By Article50.io · Last updated 11 November 2026
In practice, yes, if people in the EU can chat with your ChatGPT Business Agent, because Article 50(1) requires that they are told it is AI. That duty falls on the provider: most likely OpenAI, not you. But you write the agent's instructions, and OpenAI's policy requires it to say it is AI. Business Agents are US-only for now, so EU exposure is a possible gap, not a proven breach.
Key answers
- Who is the provider? Most likely OpenAI, which builds and runs the agent in ChatGPT. The business is most likely a deployer, because it decides to publish the agent and how it is used (guidelines paragraphs 10 to 12).
- What should the agent say? That it is AI and which business it acts for, at the latest when each conversation starts (Article 50(5); guidelines paragraphs 31 and 143).
- Is it obvious inside ChatGPT? Don't count on it. The exception is read restrictively, and helpdesk-style chatbots are listed as not obvious (paragraph 45).
- Does it apply to US-only agents? Only if EU users can foreseeably reach them. OpenAI has not said whether they can (Article 2(1)(c); paragraph 10).
- What to do now: give the agent a first message that says it is AI and names your business, and no human name or photo.
This is general information, not legal advice. For the wider rules on chat widgets, see the chatbot disclosure guide.
What are ChatGPT Business Agents?
Business Agents are AI chats that answer customer questions for a named business inside ChatGPT. OpenAI launched Business Agents on 10 November 2026 as part of Businesses in ChatGPT, a beta. Through Business Hub, a business can claim its profile, update its details and product feeds, and see aggregate insights: mentions, website clicks and discovery topics. OpenAI says businesses do not receive individual prompts or Business Agent chats.
To claim a profile, a business confirms ownership with one of the methods offered to it, such as a business phone, a business email, a DNS TXT record on its website domain, or a Shopify store. To publish changes, connect a product feed, see insights or create an agent, the person managing it must then pass an identity check with Persona. Claiming does not guarantee more visibility in ChatGPT's answers, OpenAI's help page says.
The Business Agent is the part that matters for Article 50, because it chats with customers about a business's products and policies, in a separate conversation inside ChatGPT. OpenAI says that conversation is "clearly labeled, separate from the customer's original ChatGPT conversation". For now, Business Agents are only open to eligible ecommerce and online-service businesses in the United States with a website. The agent cannot take orders or bookings; it links to the business's website.
Does the EU AI Act apply to ChatGPT Business Agents?
Yes. Article 50(1) covers AI systems that chat directly with people, and a Business Agent is one; the open question is whether people in the EU can reach it. The guidelines say AI agents are covered, and list e-commerce and customer-support chatbots as examples (paragraph 31). A Business Agent fits that description.
Reach depends on Article 2. It brings in providers wherever they are based if they place a system on the EU market, and third-country providers and deployers where the system's output is used in the EU (Article 2(1)(a) to (c)). The guidelines add that "incidental, unforeseeable or unauthorised downstream use" should not alone trigger the obligations for systems not placed on the EU market (paragraph 10).
That gives three practical cases:
- A US business with a Business Agent. If people in the EU can open the agent, and that is foreseeable, the output is used in the EU. OpenAI has not said whether users outside the US can see US Business Agents. Until it does, this is a possible gap, not a proven breach. Our guide to whether Article 50 applies outside the EU explains the "output used in the EU" test.
- An EU business waiting for rollout. OpenAI says it plans to expand Business Agents "to more markets and business categories" but has given no date. EU businesses can already claim a profile in Business Hub; only the agent itself is US-only. An EU business that later publishes one is plainly within the Act's reach (Article 2(1)(b)), though as a deployer its own Article 50 duties are in 50(3) and 50(4). The questions below are worth settling first.
- An agency building agents for clients. OpenAI lets the business owner grant an agency access to manage its agent. A legal person stays the deployer even when contractors operate the system on its behalf and under its responsibility and control (paragraph 14), so the client usually keeps that role.
Article 50 has applied since 2 August 2026, and the interaction disclosure gets no transition period (guidelines paragraph 153). For Article 50, the 2 December 2026 date covers only machine-readable marking (Article 50(2)) by generative systems already on the market before 2 August 2026 (Article 111(4)).
Who is the provider of a ChatGPT Business Agent: OpenAI or the business?
Most likely OpenAI is the provider and the business is the deployer, though the guidelines don't address this platform directly. The answer matters because Article 50(1) is written for providers. A provider develops an AI system, or has it developed, and places it on the market or puts it into service under its own name or trademark (Article 3(3); guidelines paragraph 10). A deployer uses an AI system under its authority, which the guidelines read as deciding to deploy it and how it is used, with or without technical control (Article 3(4); paragraph 12).
