On September 16, OpenAI unveiled a set of new capabilities for ChatGPT Ads. For people who work on ads or growth, the key is not just that “AI can write a few more ad copies,” but that the ad’s workflow chain begins to connect to consultation, leads, and subsequent follow-up.

Sponsored Agents are currently being tested with some advertisers in the United States: after a user clicks an ad, it can proactively start a brand-sponsored Agent conversation with clear labeling, ask whether the product is a good fit, and then decide whether to go to the merchant’s website. OpenAI explains that this conversation is separate from ChatGPT’s independent responses and also separate from the user’s original conversation.

The same batch of updates also includes: advertisers can use the ChatGPT Ads Manager plugin to create, update, and analyze ads using natural language. The tool will suggest ad copy and images based on the landing page and campaign objectives, but the advertiser must review and edit first before deciding whether to include the ads in the campaign. HubSpot users can connect accounts in context on their own CRM, create ads, view performance, and follow up on leads. U.S. Shopify merchants can use a new application to create and manage campaigns; it will be available starting September 23 for international markets where ChatGPT Ads has been opened.

These are the scope of product launch and testing, not promises of ad performance. Sponsored Agents are currently being tested with only some advertisers in the United States; the original text did not disclose pricing, conversion rates, or merchant ad delivery ROI.

After an ad is connected to the conversation and the CRM, what the customer truly needs to manage is a chain of tasks: what was said in the ad, whether the landing page can prove it, what questions the user will follow up with, which ones can be answered and which must be handed to a human, and who follows up on the leads, when. If you only provide a title, body, and images, things are easiest to get out of control in the second half: product information expires, answers go beyond what was promised, and key questions don’t make it into the follow-up queue.

You can break an AI advertising service into four deliverables:

1. Ad fact cards: price and applicable conditions, inventory or service regions, delivery timeline, refund rules, and exceptions that require human confirmation. Both the ad copy and the dialogue answers must follow these.
2. Question map: questions that can be answered directly; questions that require collecting information and then handing them to sales or customer support; and questions that must be refused or require a human confirmation prompt—each clearly stating the responsible party.
3. Review checklist: for discounts, efficacy, compliance, price, region, image copyright, and translation—what changes must be re-approved. AI-generated content doesn’t mean it’s automatically allowed to be run as ads.
4. Lead handoff page: source, product, the user’s key questions, current status, next person responsible, and the expected follow-up time. Whether or not HubSpot is used, the handoff rules should be accepted separately.

Quoting can also be broken into “asset and facts collation,” “dialogue and review workflow design,” “CRM fields and handoff configuration,” and “post-launch verification.” This way, the customer can understand the service scope, and the service provider won’t blend unlimited times of copy changes, indefinite follow-ups, and platform ad delivery fees into one bundled price.

Don’t promise “using AI can lower customer acquisition costs” or “connecting an Agent will increase conversions.” This release didn’t provide those results. What can be verified is whether the materials align, whether the questions have clear boundaries, whether approvals are traceable, and whether the leads are picked up by someone; ad delivery performance still needs to be validated together by the actual account, budget, product, and market.

Sources: OpenAI, (Reimagining advertising with AI), September 16, 2026.