Can I upload this to AI? Classify the data before you paste it
By Kike Gandia · AI & Security · 10 min read
Someone on your team has just pasted a client contract into ChatGPT to get a summary before a meeting. Nothing unusual about it: they do it every day with internal documents and nothing has ever gone wrong. But this time the document has a confidentiality clause, and the account they used is their personal one, the one they opened with their own email two years ago. Nobody decided this was forbidden. Nobody decided it was allowed either: the question simply never came up.
The answer to "can I upload this?" almost never depends on one thing. It depends on three, and each one is governed by a different law.
Three questions, not one
Most guides on AI in the workplace answer only one of the three questions you need to ask, which is why they fall short. What the data is is decided by the GDPR: an anonymized technical log is not the same as an employee's medical record. What you're going to use it for is decided by the AI Act, and here the technology doesn't matter, the purpose does: the same model can be harmless drafting an email and outright prohibited if you use it to score people. Which tool you use is the question almost nobody asks, and the one that causes the most trouble in practice: a free personal account has no contract with your company, so nobody is accountable for what happens to the data once it leaves your screen.
The three questions are independent. Getting two right and the third wrong is as serious as getting all three wrong: harmless data, for a permitted use, uploaded to someone's personal account with no contract, is exactly the most common case of AI-driven data leakage, and the most avoidable one.
- 20%
- of organizations suffered a breach caused by unauthorized AI (IBM Cost of a Data Breach 2025)
- $670,000
- average added cost when shadow AI is behind the breach (Against a global average of $4.44M)
The full breakdown of these figures is in Deploying AI in the enterprise: security and governance.
The classifier: cross-check your three answers
What follows is not a list of rules to memorize: it is a tool that cross-checks the three questions and gives you a verdict with the reasoning in plain view, so you can verify it instead of trusting a black box. Pick the most sensitive piece of data in what you're sending, the actual use you'll make of it and the tool you would use. The result ranges from "prohibited" to "go ahead", with the nuances that really matter in between.
It is not a substitute for legal analysis of your specific case. It does not collect, transmit or store anything you select: all the processing happens in your browser.
An interactive classification tool is available in the web version of this article. It cross-checks three questions: what the data is (from "no personal data" to special categories, criminal records data or children's data), what it will be used for (from drafting and analysis to screening candidates, assessing creditworthiness or outright prohibited uses such as social scoring) and which tool is used (free personal account, business plan with a contract, API with a data processing agreement, or a model deployed on your own infrastructure). The verdict ranges from "prohibited" to "go ahead", via "not here" (the tool has no contract), "with conditions" and "high risk", with the GDPR and AI Act reasoning shown for each case.
The five possible verdicts
- Art. 5 Prohibited
- The use has been banned across the EU since February 2025. No contract, consent or configuration can fix it: the project has to stop.
- Art. 28 GDPR Not here
- The data and the use are fine, but the tool is not. Without a data processing agreement, nobody is accountable for what happens to what you upload.
- Arts. 6, 9 and 35 GDPR With conditions
- It can be done, once the paperwork is in place: a legal basis, a signed contract, training turned off and, if the data is sensitive, a prior impact assessment.
- Annex III High risk
- The use is on the AI Act's list: recruitment, credit, biometrics, education. Risk management and human oversight are mandatory.
- Go ahead
- With no personal data and a minimal-risk use, the way is clear. You still need to train the staff who use the tool.
The five outcomes the classifier can give you, depending on how your three answers combine.
The three uses that are not up for debate
Before you even look at the data or the tool, the AI Act rules out three uses outright. It doesn't matter whether you have consent, a contract or the best vendor on the market: they have been prohibited throughout the European Union since February 2, 2025, with fines of up to €35 million or 7% of the group's worldwide turnover.
No exemption for size or sector. Scoring customers or employees on their behavior and treating them differently as a result (Art. 5(1)(c)), detecting emotions in the workplace or in the classroom outside a medical or safety context (Art. 5(1)(f)), and mass-scraping faces from the internet or from cameras to build a facial recognition database (Art. 5(1)(e)). All three appear among the classifier's "use" options precisely so they stand out before you go any further.
The tool matters more than the data
This is the dimension almost nobody reviews, and it accounts for half of the classifier's "not here" verdicts. The difference is not how advanced the model is: it is whether there is someone on the other side you can sign something with.
Personal or free account The one each person opens with their own email
- No contract with your company: there is no processor accountable for the data.
- On the free plan, content can usually be used to train the model. Once it's in, it doesn't come out.
- There is no way to audit retention, server location or who it is shared with.
Business plan, API or in-house model With a data processing agreement
- A signed contract (Art. 28 GDPR) listing the sub-processors, with training turned off by default.
- Configurable retention, ideally zero, verifiable by contract and not just in the interface.
- Traceability of where the data is processed: if it leaves the EU, a transfer mechanism is required (Arts. 44 et seq.).
A personal account is not "the same AI on the free tier": it is a different contractual relationship, with different rules.
The timeline already under way
The AI Act did not come into force all at once, and the AI Omnibus shifted part of the timeline in July 2026. This is what already applies and what is still to come:
AI Act timeline: entry into force on August 1, 2024; Article 5 prohibitions and staff AI literacy (Article 4) from February 2, 2025; obligations for general-purpose AI models from August 2, 2025; Article 50 transparency obligations (disclosing that people are interacting with an AI, labeling synthetic content) from August 2, 2026; Annex III high-risk systems postponed to December 2, 2027 by the AI Omnibus; and high-risk AI embedded in already regulated products (Annex I) from August 2, 2028.
What to do based on your result
- Stop — Prohibited
No configuration will fix it. If it is already running, document when it started, and find another way to solve the business problem. - Switch tools — Not here
Move to a business plan, to the API under contract, or to an in-house model. Banning it without offering an alternative is the surest way to make sure it carries on exactly as before, only without anyone knowing. - Get the paperwork in place — With conditions
A signed data processing agreement, an identified legal basis and, with sensitive data, an impact assessment before you start, not after. - Apply the full regime — High risk
Decide whether you are the provider or the deployer, set up risk management and define who exercises genuine human oversight. - Train and document — Go ahead
The Article 4 AI literacy obligation applies regardless of the risk level. It is the cheapest one to comply with and the one most often skipped.
The verdict changes the action, not the need to put it in writing.
None of these steps requires shutting the company down for two months. They require knowing what is being uploaded, from which account and for what purpose, which is exactly what the classifier above answers. If the result has left you with more questions than answers, or if the verdict was "high risk" and you don't know where to start with the file, that is what AI Act compliance consulting is for: we review your real use cases, not a generic checklist, and tell you what needs doing and in what order. If what you're missing is day-to-day technical control (which tools each team can use and with what data), that is covered by AI governance and secure use, which is also where we go deeper into the gateway, the inventory and the metrics that keep all of this sustainable from the second month onward.
Related service
Topics
#Artificial-Intelligence #AI-Act #GDPR #Shadow-AI #Data-Protection #Compliance
Keep reading
- Deploying AI in the enterprise without opening a hole: control, governance and measurement
- European compliance calendar 2026-2030: CRA, AI Act, Data Act and everything still to come
- Anatomy of 0data: dissecting a data and AI platform that never leaves your house
- Docker Model Runner vs. Ollama
- Can emojis hack AI models?