Brand Safety Institute Blog

OpenAI’s Business Agents: What Brands Need to Consider

Written by AJ Brown | Nov 18, 2026, 4:07:34 PM

The brand supplies the information and OpenAI's model decides what to say. If the answer causes a problem, the brand may have to investigate it without access to the conversation.

This piece steps outside the five parts we planned for this article series to respond to a timely announcement, and we expect to return to that structure with our next piece.

On November 10, 2026, OpenAI launched Businesses in ChatGPT, starting with eligible ecommerce and online-service brands in the United States. Brands can claim their profile, keep their information current and create a Business Agent that customers can talk to inside ChatGPT. They can also sponsor the agent and use it as the destination for a ChatGPT ad, so someone comparing products can move into a conversation with the brand without opening another site or starting the search again. OpenAI requires the agent to be clear that it is AI representing the brand, and the branded conversation stays separate from ChatGPT's independent answers.

The appeal for a brand is obvious. It can join the conversation when someone is already considering a product or service, answer the follow-up questions that usually get asked somewhere else, and reduce the number of steps between interest and action.

Parts 1 to 3 of our Brand Safety in the Conversation series focused on advertising within an AI conversation. In the case of standard ChatGPT ads, the brand still controls the content of its ad, and the uncertainty arises from what’s happening beside it and, as we discussed in Part 2, what ad adjacency means in a chat context. Then, as we discussed in Part 3, that uncertainty could be exacerbated by the user’s state, the role the assistant is playing, and whether the conversation drifts. But until this point, the brand's message had been written and approved before any of those considerations entered the picture. Business Agents change that because the model is now producing the brand's side of the conversation too.

With Business Agents, the brand provides information and guidelines, OpenAI's model writes the response, and the customer receives that response from an agent carrying the brand's name. Under the Businesses in ChatGPT Terms, the brand is responsible for the output, including answers it did not provide, review or approve, and under OpenAI's privacy approach, the brand does not receive individual chats or transcripts. The product is in beta and some of this may change, but the combination of limited visibility and full accountability raises new considerations for brands: what they can see, what they’re responsible for, whether that trade-off is worth it, what happens when a user takes the conversation off-script, and how to keep two versions of the same agent aligned.

Limited Brand Visibility

Brands currently create a Business Agent through OpenAI's Agent Builder and cannot bring their own agent or connect their own plugin. They provide business information, links and guidelines informing the agent's tone, when it should ask a question and what it should do when the information does not support an answer. They can then preview the agent, try sample questions and update the information or instructions when an answer is wrong.

OpenAI's own guidance uses a brand selling products in different finishes and colours as the example. It suggests asking whether a customized order can be returned, checking the answer against the returns policy and updating the business information if the agent gets it wrong. It also recommends asking about an unannounced product release, to make sure the agent says it has no confirmed date instead of inventing one.

All of that control happens before launch, and testing only shows how the agent handles the questions a brand thinks to ask. Once the agent is live and answering real customers, the brand’s view narrows considerably. OpenAI’s launch announcement says it does not share users' conversations or Business Agent chats with brands, and the guidance for sponsored agents says brands receive aggregated reporting on patterns across conversations, where available, rather than individual chats or transcripts. More detailed exploration becomes available once the agent has handled more than 25 conversations, but the brand can’t access the contents of any one conversation.

Aggregated reporting should help a brand spot recurring issues. If people keep asking about returns and the agent can’t answer, the pattern can show up in these insights, and the brand can improve the information it provides to OpenAI.

Aggregated reporting is much less useful when the problem is one answer given to one customer. The customized-order example is harmless while it stays in testing, but once the agent is live, a customer might rely on its answer, place the order and later be told the brand’s policy says something different. The first evidence the brand sees may be that customer’s screenshot. It might show the sentence that caused the problem, but not the earlier questions, the link the agent relied on, the version of the policy in place at the time, or what happened after the disputed answer.

The Brand is Accountable

Limited visibility into individual conversations doesn’t limit the brand’s accountability for them. Section 5.2 of the Businesses in ChatGPT Terms says that, as between OpenAI and the brand, the brand is responsible for the Business Agent, including its content, configuration, safety testing, actions and outputs, whether or not it provided, reviewed or approved them. The brand must correct identified problems within its control and notify OpenAI when it becomes aware of an inaccuracy or unauthorised behaviour.

