Someone on your team wants to paste a spreadsheet of customer details into ChatGPT to clean it up. It would take four minutes instead of an afternoon, and they are right about that.
The question is whether you are allowed to, and the honest answer is: sometimes, under conditions, and not in the way most people are currently doing it.
This is not legal advice. It is the practical shape of the problem, which is usually enough to make a sensible decision or to know that you need someone qualified.
Why this is not a simple yes or no
Putting personal data into a chatbot is processing personal data. GDPR does not care that the interface looks like a chat window rather than a database.
So the usual questions apply, and they are the same ones that apply to any other system you use. What is your lawful basis. Have people been told. Where is the data going. How long does it stay. Who else can see it. Can you delete it when asked.
Most businesses have solid answers to those for their CRM and no answers at all for the chatbot, despite putting similar data into both.
The distinction that decides it
Not all accounts are the same product, and this is the single most misunderstood point.
A free or personal consumer account is a consumer service. Historically these have used conversation content to improve the underlying models unless you actively turn that off, and there is no contract between the provider and your business governing that use. Putting customer data in is, in practice, disclosing it to a third party under terms you never negotiated.
A business, team or enterprise tier is a different arrangement. These generally exclude your content from model training by default, and come with the contractual terms a business actually needs to rely on, including a data processing agreement.
That gap is the whole answer. The same paste, the same data, one account type defensible and the other not.
Terms change, so verify the current position for the specific tier you are on rather than trusting this article or a colleague’s memory. But the principle holds: if you have not signed something, you have not agreed anything.
The consent myth
The most common thing said about this is also the most commonly wrong: that you need a customer’s explicit consent before anything about them touches an AI tool.
There is no such rule. The EU AI Act does not impose one, and GDPR does not either. What GDPR requires is a lawful basis, and consent is only one of six:
- Consent
- Performance of a contract
- Legal obligation
- Vital interests
- Public task
- Legitimate interests
For most ordinary business processing, contract or legitimate interests is the honest answer rather than consent. Consent has to be freely given, which makes it a poor fit where the person has no real choice, and it can be withdrawn, which makes it a fragile foundation for something you need to keep doing.
Reaching for a consent checkbox because it feels safest usually creates a worse position than picking the basis that actually describes what you are doing. It is also worth keeping separate things separate: AI Act transparency, GDPR transparency, a GDPR lawful basis, cookie consent, and employee monitoring rules are five different questions that get discussed as though they were one.
The four questions worth answering
- Is there a data processing agreement? If you are processing personal data through a provider, you need one. Consumer tiers do not come with it.
- Is your content used for training? On business tiers, generally no. On consumer tiers, historically yes unless switched off. Check the setting rather than assuming.
- Where is it processed? Transfers outside the EEA need a proper mechanism. Reputable providers have one. It is still your job to know it exists.
- Have you told people? Your privacy notice describes what you do with customer data. If it does not mention AI processing, it is now incomplete.
That last one catches nearly everybody, and it is the cheapest to fix.
What Ireland adds
Irish businesses fall under GDPR as enforced by the Data Protection Commission, and the DPC is not a passive regulator. It supervises several of the largest technology companies operating in Europe, which means enforcement thinking here tends to be well developed rather than theoretical.
The practical consequence for a small business is not that you are likely to be investigated. It is that “everyone does it” is a weak position in a jurisdiction where these questions get examined properly.
Separately, the AI Act’s literacy obligation now expects staff to understand the tools they use. Knowing what may and may not be pasted in sits squarely inside that.
What to actually do
Four steps, none expensive.
- Move business use onto a business account. The cost per seat is small against the exposure it removes, and it is the single highest-value change available.
- Write the rule down. What may go in, what never may. Customer names and contact details, financial records, anything under a confidentiality agreement: those belong on the never list unless you have specifically established otherwise.
- Strip identifiers first. Most tasks do not need real names. If you want text improved, the tool does not need to know whose text it is.
- Update your privacy notice. One honest paragraph describing that AI tools are used and what that means for personal data.
Step three is the one people skip and the one that resolves most cases. A remarkable proportion of “can we put this in” questions disappear once you ask whether the identifying parts are needed for the task at all.
When the answer is a different tool
If the useful work genuinely requires real customer data at volume, a public chat window is the wrong shape for it regardless of tier. That is a case for processing inside systems you already control, where the data does not leave an environment you have already assessed.
That is a build rather than a subscription, and choosing between the two is a decision worth making deliberately rather than by default.
The honest summary
Yes, with a business account, a written rule, identifiers removed where they are not needed, and a privacy notice that reflects reality.
No, if the account is a personal one, nobody has agreed anything in writing, and the data is going in because it is quicker.
If the honest position is that this has already been happening for months, that is extremely common and not a catastrophe. It is worth fixing forward rather than worrying backward, and describing what has been going in is a reasonable place to start. How AI work is approached here is set out separately.
Common questions
Is it illegal to put customer data into ChatGPT?
Not inherently. It is processing personal data, so it needs the same things any other processing needs: a lawful basis, a data processing agreement, and a privacy notice that reflects what actually happens.
Does a paid account fix it?
A business or enterprise tier is the substantive difference, because those exclude your content from training by default and come with the contractual terms a business can rely on. A personal paid account is still a consumer service.
What if this has already been happening for months?
Extremely common, and not a catastrophe. Move business use onto a business account, write down what may and may not go in, and update the privacy notice. Fix forward rather than worry backward.
Can we avoid the question altogether?
Often, yes. Most tasks do not need real names. If you want text improved, the tool does not need to know whose text it is, and stripping identifiers removes the problem entirely.