Meta AI business assistant: what works today and what is in testing

On October 6, 2026, at Advertising Week New York, Meta showed the next step for the AI business assistant in Ads Manager: an agent that builds campaigns, updates targeting and changes budgets through conversation. The news traveled as if it had already shipped. It has not. What Meta released to everyone was a different part of the package; the agent's actions are in testing. For anyone running accounts, the useful work is separating one from the other and deciding which guardrails to demand before letting any AI, Meta's or someone else's, make changes.

What Meta announced on October 6

The announcement mixes features that are ready, features in testing and features that are promised. Sorted by status:

Meta also said advertisers who adopt the assistant's recommendations typically see 5% more conversions at the same cost per conversion. That is Meta's own number, with no published method. And there is an obvious bias: the advertisers who accept recommendations tend to be the ones who already manage their accounts more closely.

What you can use right now

Two things are worth testing this week.

Customer Lifecycle Strategy. When you mark a campaign as acquisition, Meta recommends how to exclude existing customers. If your "new customer" campaign keeps selling to your current base, this goes straight at the problem. Measure before and after: the share of sales from new customers in your CRM, not just cost per purchase.

Video from static images. If you have strong statics and little video, this is a cheap way to get a new format into rotation. Treat it like any creative: run it, compare at the ad level and judge it on results, not novelty.

The rest of the assistant (asking why costs rose, asking for recommendations, fixing account issues) already works as an advisor. It recommends; you apply.

What is still in testing, and why that matters

The agent's actions are the part that changes the job, and they are exactly the part Meta describes as "starting to test" and "soon." In practice that means three things:

• It may not show up in your account, and there is no public date for everyone.
• What appears in the test can change before it reaches all advertisers.
• Scheduled tasks and automatic alerts are also in testing. Do not build client workflows on them yet.

Meta says the actions will come with previews, approvals and controls so you decide what to hand off. That is the right direction. Just do not mistake an announced guardrail for a guardrail working in your account.

An agent inside Ads Manager or an outside AI

These are different tools, and Meta treats them that way. In its Meta ads AI connectors announcement (April 2026), Meta positions the assistant as the place for guidance, issue resolution and recommendations inside Ads Manager, and the MCP connectors as the route for cross-channel insights and custom workflows. Its line is that you do not have to choose.

The difference that matters most is where the data stops. Meta's assistant optimizes inside Meta. It does not know Google brought in the same lead for less, or that leads from campaign X never close. For that question you need an AI that reads your channels and your CRM together. If MCP is new to you, what MCP means for marketers covers it in a few minutes, and Meta Ads MCP server compares running your own with a hosted connector.

The guardrails that apply either way

Meta's agent or someone else's, require the same things before you allow write access:

1. A draft or paused state before anything goes live. Nothing ships without you seeing it.
2. Explicit approval per action. A blanket "go ahead" is not enough.
3. Limits on what the agent can do. In July 2026 Meta added ads MCP server rules: anyone with full control of the business portfolio can set what AI agents may do on the ad account, including budget changes.
4. A dollar cap per action. A typo should never turn into ten times the budget.
5. A change log with before and after. So you can undo it and explain it to the client.

If any of these is missing, keep the AI read-only. AI that executes in Ads Manager lays out the order in which to unlock each action, and AI data security for marketers covers what it can see.

Your next step

Try Customer Lifecycle Strategy on one acquisition campaign and track the share of new customers in your CRM for four weeks. While Meta's agent actions stay in testing, write down the five guardrails you will require from any AI. To read Meta, Google, TikTok and your CRM in one conversation, see how to connect Meta Ads to Claude and the plans, from US$ 31.20 a month on the annual plan.