Meta AI ads can now run the account. The guardrails are still yours to write

What Meta actually announced for Meta AI ads#

Headlines compress. A fox on the prowl reads the tracks, not the tale told about them, and the tale this week is that Meta AI ads can now run your campaigns for you. The tracks say something more careful. At Advertising Week in New York on 6 October, Meta's own announcement says it is "beginning to test" new capabilities in its business assistant. Testing is not launching, and that one verb is the gap between the headline and the facts.

What does the assistant get? Meta says it now holds the context of what you tried last quarter, your seasonality and the goals you have described. Social Media Today reported that the next step lets the assistant generate creative, update targeting and even change a budget through conversational prompts. Only one item in the package is unambiguously live: advertisers and agencies can now generate video from existing static assets.

That post is the headline version of Meta AI ads, and it is worth reading slowly because it folds three release states into one verb. We checked the Meta half against Meta's own words and could not verify the TikTok half, so we leave it alone. The table below separates what is live, what is in testing and what is limited, because an advertiser who plans around the wrong column will be surprised in the wrong direction.

What an outside agent can touch through Meta's Ads MCP server#

Behind the newest Meta AI ads sits the model context protocol, MCP, which is an open-source standard for connecting AI applications to outside systems, and Meta's server turns an ad account into one of those systems. Meta's developer documentation says the server works with any MCP-compatible AI agent, which is a wide door. The same developer blog lists what walks through it: agents can create, edit and delete campaigns, ad sets and ads.

Two safeguards matter for Meta AI ads, and they are not the same kind. The first is approval. Meta's tool documentation says write tools create entities in a paused state, and that your AI client asks for confirmation before activation. Read that carefully: the confirmation lives in the client, which is a convention, not a lock on Meta's side. The second safeguard is the rules endpoint, where a rule denies a specific agent action on a specific account. Meta's example is to deny any budget increase above 20%. That is a genuine guardrail, and it is labelled limited availability.

There is one more useful lever for audits. Meta's server documentation includes a tool to get visibility into activity log changes for an ad account, mirroring the history page in Ads Manager. If an agent changed something, there is a record, provided somebody reads it. The rules documentation also notes that there is no batch endpoint, so an agency with fifty accounts sets rules fifty times.

Guardrails first: Meta ads automation that survives an audit#

Meta is not the only party saying so about Meta AI ads. The MCP specification for tools states that there should always be a human in the loop, and its security guidance warns that poor scope design obscures audit trails. Meta's own developer blog offers the matching advice: start in a read-only or approval-required posture and widen permissions as confidence grows. That is the whole playbook for Meta ads automation in one sentence.

Meta also describes a human-readable version of the same idea. Its announcement says every action is gated behind advertiser-set permissions and explicit approval, and that agent-staged work lands as a draft in Ads Manager. Drafts are a good habit. They turn an agent's output into something a person can read before it spends.

Meta itself is candid about the risk it is selling into. Its latest quarterly filing lists, among AI risks, agents acting beyond their intended scope. A company writing that sentence about its own agents is giving you permission to be careful with yours. For agencies running many accounts, this is also where a paid social partner earns its fee: not clicking the buttons, but owning the rules, the log and the review rhythm.

What the evidence says about Meta Advantage Plus and automation#

Automation arguments always end at the same question: do Meta AI ads work? Meta says yes, with caveats it does not stress. Its announcement says advertisers who adopt the assistant's recommendations are typically seeing 5% more conversions at the same cost per conversion, but the comparison is with advertisers who do not adopt them, which is a self-selected group. Meta's September case-study page reports one advertiser with 64% lower cost per purchase. Those are advertiser-reported results, offered as marketing, and should be read that way.

The independent view of Meta AI ads automation is more mixed. The measurement firm Haus ran 640 incrementality experiments and published the result in July 2025, so it is more than a year old and predates the current tools. Its headline finding: Advantage+ performed 9% better than manual campaigns at the midpoint but 12% worse by the end. Haus sells measurement and its customers self-select, so this is a trade study, not a peer-reviewed one.

Haus is careful not to turn that into an indictment. The same report says for 42% of brands in the study, Advantage+ outperformed manual on incremental return, and the average advertiser in its sample spends $14 million a year on Meta. That is a larger budget than most readers manage, so apply the lesson, not the number.

The practical answer is boring and powerful: run your own holdout. Take one audience, split it, let Meta's automation run on one half and your manual set-up on the other, and read the result at the midpoint and at the end, because Haus's data says the two can disagree. Even before agents arrived, researchers found that ad delivery systems can skew outcomes in ways advertisers did not choose, which is a reminder that optimisation has always been a negotiation with the platform.

AI ad creation and the European rulebook#

Generated video is the live item among the Meta AI ads, so AI ad creation is where compliance meets craft first. Two EU rules do not care which tool built the ad. Article 26 of the Digital Services Act, as reproduced at eu-digital-services-act.com, requires that people can tell an ad is an ad, and that they can reach meaningful information about the main parameters used to determine the recipient. If an agent changes your targeting overnight, you still need to be able to explain who saw the ad and why.

A short word on tools, because the question is always asked. People searching for the best AI for ads are usually really asking which interface to trust with money, and the honest answer is that the choice of model matters less than the quality of the account's guardrails. For measurement discipline across channels, our notes on Pinterest's new auction and the TikTok ad network opening to US advertisers follow the same rule: add the report column before the spend arrives. And when the agent in question is a language model with its own limits, our AI consultancy work helps teams decide what to hand off at all.

Meta AI ads will keep moving, probably weekly, and Meta AI ads guidance will date as fast. The durable move is to own the parts that do not change: the permission ladder, the deny rules, the weekly read of the log and the holdout test. A moonlit hedgerow looks the same whichever bird is flying over it, and a good account structure does too.

Frequently asked questions#