Since August 2 your agent has to say it is an AI: who answers when you build it and your client runs it

Paragraph 1 is the one that reaches anyone building a conversational agent. It is addressed to providers of the system, and this is the official English text:

Paragraph 2 requires synthetic audio, image, video, or text output to be marked in a machine-readable format so it can be detected as artificially generated. Audio is expressly inside, so a voice agent is caught twice over: once by the spoken disclosure in paragraph 1, once by the technical marking in paragraph 2. Paragraph 3 puts the duty to inform on the deployer when emotion recognition or biometric categorization is in play. Paragraph 4 makes the deployer disclose deepfakes and AI-generated text published on matters of public interest, with one important carve-out: it does not apply where the content went through human review or editorial control and somebody holds editorial responsibility for the publication. Paragraph 5 fixes the timing and the form: the information has to be given in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure. And paragraph 6 shuts the usual back door: complying with Article 50 does not release you from Chapter III or from other transparency obligations under Union or national law. Where the GDPR applies to you or your end users, it stays exactly where it was.

On consequences. Article 99 places breaches of the Article 50 transparency obligations in the band of fines up to €15,000,000 or up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. Now the nuance almost nobody writes down, and the one that changes the conversation with a small-business client: for small and medium enterprises, including startups, Article 99 applies the lower of the two figures, not the higher. And the fines are imposed by national authorities under whatever regime each Member State sets, not by the Commission.

Two documents help you read all of this, and they are not the same thing. The European Commission guidelines on transparency of AI-generated content, adopted in July 2026, and the Code of Practice from June 2026. The Code is voluntary, as Spain's AI supervisory agency confirms. But not signing it does not get you off the hook: under the guidelines, whoever does not adhere has to demonstrate compliance with the marking and labeling duties by alternative adequate means. Either you sign the Code or you carry the burden of proving on your own that what you do works just as well.

Three-month and six-month figures for that transition period circulated in the press. The correct one is December 2, 2026.

The regulation reasons with two figures: provider and deployer. The split is clean on paper. The provider designs the system so people can tell it is AI and marks the outputs; the deployer informs users about emotion recognition and biometrics and labels deepfakes and public-interest text that nobody reviewed. Everything published about this reasons with that binary matrix and quietly assumes there is a single company behind it.

Your chain has three links. The platform that provides the system. You, who configure it, write the script, wire it into the channels, and hand it over under your client's brand or your own. And the end business, which runs it and faces its own users. None of the official sources we checked settles which figure lands where in a white-label build, and be suspicious of any blog that settles it for you in one paragraph: that is a legal characterization, not an implementation tip.

What you can do, and what actually saves you the trouble, is to stop arguing about the label and make sure the disclosure exists, works, and survives contact with the client. The three failure modes that keep repeating in a three-party build are these:

One case shows where the reasoning is heading, even though it does not bind anyone here. In February 2024, the Civil Resolution Tribunal of British Columbia, a Canadian small-claims body, ordered Air Canada to pay a passenger C$812.02 (C$650.88 in damages plus interest and fees) over wrong information given by the chatbot on its website. The tribunal summarized the airline's position as suggesting, in effect, that the chatbot was a separate entity responsible for its own actions, called that submission remarkable, and rejected it: the chatbot is part of the website, and the company is responsible for all the information on it, whether it comes from a static page or a chat window. The most uncomfortable detail is that the chatbot linked to the correct page and there was still an award, because the tribunal would not accept that a customer has to cross-check one part of a website against another. It is not US case law and it binds no court here, but it marks the direction of travel: whoever puts the agent in front of customers answers for what it says.

The practical consequence for an agency is dull and unglamorous. Whoever can show, with dates, what the opening block said on any given day is the party that wins the argument six months later. Everyone else is relying on memory.

• The platform gives you no fixed place to put the disclosureIf the only place the disclosure lives is the prompt, anyone who edits the prompt can delete it without noticing.
• The client removes it while adjusting the toneThe most common one by far. The business wants it to sound warmer, rewrites the opening, and the disclosure disappears. Nobody finds out until somebody asks.
• Nobody rechecks the disclosure when a new channel opensThe agent launched on WhatsApp with the disclosure in place, and three months later the client wires it into Instagram and a phone line by copying over the answer block alone.

This is what people are actually looking for when they search this obligation: the text. Below are six short versions, one per situation. They are written to fit inside a WhatsApp template without eating the message, and to be read aloud on a call without sounding like a phone company disclaimer.