Section 5.6 warns that Business Agents may produce inaccurate or unexpected responses, may fail to follow their instructions and are not guaranteed to be accurate, lawful or suitable for the brand. Section 9 requires the brand to indemnify OpenAI against certain third-party claims involving the agent, including allegations of false or misleading advertising and consumer-protection failures.

Those terms set the contractual split between OpenAI and the brand. They don’t decide what a regulator or court would conclude about a particular incident, and we’d be careful about making that leap without knowing the facts and the law that applied (we’re also not lawyers, for the record), but OpenAI’s position is clear: the brand chooses to publish the agent and carries responsibility for what it says, even when the model doesn’t follow its instructions.

None of this is entirely new. Brands already manage similar risks when they train staff or run conventional chatbots on their O&O properties. An employee can misunderstand a policy, an old script can stay in circulation, and a chatbot can pull an answer from a wrong or outdated resource page. The brand still has to correct the customer and deal with the fallout, so the responsibility itself isn’t what’s different.

There’s even an earlier example for a brand being held responsible for an automated answer, though the comparison only goes so far. In Moffatt v. Air Canada, a British Columbia tribunal ruled in February 2024 that the airline was liable for wrong bereavement-fare information given by the chatbot on its own website, and rejected the idea that the chatbot was a separate entity responsible for its own actions. The plaintiff had sent Air Canada a screenshot of what the chatbot told him. It was a small claims tribunal in Canada and the chatbot sat on the brand's own site, so we would not assume this precedent automatically carries over to a Business Agent built on OpenAI's model, but it does show where the burden is likely to fall.

The key difference with Business Agents is that the brand is responsible for what its agent said without being able to see what its agent said. A chatbot on a brand’s own site leaves a transcript the brand can check, but a Business Agent doesn’t give the brand one. Anyone who has investigated a brand safety incident knows how quickly missing context becomes the argument. Brands unable to supply a platform with a screenshot, URL, or other context behind an alleged incident are often disappointed by the resulting inaction.

With Business Agents, the brand is expected to identify the problem, correct it and tell OpenAI, while OpenAI treats the output as the brand's responsibility and keeps the conversation private from it. Tweaking the language in a brand’s profile might prevent a wrong answer from happening again, but it will not show whether the cause was outdated information, an ambiguous question or the agent ignoring its guidelines. That difference feels academic in testing, but much less so when a customer, a regulator or a journalist is asking.

Trust Paradox Revisited: Privacy as a Potential Adoption Deterrent

In Part 1, we defended OpenAI’s decision not to share conversation transcripts with advertisers whose ads appear in them. Users who trust that their conversations stay private are more willing to use ChatGPT openly, and that trust is what makes the platform valuable to users and advertisers alike. The trust paradox carries over to Business Agent chats too, but the stakes for the brand, and its available alternatives, are different.

With standard ChatGPT ads, a brand that wants to reach people in ChatGPT has no real alternative to accepting OpenAI’s limits on context. With Business Agents, it might. A ChatGPT ad can just as easily send someone to the brand’s own website, where a chatbot the brand runs can answer the same follow-up questions. On its own properties, the brand can retain and access full chat transcripts, and see which policy and sources its agent drew on when a customer disputes an answer. Using OpenAI’s Business Agent instead means giving up that visibility while still taking responsibility for what the agent says on your behalf.

What OpenAI offers in return is convenience: the customer never leaves ChatGPT, and there’s one less step between interest and action. For some brands, including those that don’t have their own chatbots, that will be worth it. For others, especially those whose legal or customer-service teams have to stand behind every answer, an ad that directs to a chatbot the brand fully controls may be the safer route to the same customer. So it’s an open question how widely brands will adopt branded chats inside ChatGPT, given the privacy stance OpenAI has taken.

To be clear, we don’t think the answer is handing brands full transcripts, and we don’t think OpenAI’s privacy stance is necessarily wrong. People may mention their health, finances or other sensitive personal details in these conversations, and sharing those details with brands could change how people use the platform, to OpenAI’s detriment. But there’s plenty of room between sharing full transcripts and aggregate themes. Examples include a record of which business information and guidelines were referenced when an answer was given, an incident record preserving the sources and instructions behind a disputed answer, or an option for a complaining customer to authorize release of a particular exchange.

Controls like these may already exist elsewhere in the product or arrive as the beta develops. We could not find them in the public material, so a brand deciding whether to publish an agent today should ask rather than assume. Aggregated insights help improve the agent over time, but they may not be sufficient to convince a brand with its own chatbot to choose OpenAI’s instead.

