Magic TeamsWhat Do GDPR and EU AI Act Rules Say About AI Email Disclosure?
GDPR never asks you to say an email was written by AI. The EU AI Act asks only in narrow cases, and a reviewed one-to-one business email is usually not one of them. GDPR governs what you do with people’s data. Article 50 of the AI Act governs whether the person on the other end can tell they’re dealing with a machine.
Two duties, two triggers, two fine ceilings. Founders keep fusing them into one imaginary rule that says “label every AI email in Europe.”
It stopped being theoretical on 2 August 2026, when Article 50 started to apply. Enforcement sits with national market surveillance authorities, the AI Office for systems under its supervision, and the European Data Protection Supervisor (European Commission).
Here’s the split in one view, then what each law actually says.
| GDPR | EU AI Act Article 50 | |
|---|---|---|
| What it governs | How you handle personal data | Whether people can tell they’re dealing with AI |
| Trigger | You process personal data of people in the EU | Your AI interacts with people, or publishes public-interest text or deepfakes |
| Requires an “AI wrote this” label on email? | Never | Only for autonomous AI conversations |
| Max fine | 20m euros or 4% of global turnover | 15m euros or 3% of global turnover |
Fine ceilings come from GDPR Article 83(5) and AI Act Article 99(4), both “whichever is higher.”
Does GDPR require you to disclose that AI wrote an email?
No. There is no provision anywhere in the regulation about authorship, drafting tools, or software credit.
Article 13 lists exactly what you must tell someone whose data you collect: who you are, the purposes, the legal basis, the legitimate interests you rely on, the recipients, any transfer to a third country, retention, and their rights (GDPR Art. 13). Nothing in that list is “the tool you typed with.”
So “GDPR requires an AI disclaimer” is folklore. What’s true is that GDPR gets loud the moment an AI system reads, enriches, or acts on an email containing personal data. It just asks different questions.
The four GDPR questions that do apply to an email operation
Do you have a legal basis? For B2B outreach most senders lean on legitimate interests under Article 6(1)(f), which means a documented balancing test, not a vibe.
Marketing consent is a separate question under the ePrivacy Directive, where Article 13 sets an opt-in default with a narrow soft opt-in for existing customers. The Court of Justice looked at that exception in Inteligo Media (C-654/23, 13 November 2025) and held that a free account can count as a “sale” for these purposes, because the payment can be indirect. It also stressed that the exception must be read strictly, and that a commercially motivated newsletter is still direct marketing (Matheson).
Did you say where the data came from? If you didn’t collect it from the person directly, Article 14(2)(f) requires you to disclose the source, including whether it came from publicly accessible sources, and Article 14(3)(b) sets the deadline at the time of the first communication (GDPR Art. 14). Scraped or vendor-supplied prospect lists land squarely here.
Is anything being decided automatically? Article 22 gives people the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them (GDPR Art. 22).
AI lead scoring that only ranks a queue rarely clears that bar. AI that auto-declines an application or auto-cancels an account might. Where Article 22 bites, Articles 13(2)(f) and 14(2)(g) also require meaningful information about the logic involved.
Where does the data physically go? If your model runs in a US cloud, that’s a third-country transfer needing its own disclosure and transfer mechanism.
The European Data Protection Board’s Opinion 28/2024 sets out the three-step legitimate interest test supervisory authorities should apply to AI models, and puts real weight on whether people could reasonably expect their data to be used this way (EDPB).
This is the checklist we work through before an inbox layer touches a single European contact.
- Documented legal basis for the outreach, with a written balancing test if you rely on legitimate interests
- Privacy notice that names the source of the data when you did not collect it directly
- First-contact link to that notice, not a buried footer
- A written line on whether any decision is solely automated under Article 22
- Third-country transfer disclosed, with a valid transfer mechanism for your model vendor
- Processor agreement with every AI vendor that sees message content
- Retention limit on stored prompts, drafts, and email context
- A working, honored opt-out on every send
If you want the wider version of this, we cover it in AI data privacy for agencies.
What does the EU AI Act’s Article 50 actually require?
