If you handle other people's confidential information for a living, you have probably already asked yourself some version of this question: can I paste this into ChatGPT? The document is sitting in front of you, the deadline is real, and the AI would save you an hour. The only thing stopping you is a quiet worry about whether you are allowed to.
This article is an honest attempt to answer that question under the GDPR. It is not legal advice, and we are not lawyers. What follows is a plain-language explanation of what the regulation actually requires and the questions you need to answer for your own situation. For a definitive answer about your specific case, your data protection authority or a qualified lawyer is the right source.
First, the part people get wrong
The most common mistake is treating "is it secure?" and "is it lawful?" as the same question. They are not.
A cloud AI service can be perfectly secure, encrypted in transit, encrypted at rest, run by a serious company with good engineers, and using it can still put you in breach of the GDPR. Security is about whether someone can steal the data. Lawfulness is about whether you were allowed to send it there in the first place, on what basis, with what safeguards, and whether the people whose data it is would have any idea you did.
When you put a client's document into a third-party tool, you are not just reading it anymore. You are transferring personal data to another company, the processor, who then handles it on your behalf. That transfer is a processing activity the GDPR cares about, regardless of how well the tool is secured.
What the GDPR actually requires
A few concrete obligations come into play the moment a document contains personal data, which most client documents do.
You need a lawful basis for the processing. You almost certainly had one for your own work, a contract with your client, a legal obligation, your legitimate interest. But sending that data onward to an AI vendor is a new processing activity, and your original basis does not automatically stretch to cover it.
You need a data processing agreement with the vendor. Article 28 of the GDPR requires that when a processor handles personal data on your behalf, there is a contract governing what they can do with it. Many consumer AI products are not offered with a DPA by default. Business and enterprise tiers often are. The free tier you signed up for in two minutes usually is not. If there is no DPA in place, the processing has a problem before you even consider where the servers are.
You need to think about international transfers. If the vendor processes data outside the European Economic Area, the GDPR adds another layer of requirements around how that transfer is protected. This is the area that has shifted the most in recent years, and it is genuinely complex. It is also one of the main reasons "the data goes to a server in another country" is not a detail you can wave away.
You need to honor the rights of the people in the document. Under the GDPR, individuals can ask what data you hold about them, ask for it to be corrected, and in some cases ask for it to be deleted. If their personal data is now sitting in a third party's systems, possibly retained, possibly used to improve a model, you have made those rights harder to honor, and you may not even be able to say with confidence where the data ended up.
"But they say they don't train on my data"
This line appears in a lot of AI vendors' marketing, and it is worth taking seriously, but it does not settle the question.
"We don't train on your data" is a statement about one specific use of the data. It says nothing about whether a DPA exists, where the data is processed, how long it is retained, who can access it for support or abuse monitoring, or whether the setting that disables training is on by default or something you had to find and enable. A promise not to train is a narrow promise. The GDPR asks broader questions, and a single reassuring sentence on a pricing page does not answer them.
It is also worth remembering that these settings and policies change. The configuration that was compliant when you set it up can quietly change with a product update or a new terms of service. Compliance built on "I read the settings once" is fragile.
The questions to answer before you paste
Rather than a yes or no, here is a short list you can actually work through for your own situation:
Does the document contain personal data? If it names people, describes them, or could identify them, it almost certainly does. Financial records, contracts, medical notes, and HR files nearly always do.
Do you have a DPA with the vendor for the tier you are using? Not the company in general, the specific product and plan you are about to use.
Where is the data processed, and is that transfer covered? If you do not know where the servers are, that is itself an answer.
What is your professional duty on top of the GDPR? Lawyers, auditors, accountants, and healthcare professionals often carry confidentiality obligations that are stricter than data protection law alone. The GDPR is the floor, not the ceiling.
Would your client be surprised? This is not a legal test, but it is a useful gut check. If you would not be comfortable telling the client "I ran your file through an American AI service to save time," that discomfort is information.
The option that sidesteps most of this
There is a category of tool the questions above mostly do not apply to: AI that runs on your own computer, where the document never leaves the machine.
If the file is never transmitted to a third party, there is no onward transfer to find a lawful basis for, no processor to sign a DPA with, no international transfer to assess, and no vendor retaining a copy you cannot account for. The analysis happens locally, the same way a calculation in a spreadsheet happens locally. You still have your own obligations as the data controller, but you have removed the entire category of risk that comes from handing the data to someone else.
This is the reason local AI exists as a category, and it is why we built SoloScan to work this way: the document is read and analyzed on your computer, and nothing is uploaded. We are obviously not a neutral party here, so treat that as disclosure rather than a recommendation. The general point stands regardless of which tool you use: if the data never leaves your machine, most of the cloud-AI compliance problem never starts.
Where this leaves you
The honest summary is that "can I use ChatGPT for confidential client documents?" does not have a clean yes or no answer. It depends on what is in the document, what agreement you have with the vendor, where the data goes, and what professional duties you carry. For many casual, free-tier uses on documents full of client personal data, the realistic answer leans toward no, or at least not without doing the work first.
What you should not do is assume that because a tool is popular and well-engineered, it is automatically fine for confidential files. Security is not the same as lawfulness, and a reassuring sentence about training is not the same as compliance.
If you are unsure about your specific situation, the right next step is a conversation with your data protection authority or a lawyer who knows your field. And if the simplest way to stop worrying about it is to keep the documents on your own machine in the first place, that option is available too.
This article is general information, not legal advice. Data protection rules are complex and depend on your specific circumstances. Consult your data protection authority or a qualified lawyer for guidance on your situation.