ChatGPT Ads Manager: the plugin with write access

For years, buying digital media meant opening an interface: log into Ads Manager, touch a control, save, confirm. That sequence is what OpenAI just rewrote for its own ads platform. The product changelog that reached advertisers describes a plugin that lets you "create, manage, and analyze campaigns from ChatGPT or Codex," using plain language to turn a website or brief into ready-to-launch ads, create variants, update campaigns, troubleshoot delivery, and get prioritized recommendations. The text adds a detail that matters: "the plugin previews recommended changes and asks for confirmation before applying them."

Read that twice. This is not an assistant that answers questions about an account someone else keeps operating by hand. It is a layer that does the work and needs permission to touch the budget.

From moving controls to writing the request

The plugin's listing on chatgpt.com confirms who built it and what it can do: OpenAI built it, and it has read and write capabilities over the connected account. Among them, discovering associated ad accounts, creating and managing campaigns, ad groups and ads, uploading creatives, and analyzing performance and conversions. It is, in effect, the full buying layer of a channel exposed as a conversation: instead of navigating a panel of tabs and filters, someone describes what they want to achieve and the plugin translates that into concrete changes on the account.

The work of buying media on that channel stops being about moving controls one by one. It becomes writing the request with precision, reviewing what the agent proposes before it applies, and approving or correcting it. It is the same kind of shift we already described when every ads platform started speaking MCP: the account becomes operable by an agent, with its own permissions and read-and-write access to what only a person with panel access used to touch. The difference with this plugin is that nothing needs to be integrated from outside: OpenAI offers it natively, inside ChatGPT.

What the sources confirm, and what nobody has said yet

It is worth being precise about where each fact comes from, because the temptation to fill gaps with plausible detail is high here and the ground is still undocumented. We reviewed the entire ChatGPT Ads help center collection today, September 3rd, and there is not a single article about this plugin. Everything that can be stated with confidence comes from the product changelog and the plugin's marketplace listing, not from support documentation that would explain permission levels, setup requirements, or enabled markets. We do not know which markets have the plugin available, whether a specific role is required within a team to connect it, or how each applied change is audited. Any categorical answer on those points today would be a guess dressed up as a fact.

The same pattern we already saw, with a costlier object

This blog already wrote about an equivalent architecture: PostHog's self-driving mode, where an agent proposes changes and a human approves before they apply. There, the object of the proposal was product code, something that enters production under review and, if something goes wrong, can be reverted with a rollback. A pull request that never merges simply never happened.

With the ChatGPT Ads Manager plugin, the object of the proposal is live media budget. A badly approved bid change does not get reverted: it got spent. The logic of "the system previews and asks for confirmation" that OpenAI's own changelog describes is the same gesture as approving a pull request, but applied to a domain where the mistake is not contained in a staging environment.

That changes what approving means. In code, the human review gate is a best practice on top of a process that is inherently reversible. In paid media operated by an agent with write permissions, that gate stops being an optional layer and becomes the main control: the one point where a human sees the change before money goes out because of it.

The questions a team needs to resolve before granting write access

Without support documentation, this section does not describe how the plugin works under the hood. It raises what a paid media team should have resolved before connecting it to a real account.

Who on the team has the plugin connected to the account, and is that access separate from normal Ads Manager access or does it run in parallel?
At what permission level was the account connected: can the plugin raise budget, pause campaigns, touch audiences, or is there a configured ceiling?
What gets logged for each change a human approved: who approved it, when, and what justification was recorded for that decision?
How does anyone distinguish, six months later, a change the plugin made from one a person made directly in the panel?

None of those has a public answer today. It is worth resolving them before connecting the plugin to an account that spends real money, the same way it is worth having an objective measurement gate before letting any system, agentic or not, decide on its own whether something worked. And it is worth watching how ChatGPT Ads availability keeps moving overall: the platform keeps adding markets in complete blocks, with no reason to assume this plugin will roll out on a different curve.

Who this is for

For paid media teams already running or evaluating ChatGPT Ads who see this plugin as a way to speed up day-to-day operations: it is worth separating, from day one, who has the connected access and at what permission level, before that gets decided by default. For whoever owns tool governance inside a marketing team: it is worth writing those four questions into an internal document now, rather than waiting for support documentation to answer them. And for anyone tracking how agentic media buying is progressing overall: this plugin is one more piece of the same pattern already visible across other channels, now applied directly inside the channel where the conversational agent itself was born.

This content was developed with AI assistance and reviewed by the Zenda team. The bad ideas are 100% ours.