On October 6, at Advertising Week New York, Meta said its AI business assistant inside Ads Manager now has “a deeper memory of your business.” Each conversation picks up where the last one left off, including “what you tried last quarter, your seasonality, and the goals you’ve described.” In the same post, Meta said it is beginning to test an assistant that can plan and create a campaign, update targeting, or change a budget just by being asked.
Most coverage led with the budget part, because that is the dramatic part. For a team that shares an ad account, the more important line is in Meta’s help center, in the article on memory settings. Memories are tied to your user account and to the specific ad account where the chat happened, and they “are personal to you and won’t be seen by other users in your ad account.”
So the assistant does not remember the account. It remembers your version of the account. If a brand manager, an agency buyer, a freelancer and a regional lead all talk to it about the same account, it keeps four separate pictures of what the goals are. Today that shapes recommendations. Once the agentic features leave testing, it may shape changes.
A note on evidence. This is a primary-source review. I have not used the new features in a live ad account for this piece, and the agentic actions are not generally available to test. Everything below comes from Meta’s October 6 announcement, its April 22 expansion post, three Meta Business Help Center articles (the assistant overview, memory settings, and the ads AI connectors guide), and two Meta developer pages on the ads MCP server, all read on October 12. I cross-checked against Social Media Today, MediaPost, BestMediaInfo, Relevant Audience, Mockingbird, and PPC Land.
What it is
The Meta AI business assistant is a chat panel that lives where advertisers already work: Ads Manager, Meta Business Suite, and Business Support Home. Meta launched it with small businesses in the US in October 2025 and widened the beta to advertisers and agencies of all sizes across major markets on April 22, 2026. It is still labeled a beta.
It reads your campaign and account data. The help center lists what it does today: charts and trend summaries, analysis of the campaigns you select, comparisons against industry benchmarks for cost per result, click-through rate and return on ad spend, budget and audience recommendations, help with pixel and Conversions API setup, and fixes for disabled accounts, spend limits and payment errors.
There is a second door into the same account. Meta’s ads MCP server lets outside AI tools such as ChatGPT, Claude, Claude Code and Perplexity, plus partner products, read reporting and create or edit campaigns, ad sets, ads, custom audiences and catalogs. Meta opened that server to any developer with a Meta app on July 16. The two doors have very different controls, which matters below.
What changed on October 6
Five things for the assistant, with three different statuses.
Live, according to Meta: creative insights from Ads Creative Studio now appear inside the assistant, and the deeper memory is described as something Meta has already shipped.
In testing: importing “the strategy, voice, and brand guidelines from the AI tools you already use,” and scheduled tasks and alerts for performance monitoring, weekly summaries and reminders to review budgets.
Coming “soon,” and only beginning to test: planning and creating a campaign, generating creative images and text, updating targeting, and changing a budget, all by conversation. Meta says “previews, approvals, and controls let you decide what to hand off and what to own.”
For the MCP door, Meta said agent-staged work now lands as a draft in Ads Manager, so people and agents can co-author the same campaign. Agents can also upload creative, preview ads as they will appear on Facebook and Instagram, and track performance. Meta named ten partners that have built it into their products, including Databricks, Snowflake, Kochava, Perplexity and Triple Whale.
What works
The memory has real controls. You can tell the assistant “Remember this metric” or “Forget this audience” in plain language, and you can review and delete every memory under Settings, then Memory, inside the chat panel. That is more visible than most AI memory features, where you cannot easily see what the tool kept.
Memory stays inside one ad account. The help center says memories differ on each ad account you use. For an agency buyer who works across ten clients, that is the right design. What one client told the assistant about margins should not show up in a recommendation for another client.
Creative insight now sits where budget is decided. Ads Creative Studio, which Meta introduced this summer and is now rolling out to more advertisers, shows which creative is working by objective. Putting that inside the assistant means a planner asking where to move money can see creative performance in the same conversation, instead of exporting one report into another.
The MCP door has hard, server-side limits. This is the most concrete control in the whole package. Meta’s developer guide describes rules that deny specific agent actions on a specific ad account: creating campaigns, ad sets or ads; editing targeting, creative or delivery status; or any budget increase above a percentage, above a fixed amount, or above a ceiling. The help center example is simple: set a maximum of USD 1,000 and any agent request to set a budget above that is refused. Rules can be managed one account at a time in Business Suite settings or programmatically across many accounts through the Marketing API. The connectors guide adds that ads created this way start paused until you set them live.