Article 50 sets four transparency duties, and they bind different parties. Working out which one lands on you is most of the job.
| Article 50 duty | Who it binds | Does it touch your email? |
|---|---|---|
| 50(1) Tell people they’re interacting with an AI system | Provider, designed into the system | Yes, if an AI agent converses with people on its own |
| 50(2) Mark synthetic output in machine-readable form | Provider of the generative AI system | Usually your model vendor, not you |
| 50(3) Disclose emotion recognition and biometric categorisation | Deployer | Only if you run sentiment tools that qualify |
| 50(4) Label deepfakes and AI text published on matters of public interest | Deployer | Almost never for one-to-one business email |
Read 50(4) closely, because it’s the one people misquote. The obligation covers text that is published, to inform the public, on matters of public interest.
The Commission’s guidance ties that last phrase to politics and democratic processes, public administration, justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, scientific or cultural developments (European Commission FAQ).
A follow-up to a prospect about a proposal is not published, not aimed at the public, and not a matter of public interest. It fails all three legs.
The editorial-review carve-out is stricter than it sounds
There’s also a carve-out inside 50(4). Published text does not need a label where it “has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication” (EU AI Act, Article 50).
The Commission is blunt about what that means. It takes real authority to approve, alter or reject the substance, including fact-checking, and “superficial, solely formal, or procedural checks (e.g. spell-checking or grammatical correction) are not considered to be human review or editorial control” (European Commission).
That bar is higher than most people assume. Skimming a draft and hitting send is not editorial control.
Are you the provider or the deployer of your AI email agent?
This is the question that decides whether Article 50(1) is your problem or your vendor’s, and it’s where most summaries stop too early.
The usual line is that 50(1) binds providers, so a business using AI is safely a deployer. That holds only when someone else is the provider of the system you’re actually running.
Article 3(3) defines a provider as anyone who develops an AI system, or has one developed, and “places it on the market or puts the AI system into service under its own name or trademark.” Article 3(11) defines putting into service to include supply “for own use in the Union for its intended purpose” (EU AI Act, Article 3).
Read together, those two definitions split the room.
Buy a finished chatbot product and switch it on, and you’re a deployer. The vendor designed the notice; your job is to check it works and keep it in the contract.
Build your own email agent on a model API, name it, and point it at your inbox, and you have developed an AI system and put it into service under your own name. On that reading you’re the provider of that system, and 50(1) becomes a design duty you own. The model company is the general-purpose AI provider underneath, which does not cover the interface you built on top.
Role classification is fact-specific, and the Commission’s guidelines are formally non-binding even though they shape how national authorities enforce (Stephenson Harwood). If your setup is near this line, get it classified by a lawyer rather than assumed by a blog.
When does an AI email actually need a label?
The real trigger is Article 50(1), and the hinge is autonomy.
The Commission sets four cumulative conditions: the thing must be an AI system, designed for a genuine two-way exchange, communicating directly rather than through a human intermediary, and with natural persons. Systems “operating solely in the background, through machine-to-machine communication, or without direct contact with people, fall outside the scope of this obligation” (European Commission FAQ).
Law firm guidance reads this as covering chatbots, voice assistants, social bots, and “agentic AI systems that autonomously contact individuals (e.g., an AI agent making phone calls or sending emails on behalf of a business)” (Travers Smith).
So the line runs between drafting and conversing. An AI that writes a draft a human reads, edits, and sends is a writing tool. An AI that holds a thread with a stranger for four rounds while everyone’s asleep is the thing 50(1) was written for.
Where a disclosure is owed, Article 50(5) requires it “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure” (EU AI Act, Article 50). A footer on message five doesn’t cure message one.
The Commission’s final guidelines, adopted 20 July 2026, get specific about form. A plain-language notice at the start of a session works. A generic reference to an “assistant,” a machine-readable marker alone, or a line buried in terms does not, and an agent acting autonomously also has to say on whose behalf it acts (Stephenson Harwood).
That last requirement is the one email people miss. “You’re talking to an AI assistant” isn’t enough if the reader still can’t tell whose AI it is. For placement mechanics in an actual message, see where to put an AI disclosure in an email.
This is why the Magic Teams install defaults to draft-and-approve rather than autosend for anything that reaches a client, a prospect, or a regulator. The reason is quality first. A founder catching a wrong number before it ships is worth more than any label. The simpler legal position is a side effect we’ll happily take.
Do these rules apply to a US agency?
Often, yes. Article 2(1)(c) extends the AI Act to “providers and deployers of AI systems that have their place of establishment or are located in a third country, where the output produced by the AI system is used in the Union” (EU AI Act, Article 2).
An agency in Denver running an autonomous email agent that converses with a prospect in Lisbon is producing output used in the Union. Being American doesn’t exempt you.