The three moments in that last row are not arbitrary: Meta's own policy points to the start of the conversation, resuming after a significant lapse of time, and handing off from a human to automation as the points at which to inform. It is a good map of where to repeat the disclosure without turning it into a tic on every turn.

And what does not count, in order of frequency. Calling it a virtual assistant and stopping there does not count: virtual does not mean AI, and for fifteen years a virtual assistant was a person working remotely, often in another country. Putting it in the website footer does not count, because someone messaging on WhatsApp never went through the website. Burying it in the privacy policy does not count. And a sixty-word paragraph with regulation citations does not count either: the information has to be clear and distinguishable, and a wall of text nobody reads is neither. If your agent runs from a prompt, the disclosure belongs to the fixed identity block, not the tone block, so that a rewrite of the personality cannot take it out.

The exception is real and it sits inside paragraph 1 itself, but it arrives with a list of adjectives that most summaries trim off. It is not enough that it be obvious to anyone: the text speaks of a natural person who is reasonably well-informed, observant and circumspect, and it adds that you have to take into account the circumstances and the context of use. If you are going to cite it, cite all of it or leave it alone.

The Commission, in its FAQ on the Article 50 transparency obligations, holds that the exception must be read narrowly and that the yardstick is an average person who is reasonably well-informed, circumspect and observant. That is Commission interpretation, not enacted text, but it is the interpretation a national authority will be working from.

Translated into your daily work: naming the agent Assistant at [Business] does not make it obvious. Answering fast and well does not either; if anything, the better it works, the less obvious it is. A user who messages a clinic at 11:40 p.m. and gets an answer in two seconds can reasonably assume somebody is on call. And on voice, with decent synthesis, the obviousness is lower still.

The economic argument is stronger than the legal one. Complying costs you a twelve-word sentence at the top of the conversation. Leaning on the exception costs you having to defend, if anyone ever asks, why you considered it obvious on that channel, with that script, for that user. One sentence against one legal opinion. There is no debate.

Time to be honest about the scope of this check: territorial reach is set by Article 2 of the regulation and we have not read it line by line, so we are not going to declare who falls inside and who falls outside. What can be said without risk is that the question is not answered by looking at where your client's business is incorporated, and that assuming it does not apply because the company is in Austin or Miami while the agent serves users sitting in Ireland or Spain is, at best, an expensive bet.

There is a reflex you already have on the cost side that works as an exact analogy. On WhatsApp, the rate for a message depends on the country code of the number receiving it, not on the country of the company sending it. You have spent years looking at the recipient rather than the headquarters to work out what a message costs you. Transparency deserves the same reflex.

Then there is the practical argument, which wins nearly every time. If the same agent, with the same script, serves customers on two continents from the same number, you are not going to maintain two versions of the opening or fork the flow by country code to save twelve words. And in the other direction, US state-level transparency rules exist, they vary, and we have not verified them here, though it is telling that Meta's own policy singles out California. Putting the disclosure in every time is cheaper than auditing every market you touch.

An article that only says you must comply is a brochure. These are the cases where this does not reach you, or does not reach you yet, and the one where you would be doing more work than the situation deserves:

• If what you built is not an AI systemA button tree with hand-written answers is not the same thing as a model generating text. Whether something qualifies as an AI system for purposes of the regulation is not settled by this article and turns on the Article 3 definition, but it is a different case and deserves a look before you assume you are inside.
• The technical marking, if your system predates August 2If the system was already placed on the market before August 2, 2026, the machine-readable marking duty in paragraph 2 is not required of you until December 2, 2026. The paragraph 1 disclosure is required now.
• The high-risk block, which did not start this monthThe omnibus pushed Annex III systems to December 2, 2027, and Annex I systems to August 2, 2028. Saying the whole regulation applies as of August is the mistake being repeated in both directions.
• The Code of Practice, which is voluntaryNobody forces you to sign it. But not signing does not exempt you: you will have to demonstrate compliance with marking and labeling by alternative adequate means. Voluntary does not mean free.
• When the agent is only a doormanIf it collects a name and a phone number and hands off to a person in three messages, you do not need a sixty-word legal block. You need one sentence. Excess protects nobody here: it raises drop-off on the first message and the user reads it less carefully, not more.

What does not stop applying in any of those cases: paragraph 6 of Article 50 itself, which reminds you that meeting the transparency duty does not release you from Chapter III or from other transparency obligations under Union or national law. And where the GDPR applies to you or your end users, it runs on its own track regardless of any of this.