Where it breaks or is fenced
Private memory means private goals. The assistant remembers “the goals you’ve described,” and each person describes their own. A client marketer who told it in August that the goal was new customers at any cost, and an agency buyer who told it last week to hold cost per acquisition flat, are both right in their own memory. Neither can see the other’s. The help center describes no shared account memory, and Meta has not said whether the guidelines import, still in testing, will be personal or shared. Until it says, assume personal.
The 5% figure is not a forecast. Meta says advertisers “who adopt its recommendations are typically seeing 5% more conversions at the same cost per conversion than those who don’t.” Relevant Audience and Mockingbird both point out that Meta published no method, period or sample. Advertisers who choose to follow recommendations may differ from those who do not. In April, Meta cited a 12% drop in cost per result for small businesses that applied opportunity score recommendations, also without a method. Treat both as platform claims.
The agentic controls are a sentence, not a spec. “Previews, approvals, and controls” is all Meta has said about the in-Ads-Manager agent. It has not said whether the MCP rules above also apply to the assistant inside Ads Manager, whether approval can be switched off, or whether a budget change made by the assistant is marked differently in account history. Meta’s July developer post describes starting an account “in a read-only or approval-required posture” and widening permissions “as trust develops,” which suggests approval is a setting rather than a constant.
The strongest guardrail is not available to everyone. Meta’s developer guide says MCP server rules “are in limited availability,” and the help center says that if you cannot see the Ads MCP server option in settings, you do not have the feature yet. On the advertiser side, managing them takes full control of the business portfolio. And when an agency sets rules on a client’s shared account, those rules apply only to the agency’s own employees. The client and the agency can hold different rule sets on the same account.
Agents are already doing real work. PPC Land reported on October 7 that Ads Uploader, a bulk upload tool, said 17.5% of the 36,289 Meta ad batches submitted through it in September came from AI agents. One vendor, one month, unaudited. It still means this is not a 2027 problem.
Beta terms apply. Both assistant help pages I read carry the same note: guidance is a suggestion with no guaranteed outcome, interactions “may be used to improve AI at Meta,” and using it means agreeing to Meta’s beta and AI terms. If a client contract limits what account data can be used for, read those terms before the guidelines import arrives. Meta also listed no countries for any of the October 6 features.
Who it is for right now
It fits best for in-house teams with one or two people running the account, who already ask the assistant for reporting and want it to stop forgetting what they tested last quarter. Small advertisers who live in the panel will get the most from memory, and Meta built Muse for Small Business for the owner who wants more of the work done for them.
It fits less well, today, for accounts with many hands: a brand plus an agency, multiple regions, or contractors. There, private memory produces private strategies. Agencies that already run their own agents should look at the MCP door and its rules first. That is where the documented limits are.
What it costs
Meta lists no price for the assistant, the memory, Ads Creative Studio, or the ads MCP server, and Relevant Audience noted that Meta did not say whether the new features carry a cost. What you pay is your ad budget, which is exactly what the agentic features are being built to move. Third-party tools that sit on the MCP door have their own pricing.
What to do this week
Treat the assistant’s memory like a strategy document that each person keeps privately, because that is what it is.
- Write one shared account brief per ad account: goal, target cost per result, seasonality, what was tested last quarter, and what is off limits. Paste it into the assistant at the start of each conversation, so everyone works from the same version.
- Ask every person who uses the assistant on a client account to open Settings, then Memory, and delete anything that contradicts the brief. Make it a monthly habit.
- Decide now who may ask the assistant to change a budget once that feature reaches your account. Write it down before the button appears.
- If you have full control of the business portfolio, check whether the Ads MCP server option appears under Integrations in Business Suite settings. If it does, set a budget ceiling and block status changes for every account an agent can reach.
- Agencies: ask each client what rules they have set, and share yours. Rules on one shared account do not cover the other side.
- Do not plan around the 5% figure. If you want to know whether recommendations help, apply them to half your campaigns for a month and compare.
The October 6 update makes Meta’s assistant more useful, and its memory controls are better than most. But a shared account with private memories is a team with private strategies. Get everyone on one brief this week, while the assistant is still asking before it acts.