Questions a Business Agent Can’t Answer

OpenAI’s Business Agents Policy does a lot to keep brands, and the agents speaking for them, from steering a conversation somewhere it shouldn’t go. Brands can’t instruct an agent to ignore or downplay an immediate safety concern, such as a possible medical emergency or danger to someone. Agents can’t exploit vulnerability, create false urgency or keep pushing a sale after someone declines. Brands also can’t have their agent ask for unnecessary health information, financial details, identity documents or other high-risk personal data, nor can they have it share a user’s information with the brand or submit a form, booking or purchase without the user’s explicit authorisation through an approved feature.

Those are good rules, but they only govern the brand’s side of the conversation; they can’t control what the user does. A customer can start with an ordinary commercial question and take the conversation somewhere the brand never designed its agent to go. They might volunteer sensitive information the brand didn’t ask for, or ask questions the brand has no business answering. Someone asking a travel brand about changing a booking might explain that the change is connected to a personal safety issue. Someone else might tell a retailer’s agent they’re in crisis. In those moments, the right response could mean dropping the sales conversation and pointing the person to other sources of help, whether that’s ChatGPT itself or support outside it. To ensure that a brand’s agent isn’t left speaking in a conversation it was never built for, the agent has to recognise that the purpose of the exchange has changed and that it’s time to (mindfully) end it.

The public material is vague about what happens when the conversation reaches that kind of boundary, how a hand-off from Business Agent chat to ChatGPT would work, or what the brand can later see to confirm the situation was handled properly. OpenAI may manage the safety response entirely inside ChatGPT, and that could be the right privacy decision, but the brand is still represented in the conversation and, under the terms, still responsible for the output until that hand-off.

This matters even if OpenAI keeps ads away from ChatGPT conversations that are sensitive in their own right, because a conversation can become sensitive after the customer enters the Business Agent, and by then the user is in a conversation carrying the brand’s name, whether they arrived through an ad or through the brand’s profile. A brand needs to test when its agent should carry on, when it should change what it is trying to do, and when it should stop speaking for the brand and hand the user back to ChatGPT.

Disconnect Between Sponsored and Organic Agents

OpenAI says that an organic agent created through Business Hub and a sponsored agent created through Ads Manager do not currently synchronise. A brand using both has to reuse its information and guidelines, then test and publish the sponsored version separately, so the same brand could answer the same question differently depending on whether the customer arrived through its profile or its ad.

That is manageable if a single person or team owns both agents and keeps a record of every change, but it gets harder when advertising, ecommerce, customer service and legal each assume someone else is doing it. A corrected returns policy in Business Hub does not help much if the sponsored agent is still running on the earlier version in Ads Manager. We suspect this disconnect will be bridged at some point, but it’s a gap at the moment and worth noting.

Even if OpenAI closes that gap, the same problem will arise for brands seeking to maintain consistency among their agents across platforms if Google and Meta roll out similar tools. Brands will need one place to maintain their policies, prices and offer terms, and a way to check that each agent is using the current version.

Questions to Ask Before You Publish

Testing the questions customers usually ask will confirm the basic information is there. The more useful tests involve a customer who starts from a false premise, asks for something the brand cannot promise, challenges an earlier answer, gets angry, or takes the conversation somewhere the agent was not designed to go. It is also worth testing what happens when a customer says the agent gave them a different answer yesterday, since that is likely to show the gap between how the system is tested and how the brand will be asked to defend it.

Before putting an agent in front of customers, we would want answers to the following questions:

  • Which version produced the disputed answer? With an organic agent and a sponsored agent that do not synchronize, the answer has to cover both.
  • What information and guidelines were referenced when the answer was given? You will want a record of what the agent was working from at the time, rather than only what it uses now.
  • Did the agent follow the approved guidelines? Without the conversation, you will need to know how that gets checked.
  • What evidence would remain if you had to investigate a complaint, and who holds it? OpenAI says brands receive patterns rather than transcripts, so ask what evidence is preserved that could help triage a specific incident.
  • How quickly can you suspend the agent while you work out what happened? Settle who owns that call before launch, since advertising, ecommerce, customer service and legal could each assume it is someone else.

With our next piece, we expect to pick the series back up with Part 4, on the policies and regulation shaping ads in ChatGPT, and we’ll come back to Business Agents where they fit. As with the rest of this series, we don’t think we’ve got all of this right. If you’ve tested a Business Agent or see any of this differently, we’d love to hear from you.

 

* Image generated by ChatGPT. The irony is not lost on us.