On what OpenAI publishes, OpenAI looks like the provider. It builds the agent experience, runs it inside ChatGPT, controls the conversation label, and "may create and configure a Business Agent for your business" (Businesses in ChatGPT Terms, section 5.1). The guidelines' basic example of a provider is a company offering a chatbot or AI agent under its own name to businesses that use it without modifying it (paragraph 11).
The business looks like a deployer. It decides whether to publish, writes the instructions and supplies the information the agent uses. That is "authority" over the decision to deploy and the manner of use (paragraph 12). Deployers' own Article 50 duties are in 50(3) and 50(4) (paragraph 13), which a sales agent rarely triggers.
Two points are unclear, and we say so plainly:
- The agent speaks under the business's name. Paragraph 11 says a company that modifies an existing system, for example with new training data, and puts it into service under its own name becomes a provider of the new system. Writing instructions and loading a product feed is far lighter than new training data, but the guidelines do not address branded agents on a third-party platform directly. OpenAI's terms also let it use the business's names and trademarks in providing the service (Terms, section 2.3). One operator can hold both roles (paragraph 15).
- OpenAI's terms put responsibility on the business. Section 5.2 of the Terms makes the business responsible, as between it and OpenAI, for the agent's content, configuration and outputs. That allocates risk in the contract. It does not decide who is provider or deployer under the AI Act.
For enforcement, the guidelines name the EU AI Office as the authority for AI systems built on general-purpose AI models by the same provider (paragraph 151, footnote 47). OpenAI has not said which model powers Business Agents, so whether the AI Office or a national authority would supervise them is unconfirmed.
For the same question on ordinary website widgets, see do I need Article 50 compliance if I only use a chatbot?
Do I need to disclose an AI agent in ChatGPT?
In practice, yes: make sure your Business Agent says it is AI and names your business in its first message. The Article 50(1) duty sits with the provider, most likely OpenAI, but the business shapes what the agent says, and OpenAI's own policy already requires it: "An agent must be clear that it is AI representing the business; do not make it appear to be an independent adviser, a human, or endorsed by OpenAI" (Business Agents Policy, updated 10 November 2026).
What Article 50 expects of the notice:
- What it says. People must be told they are interacting with an AI system (Article 50(1)). For AI agents, the guidelines add that the agent should disclose both its artificial nature and the person on whose behalf it acts (paragraph 31).
- When. At the latest at the first interaction (Article 50(5)), and at least once at the start of each conversation (paragraph 143). One prominent notice before the first interaction is likely to be enough in most cases (paragraph 40).
- How clearly. It must be clear and distinguishable (Article 50(5)). A generic "assistant", a human-like name or photo, or a line in the terms is not enough on its own (paragraph 38).
- When asked. The agent should confirm it is AI whenever someone asks, or seems likely to be misled or confused about it (paragraph 40). A badge or label that stays visible throughout the chat, alongside the first-turn notice, is a recommended best practice where proportionate to the context and risks (paragraph 37).
OpenAI's conversation label may already do much of this. We could not see the label's wording, so we cannot say whether it names AI clearly enough. A first message from the agent is the part a business can influence, through its instructions. For the general rule, see does my chatbot need an AI disclosure?
Is it obvious you're talking to AI inside ChatGPT?
Not safely: the guidelines read the "obvious" exception restrictively, so don't rely on the ChatGPT setting alone. Article 50(1) drops the notice only where it is obvious to a "reasonably well-informed, observant and circumspect" person that they are dealing with AI. A provider relying on that must be able to assess and demonstrate it (paragraph 42).
Being inside ChatGPT helps, but it is not a safe answer. The guidelines say the exception "should be interpreted restrictively", that general awareness that chatbots exist does not mean people recognise them, and that it should be limited to cases with "almost no doubt" (paragraph 45). They list chatbots on online platforms and helpdesks, whose replies users "may perceive as human-generated", as examples where the exception does not apply.
A Business Agent is a conversation with a named business about its products and returns, which looks a lot like a sales or support chat. ChatGPT's audience is the general public, including people with lower AI literacy, which makes it harder to treat the AI nature as obvious (paragraphs 44 and 45). OpenAI describes no hand-over to human staff inside the agent chat, which lowers the risk compared with a website widget that mixes bots and people, but we would not rely on the exception.
What does the business control, and what does OpenAI control?