GDPR reaches even further through its own territorial scope, which catches you when you offer goods or services to people in the EU or monitor their behavior (GDPR Art. 3).
How do these rules land on one agency’s inbox?
Say you run a nine-person marketing agency in Austin. Two of your clients are European, you bought a prospect list from a data vendor, and you’ve just turned on an AI layer over your shared inbox.
Three things are happening at once, and only one of them needs an AI label.
The AI drafts your outbound sequences and a strategist approves each one before it sends. No Article 50 label. It’s a writing tool with a human in the chair.
The AI auto-replies to inbound support questions from your European client’s customers, holding threads without a person in the loop. That’s Article 50(1). You assembled that agent on a model API and put your agency’s name on it, so the design duty is likely yours rather than your vendor’s, and the honest version is one plain line at the top naming both the AI and the business it speaks for.
The purchased prospect list is a GDPR problem with no AI in it. Article 14 says your first email must tell those people where you got their details and point to a real privacy notice. Whether a human or a model wrote that sentence is irrelevant to the duty.
One install, three separate legal questions. Most compliance confusion comes from answering them as though they were one.
What are the deadlines, and what did the Digital Omnibus change?
Less than people hoped. The Digital Omnibus on AI became Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force from 27 July 2026.
It pushed standalone Annex III high-risk obligations to 2 December 2027 and embedded Annex I systems to 2 August 2028, and it left Article 50 alone (Hunton).
One narrow grace period survives. Generative systems already on the market before 2 August 2026 have until 2 December 2026 to meet the marking and detection duty in 50(2) (Cooley). Systems placed on the market after that date get no runway.
Penalties for Article 50 breaches run up to 15 million euros or 3% of total worldwide annual turnover, whichever is higher. There’s a break most agency owners haven’t heard about: under Article 99(6), for SMEs and startups the fine is capped at whichever of the amount or the percentage is lower (EU AI Act, Article 99).
For a 20-person shop, that means the 3% figure, not the 15 million one. Still not a number you want, and small enough that the right response is a written rule rather than panic.
If you want a documented path, the Commission’s Code of Practice on Transparency of AI-generated Content is voluntary and had roughly 190 signatories by the end of July 2026 (European Commission).
The Commission and the AI Board judged it adequate in early July 2026, which buys signatories more predictable supervision and possible penalty mitigation, without being conclusive proof of compliance (Reed Smith).
Frequently asked questions
Does a privacy policy line cover my AI disclosure duty?
For GDPR transparency, a linked notice does a lot of the work. For Article 50(1), no. The disclosure has to reach the person at or before first interaction in a clear and distinguishable manner, and the Commission specifically rules out burying it in terms.
Does using ChatGPT or Claude to write emails make me an AI Act provider?
Not on its own. Typing prompts into someone else’s product makes you a deployer, and 50(2) marking sits with the company that placed that generative system on the market. Building and branding your own agent on their API is the fact pattern that can flip you into provider territory.
Does an AI-drafted newsletter count as text published to inform the public?
Usually not. A product update is commercial communication, not information for the public on politics, public health, justice, or the other listed categories.
Note that the Commission’s list does include economic, financial, scientific and cultural developments, so AI-written market commentary on your blog sits closer to that line than any of your email does.
Do I need consent to send AI-written cold email in Europe?
The AI has nothing to do with it. That question is ePrivacy plus GDPR, and it turns on whether the recipient opted in or fits the soft opt-in for existing customers. An AI-written email and a human-written email face the identical test, which we break down in do AI email disclosure laws apply to my business.
What if a client asks whether AI wrote their emails?
Tell them. No EU rule forces that answer in a normal service thread, and every commercial instinct says give it anyway. A founder who won’t say how the work gets done has a trust problem no statute will fix, and if you want wording, we have templates for the actual sentence.
One caveat before you act on any of this
This is a plain-language summary of published regulation and official guidance, not legal advice. Article 50 only started applying in August 2026, there’s no case law on it yet, the Commission’s guidelines are non-binding, and how a specific inbox setup lands depends on facts a blog post can’t see.
Run your actual configuration past a qualified EU data protection lawyer before you rely on any reading here.
I’m Satya Phanindra Reddy, and at Magic Teams AI we install the email layer with the approval step built in, the privacy notice wired to first contact, and the disclosure rule written down rather than improvised per campaign. If you’re trying to work out which of your three AI email problems is actually a legal one, that’s a good conversation to have before the next sequence goes out.