised model is used, only pseudonymised extracts go out, and images are stripped of their metadata, including the GPS location. If a document remains identifying even once masked, the assistant flags it as an "Identifying document": it is then handled by a confidential model, or only your question is sent. For a whole case file, on a computer with Chrome, Edge or Opera, you can add the folder from your computer to a space with the "Confidential models" option: it is then never sent to an anonymised model. One space per client or per matter also keeps cases from getting mixed up in the same conversation.
In the attachments, find every deadline (time limits, settlement dates, expected reply dates). Present them in a table with the document and the page each date comes from, and point out any that contradict each other from one document to another.
In the end, you start to trust it: you copy the answer into a formal document, you send a message without rereading it, you assume the tool has masked everything.
The consequence: AI can get things wrong and make them up with total confidence: a date, a statutory provision, a fact that is not in the file. A detector can also let a detail slip through. Either way, it is your signature on the line.
How to avoid it: reread everything that comes out, check every reference against your sources, and keep control over what goes to an anonymised model. AI can even help you prepare a general question with nothing from the case file in it, on a confidential model first:
Here is my client's situation. Rewrite it as a question of principle, without any fact specific to the case: no dates, no amounts, no places, no occupation, no family relationship. I want to be able to ask it as it is to an anonymised model. Then list everything you removed, and tell me whether the question loses anything essential.
With a confidential model, your messages, your attachments and the answers are processed on servers in Switzerland and Europe, without being kept, and your history stays in your browser. Folders added with the "Confidential models" option are never sent to an anonymised model.
With an anonymised model, nothing goes out without passing through the Confidentiality filter, and you decide: the anonymised version you reviewed, pseudonymised extracts, or the message as it is if you choose to take responsibility for it. Whatever goes out is then processed by that model's provider, under its own rules.
On a shared computer, anyone using the same session can read the history: keep a personal session and log out when you leave.
For a list of what should never go out, even masked, read what not to paste into AI. For a method tailored to a profession, see for example our guide for lawyers. Then try it out on a fictitious case first: open the chat.