The business decides whether to publish the agent and writes its instructions; OpenAI runs the model, labels the conversation and decides who can open it. Based on OpenAI's announcement, Help Center article, Business Agents Policy and Terms, as checked on 12 November 2026:
| Part of the agent | The business | OpenAI | Unclear |
|---|---|---|---|
| Whether it goes live | Chooses to publish; it stays private until then | Can decline, pause or remove an agent | — |
| Instructions and information | Writes the instructions; supplies product and policy details | Sets what "functionality and settings" exist; may create a draft agent | — |
| Conversation label | — | Labels the Business Agent conversation | The label's exact wording |
| First message or greeting | Can instruct the agent | Runs the model that writes the reply | Whether a fixed custom greeting can be set |
| Model behaviour | Tests the agent in preview | Runs the model; does not promise it will always follow instructions (Terms 5.6) | Which model is used |
| Who can open it | — | Sets eligibility (US-only for now) and markets | Whether users outside the US can open US agents |
| Contractual responsibility | Responsible for content, configuration and outputs (Terms 5.2) | — | How that maps to AI Act roles |
What could a clear first message look like?
A clear first message says the agent is AI, says it is not a person and names the business, before it answers anything. This is an illustration only, not a template. Adapt it to your business and check it with your adviser:
Hi, I'm an AI assistant for [Business name], chatting with you in ChatGPT. I'm not a person. I can answer questions about our products and policies. To order or arrange a return, I'll link you to our website. Ask me any time if you want to know how I work.
Why it is written this way:
- It says AI and not a person, not just "assistant" (paragraph 38).
- It names the business it acts for, as the guidelines expect of agents (paragraph 31).
- It comes first, before any answer (Article 50(5); paragraph 143).
- It has no human name or photo, which OpenAI's policy and paragraph 38 both rule out.
Put the same line in the agent's instructions, and tell it to confirm it is AI whenever asked. Then preview the agent to see whether it actually says so. For more wording, see our AI chatbot disclaimer template.
What should EU businesses and agencies do now?
Start with the AI chats you already run on your own website, then settle your Business Agent's disclosure before it can reach EU users.
- Check your own website first. Your site's chat, voice or avatar widgets are under Article 50 today. The free scan loads the one page you enter in a real browser (it doesn't crawl your site), under a minute and capped at 45 seconds, and flags undisclosed widgets. It cannot see anything behind a login, so it cannot test a Business Agent inside ChatGPT. See product facts for what it checks and a sample report for what the paid report looks like.
- Write the disclosure into your agent's instructions before you publish, if you are a US business. Keep a dated copy and a screenshot of the preview.
- Watch OpenAI's pages for an EU rollout and for the wording of its conversation label.
- Agencies: keep a record of what each client approved, because the client is usually the deployer.
Frequently asked questions
Short answers to the questions people ask most about Business Agents and Article 50.
Does Article 50 apply to a ChatGPT Business Agent?
It can. A Business Agent is an AI system that chats directly with people, so Article 50(1) applies when it is placed on the EU market or its replies are used in the EU. Business Agents are limited to eligible US businesses for now, so EU exposure is a possible gap to watch, not a proven breach.
Who is the provider of a ChatGPT Business Agent?
Most likely OpenAI, because it develops the agent and runs it inside ChatGPT. The business is most likely a deployer: it decides to publish the agent and how it is used. The guidelines do not address this platform directly, and a business that modifies an agent and offers it under its own name could arguably become a provider too.
Does OpenAI's label count as the AI disclosure?
Possibly, but nobody can say yet. OpenAI says the Business Agent conversation is clearly labeled, but it has not published the label's wording. Article 50 needs people told, clearly and at the start, that they are dealing with AI. A first message that says so and names the business removes the doubt.
Is it obvious you are talking to AI inside ChatGPT?
Not safely. The guidelines read the obvious exception restrictively and say that knowing chatbots exist is not the same as recognising one. A Business Agent speaks for a named business, which can feel like a sales or support chat. Helpdesk-style chatbots are listed as examples where the exception does not apply.
Can Article50.io scan a ChatGPT Business Agent?
No. Article50.io loads a live public page in a real browser, under a minute and capped at 45 seconds, and it cannot see anything behind a login. It cannot open ChatGPT or test a Business Agent. It can check the chat, voice or avatar widgets on one page of your own website, which Article 50 also covers.
Sources
Sources: OpenAI's Businesses in ChatGPT announcement (10 Nov 2026), its Help Center setup article, the Business Agents Policy and the Businesses in ChatGPT Terms, checked 12 Nov 2026. EU AI Act Articles 2, 3, 50, 111 and 113, and the Commission's Article 50 guidelines, C(2026) 7702 final, adopted 5 Nov 2026, non-binding (paragraph 5).
This is general information, not legal advice.