A chatbot used to be a sounding board for advertising: you pasted in numbers and it gave advice. Now it can be let inside. Meta announced this on April 29, 2026: in an open beta, ad accounts connect to ChatGPT and Claude, and these services work with live data rather than with copies you made. It sounds convenient. But with such a connection the assistant does not only read, it acts, and an owner who clicked “Allow” for a minute usually does not know what exactly was allowed.
Below is what we managed to confirm on Meta’s official pages and in Claude’s help, plus three things worth doing today. Whatever we could not confirm is marked plainly, so you do not take someone else’s retelling on trust.
What this connection is
By Meta’s description, “ads AI connectors” let you manage advertising from the AI tools you already use. PPC Land names two services: ChatGPT from OpenAI and Claude from Anthropic. Digiday writes that availability depends on your plan inside the AI service itself.
Meta lists four areas where the assistant works:
- reports: it pulls figures and explains what is happening with your ads;
- campaigns: it creates and edits ads, ad sets and campaigns from a description in plain language;
- product catalogs: it creates a catalog, adds product data and hunts for feed errors;
- signals: it checks whether data about site events reaches Meta and how good that data is.
PPC Land points out separately that, unlike similar connections from Amazon and Google, which launched mostly in read-only mode, Meta offers write access from day one.
What exactly it can change
Meta’s developer documentation describes the toolset in more detail. It shows that the assistant can:
- create campaigns, ad sets and ads, with an ad set given its audience, budget and optimization goal right away;
- update fields of an existing campaign, ad set or ad;
- switch on what was paused: the activation tool is described in the documentation as one that “starts spending budget”;
- change and delete your audiences: update a name, rule or labels, add or remove people on a customer list, delete an audience entirely;
- boost Instagram posts, turning them into ads.
There is no separate “pause” tool in the list, and the page does not say which fields an update to an existing campaign may change. So the claim that the assistant can change the budget of ads that are already running is something we cannot confirm from the official pages. There is an indirect hint: in its July update Meta says its new rules help govern what AI does “from budget changes to catalog updates”. The practical conclusion is simple: until shown otherwise, assume the account is open to the assistant more widely than you would like.
Where Meta protects you and where it does not
One safeguard is in the documentation. New campaigns, ad sets and ads that the assistant creates appear paused, and the AI client is supposed to ask for confirmation before activation. That is good, but note what it protects against: launching something new. Whether the assistant will ask permission before every change to ads that are already running is not stated on the Meta pages we read. So “it will always ask” should not be treated as a guarantee.
The second safeguard sits on the side of the AI service itself. Claude’s help says that for each permission category or individual tool you can choose Always allow, Needs approval or Blocked. Tools are grouped by type, for example read-only and write or delete. On Team and Enterprise plans the account owner can set this for the whole team, and individual users cannot override it. This switch decides whether one careless “yes” is enough. If it says Always allow, nobody will ask you anything. For ChatGPT we could not open a comparable page (OpenAI’s help returned an access error), so we claim nothing about its modes: check them yourself in the settings of connected apps.
The third safeguard is new. As of July 16, 2026, Meta announced “rules” for connected AI agents: anyone with full control of a business portfolio can govern what agents may do in an ad account. Where exactly these rules sit in Meta’s menus is not stated on the pages we cited, so we do not invent a path. Ask the person who manages your portfolio whether they are switched on.
Why changes to ads are not free
There is a reason even an “innocent” edit by an assistant can cost money. PPC Land quotes advice from a Meta representative: remind your AI about the learning phase. Each significant change, such as to the budget or the audience, can reset an ad set’s learning, and the system again spends some time working out whom to show the ads to. An assistant that “improves” settings every day can keep your ads in that state permanently. We wrote about this at more length in the piece on editing ads every day.
What to do today: three steps
- Give the assistant read-only work. Ask it to assemble reports, explain why enquiries dropped and draft a list of ideas. Make the changes yourself, in the account, seeing every number. In Claude it is enough to set write tools to Blocked, or at least to Needs approval. When you read a confirmation request, look at what will be changed, not at the word “launch”.
- Review the list of connections. In Claude, according to the help, it is Customize > Connectors: for each service you can disconnect it, change connection settings or review permissions and access levels. Claude’s help advises disconnecting services you no longer use. Once a month it is worth looking there and in ChatGPT’s settings, and asking whoever manages your Meta portfolio which AI apps have access to the account.
- Look at the wider picture of access. If a contractor runs your ads, agree whether they may connect AI assistants to the account and at whose expense. Separately, check which people still have access: our article on the former contractor who is still an admin covers that. And what is safe to hand to AI in general is covered in the text on five tasks for an owner.
Summary
Connecting AI to your advertising is neither forbidden nor harmful in itself: an assistant really can read reports well. The risk is that since April 2026 it can also create, launch and edit. Meta guards the launch of new items, while the rest is governed by your AI service’s settings and by portfolio rules. Until you know what those are set to, the sensible choice is to let the assistant only read.
You can read how we set up advertising in the search advertising and Performance Max sections.
Want a free audit of your Google Ads account? Send us a request — we will look at your account and tell you who has access to it and what is worth fixing.