How to do it
The AI doesn't need the confidential document in full. It needs the type of document and what you have to do with it. The sensitive part (names, amounts, addresses, numbers) is precisely what you can remove without losing the help.
Decide what is really confidential. Usually it isn't the whole text: it's the data that identifies people or companies and the figures. The structure of a clause, the layout of a proposal, the tone of a letter are not secrets.
Anonymize before pasting. Replace the sensitive data with stable placeholders: "Client A," "Supplier B," "amount X," "city Y." Keep a little table of your own on paper or in a local file that links the placeholders to the real values. That table never goes into the AI.
Ask for help on the cleaned version. The AI reasons on the structure and gives you back the improvement; you re-enter the real data by hand in the original document.
For the most delicate cases, raise the wall. Turn off training, open a temporary chat, and consider an AI that runs on your computer (see below). The more sensitive the data, the more layers you add.
If you really must give a whole document and can't clean it, there's the local route: an AI that runs on your computer without sending anything over the network. Tools like Ollama or LM Studio download a model and run it locally; the text doesn't leave your machine. It's slower and less brilliant than the online services, but for absolute confidentiality it's the only one where the document doesn't leave the house.
The operational syntax for working on an anonymized document:
I'm pasting you a [type of document, e.g. supply contract] from which I've already removed the real data,
replaced with placeholders like "Client A", "amount X", "date Y".
Help me [goal: e.g. check that the payment clauses are clear and complete].
Don't ask me for the real values and don't try to guess them: work only on the structure.
Text: [paste the anonymized version]
A concrete example
Elena needs to have a consulting contract reviewed before signing it. She doesn't want to upload the client's name and the fee. She opens the file, replaces the name with "Principal A," the fee with "amount X," the dates with "date 1, date 2." She pastes only this version into a temporary chat with training off. The AI flags an ambiguous termination clause and an unclear penalty. Elena opens the real contract on her PC and corrects only those two points. The original document never left her computer.
When it does NOT work (and how to fix it)
If the document is useless once the data is removed
Sometimes the numbers ARE the problem (a balance sheet, a quote to be verified in its figures). Fix: keep the numbers but remove what ties them to an identifiable person or company; or, if you need the real figures, use a local AI (Ollama, LM Studio) so the numbers stay on your device.
If anonymizing by hand is too long
For short documents it's fine by hand. For long, repetitive files you risk forgetting a piece of data. Fix: do a first pass of replacement with find-and-replace in your text editor, then reread before pasting. There are also dedicated anonymization tools that run locally.
If the AI tries to guess the hidden data
It happens that the model "fills in" the placeholders with invented values. Fix: write into the prompt the explicit instruction not to guess the real data (as in the block above) and ignore any value the system proposes in place of the placeholders.
A tip from someone who actually uses it
Think in terms of risk levels, not with a single rule. For a generic email draft you can even paste the text after turning off training. For a contract with third-party data, anonymize. For documents covered by professional secrecy or health data, go with a local AI and let nothing out. The level of protection is dosed according to how serious it would be if that data ended up out there.
Frequently asked questions
If I use the temporary chat, can I upload the confidential document as is?
Better not. The temporary one avoids saving and training, but the document still passes through the provider's servers during processing. For third-party data, anonymize anyway; the temporary chat is one more layer, not a free pass.
Is a local AI complicated to install?
Less than it seems. Tools like LM Studio have an interface similar to the online apps: you download the program, choose a model from the list, and type. The advantage is that everything stays on your computer. The drawback is that you need a fairly powerful PC and the answers are less refined.
Anonymizing is overkill, serious companies don't look at my chats anyway, right?
That's the reasoning that lowers your guard. The risk isn't the employee who spies: it's that another person's data, once inside a system that trains itself, is no longer under your control, and that you're handling data you don't have the right to hand over. Anonymizing isn't distrust toward the company, it's respect toward those who entrusted you with that data.