← All articles

AI

AI agents in Google and Meta Ads: what we let them touch

Meta and Google put AI agents inside ad accounts, changing budgets and targeting on approval. What we let them touch, the approval gate, and the log we keep.

Pierrick NonnisPierrick NonnisCo-founder, COO6 min read
An operator seen from behind, seated in the dark before a wall of monitors

Meta’s assistant in Ads Manager will soon change your targeting and budgets from a chat window, and Google’s Ads Advisor has applied changes to Performance Max campaigns on approval since December. For a media team, that is the pacing, the Monday report and the disapproval chase handed to something that remembers the account and waits for a yes. In October, Meta also drafted terms that make you solely responsible for whatever its AI changes, then paused them before they could take effect, while a UK regulator opened an investigation. So the agent gets the keys, on conditions written before it arrives. Here are ours.

What Meta’s Ads Manager assistant and Google’s Ads Advisor can do now

On 6 October, at Advertising Week New York, Meta said the business assistant in Ads Manager now picks up each conversation where the last one ended, and that "soon, you’ll be able to plan and create a campaign, generate creative imagery and text, update targeting, or change a budget conversationally". Teams that bring their own AI go through Meta’s ads MCP server, in open beta since April: anything an agent stages lands as a draft in Ads Manager with explicit approval before it goes live, and whoever controls the business portfolio sets rules on what agents may do, "from budget changes to catalog updates".

Google got there first. Ads Advisor has been live for English-language accounts since December 2025 and, on Performance Max, "with your review and approval, it can then apply these changes directly to your account". Since May, Ask Advisor has been in beta for English-language accounts across Google’s Ads and Analytics agents, with a Merchant Center agent announced next; Google’s own example of where this goes is asking for new customers for your hair care products and letting it set up the campaign from the product data.

Google’s advisors are English-only for now, and Meta’s own FAQ says its assistant launched in beta in English and is scaling through 2026, which leaves a French-language account a little time to write the permissions list.

Why we want the agent in the account

An agent that remembers the account and waits for a yes is the right shape for the work that eats a media team’s week, including the Tuesday where spend doubled for no visible reason, and a team that keeps it out in 2027 is choosing to be slower. Meta says advertisers who adopt the assistant’s recommendations typically see 5% more conversions at the same cost per conversion; no method is published, so read it as the platform’s number, and a plausible one. We listed in September ten jobs agents already do in a growth team and drew the line at budgets; the platforms have since built, inside the account, the approval gate that moves that line, along with their own idea of what counts as a result, hence the log further down.

Meta’s new AI terms: the advertiser answers for the agent’s mistakes

Meta’s revised Commercial Terms, previewed on its legal pages, carry a new "AI Features" section. You acknowledge that those features "may make mistakes and may not always produce accurate, complete, or suitable results", and you become solely responsible for all risks associated with them and with the content they produce. Adweek’s reading, reported by PPC Land, is that the previous wording held you responsible only when you chose to use Meta’s AI creative tools; the new text has no such condition. Nobody signs anything: continued use after the notice counts as acceptance, which lawyers call efficient.

Then the dates moved: planned for 30 October, pushed to 20 November on the 23rd, paused on the 28th pending a consultation with industry groups, and on 30 October the UK’s Competition and Markets Authority opened an investigation into whether Meta’s terms, existing or revised, amount to an abuse of a dominant position, its stated concern being the clause that lets Meta crawl businesses’ websites for its own products, AI features included. Meta’s legal page still shows 20 November. Meta’s answer, that each Advantage+ feature can be switched off in Ads Manager, holds as far as it goes: that switch is also the one practitioners report coming back on when an ad is duplicated. Meta’s liability toward you was already capped at the greater of 100 dollars or what you paid Meta in the past twelve months; nobody has written a cap for your brand, which is what the four tiers below are for.

What we let an AI agent touch: four permission tiers

Four tiers, set per account before the agent is switched on, across the Meta and Google Ads accounts we run, encoded in Meta’s ads MCP server rules and in the approval Google’s advisor asks for before it acts.

  • Read everything. Reporting, search terms, feed diagnostics, delivery errors, anomalies. No approval needed, and the more it reads, the better its drafts. With your own model over the MCP server, this is also where account data leaves the platform, so read the model’s data terms first.
  • Draft, never publish. Creative variants, audiences, campaign structure, copy. The draft lands in Ads Manager and a person publishes it.
  • Change, with approval and a cap. Budget pacing inside a band written per account as numbers (a share of the daily budget, a monthly ceiling it never passes), bid adjustments inside a written range, pausing an ad set that has spent its written test budget without a conversion. Each one approved in the chat by the operator who owns the account, then logged.
  • Off limits. Conversion goals, attribution settings, tracking, account spending limits, and anything else that changes the number the agent is judged on. An agent does not mark its own homework.

If the agent can edit the conversion it optimises toward, every other control is decoration.

The change log: judge the agent on tracking you own

Change history in Ads Manager and Google Ads lists what the agent did. Whether it helped is a separate question, and the platform answers it with its own attribution. So the log we keep is plain: every agent-made change, dated, with the metric it was meant to move, read a week later on tracking you own. One column for the agent, one for humans, otherwise the account learns nothing from either. Once a quarter, a four-week holdout: an account or a region where the agent only reads, against one where it acts within its band. That is the number a CFO asks for in March, and no platform report contains it.

Keys, on conditions

What the agent buys you: recurring work runs itself, drafts arrive before the meeting, and a senior operator spends the week on structure and creative. Those gains land first in the accounts that were governed before the agent walked in, because the permissions list, the approval gate and the log take an afternoon to set up, and about a year to regret not having. If you want a second pair of eyes on your list, Meta or Google Ads, book a call and we will write it with you.

Keep reading

Want this applied to your accounts?

Thirty minutes, one-to-one with a founder, on us. Bring your numbers.