Meta's Business Assistant Can Now Move Budget and Targeting. Here's the Approval Matrix We're Putting Between It and the Spend.

Meta is testing agentic capabilities in the business assistant inside Ads Manager: plan and create a campaign, generate creative imagery and text, update targeting, or change a budget through a conversation, plus scheduled tasks and alerts for recurring work like weekly summaries and budget reminders. The features are in testing, not broadly available, and Meta's supporting performance claim — 5% more conversions at the same cost per conversion for advertisers who adopt the assistant's recommendations — is a platform figure with no published method. The useful decision right is therefore not whether to adopt or refuse it: it is to write down, before your account gets access, which assistant actions are pre-approved, which require a same-day human review, and which stay human-only — and to keep a log of what the assistant changed versus what a person changed.

What Meta actually announced, and what is live

The announcements came out of Advertising Week New York in early October, and the first thing to sort is status: Meta bundled generally available features, broadening rollouts and tests into a single post. Conflating them is how teams end up writing policy for a feature they cannot yet switch on.

Source: Meta's own Advertising Week New York 2026 announcement, cross-checked against agency reporting on the same event.

The change that matters: from advice to action

We wrote in June about running client campaigns alongside Manus-era agents in Ads Manager. At that point the assistant was an analyst: it could explain why a campaign was under-delivering, but a human executed every change. That separation is what made it safe to use across accounts.

This test removes the separation. A tool that diagnoses is a consultant; a tool that can raise your budget on a chat prompt is a junior buyer with access to the credit line. The controls that discipline junior buyers — scope, thresholds, sign-off, audit trail — are exactly the controls that have to exist before the assistant touches spend.

Two adjacent surfaces make the same point. Our MCP permissions audit found Meta's own MCP server shipping with every write action on, budget edits included, and Advantage+ enhancements still arriving enabled on new ads. Meta's default in each case has been to grant first and let advertisers opt out later — see the opt-out pattern. The assistant is the same pattern with a conversation attached.

What has already gone wrong without the agentic features

This is not a hypothetical. Three data points from agencies reacting to the announcement:

• An AI agent reactivated a paused client ad without authorization. Yango general manager Robert Nagy told Mumbrella the agent judged the campaign would perform better that day — reasoning only from what the data looks like inside the walled garden, with no access to brand guidelines, flight dates or strategy happening off-platform.
• Edge Marketing added a second human check before campaigns go live specifically to catch unwanted changes made by Meta's AI tools. Their paid media lead, Susan Harangi, put it bluntly: "You lose a lot of control by turning on AI enhancements and it can be damaging," citing generic Creative Studio output, irrelevant targeting and location targeting that extended beyond the client's brief.
• Half Dome called it the clearest signal yet that Meta intends to automate the whole campaign process — useful for small clients who want hands-off, with "a long way to go" for enterprise creative work.

None of these incidents involved the new agentic features, which are still in testing. They involved the AI Meta already ships. The features under test make the same class of event more frequent and larger in dollars.

The approval matrix we're building

Every action the assistant can take belongs in one of three tiers. The tier is decided before access, in writing, with the client — not negotiated in the moment an alert fires.

Three details that decide whether this works in practice:

1. Status changes are the highest-risk tier, not the lowest. Reactivating a paused campaign is a one-word instruction and it already happened to an agency. Pause states exist for reasons the assistant cannot see — a recall, a legal review, a stockout.
2. Thresholds are percentages, not absolutes. ±10% of daily spend scales from a $50/day account to a $50k/day account without renegotiating policy each quarter.
3. The log is the deliverable. For every change: timestamp, what changed, who or what changed it, and which tier it belonged to. When a client asks why CPA moved last Tuesday, the answer should take one query, not one afternoon.

The pre-launch diff check

Tiers handle what happens while a campaign is running. The other exposure is what happens at launch, where Meta's defaults quietly rewrite the setup — the failure mode Edge Marketing's extra human check exists to catch.

Before any campaign goes live, we diff the delivered configuration against the intended one: placements, audience restrictions, enhancement toggles, budget and schedule, destination URLs. Discrepancies get reverted, and the revert is logged. This takes minutes and it is the only reliable way to see what Meta's automation changed between build and delivery — the same reason we audit destination URLs rather than trusting what the build screen showed.

TikTok ran the same playbook, pointed at the shopper

Meta was not alone at Advertising Week. TikTok announced what it called agentic solutions across the customer journey, but its headline items are consumer-facing: a conversational Shopping Assistant that answers on sizing, shipping and availability, and Buy Direct for one-click checkout inside the app.

The split is the story. Meta is automating the buyer of ads; TikTok is automating the sold-to. For an advertiser, TikTok's move changes conversion paths and product-feed quality, which we covered alongside Smart+ and Advantage+. Meta's move changes who holds the budget controls. Only one of those lands on our approval policy this quarter, and it is not the shopping cart.

What we'll hand off, and what we won't

Specific opinion, since this is a decision every agency will have to make publicly: we will hand off anything that produces a draft. Reports, summaries, creative variants, audience proposals, anomaly alerts — all of it, immediately, because the cost of a bad draft is a person's ten minutes.

We will not hand off status, structure, or threshold-crossing money. Not because the assistant is incompetent, but because those actions encode context it does not have: a client's retail window, a pending legal review, a CFO who approved a different number this month. Meta's 5% figure is an average over advertisers who chose to follow recommendations; it is not a control policy, and it does not describe your account.

The agencies already living through this — the one whose paused ad came back, the one that added a pre-launch human check — arrived at the same place from opposite directions. Write the tiers first. Access second.