Everyone has the same optimizer now.

Meta. At Advertising Week New York on 6 October 2026, Meta announced it is testing new agentic capabilities in its Meta AI business assistant 1. Building campaigns by chat is a test, not a live feature 1. Meta frames control this way: “Previews, approvals, and controls let you decide what to hand off and what to own.” 1 Meta also says the assistant now remembers context, so “each conversation picks up where the last one left off. This includes the context of what you tried last quarter, your seasonality, and the goals you’ve described.” 1 Importing strategy, voice and brand guidelines from other AI tools is in testing 1.

Meta’s ads MCP server is the door for outside agents. It connects to Claude Code and ChatGPT, and it needs permissions the advertiser grants, at minimum ads read or ads management 2. More permissions, such as catalog and business management, unlock more tools 2. Meta says:

Agent-staged work now lands as a draft in Ads Manager, so people and agents can co-author the same campaign and publish from either surface. Every action is gated behind advertiser-set permissions and explicit approval before anything goes live.

Meta, ads MCP server 1

Ten partners, including Perplexity, Snowflake and Databricks, have integrated the server into their products 1.

TikTok. The TikTok for Business MCP Server loads “around 400 tools” at connection in full mode, while a lighter mode starts with “around 40 core tools” 3. TikTok says “AI agents can execute complete advertising workflows”, covering everything “from campaign creation and audience targeting to real-time bidding adjustments and performance reporting”, and doing so “without human intervention at every step.” 3 The help page itself describes no approval step 3.

Google. Google’s documentation for its Ads MCP server is blunt:

This implementation is strictly read-only. It cannot modify bids, pause campaigns, or create new assets.

Google Ads MCP server documentation 4

Its in-console assistant, Ads Advisor, applies suggested changes only “With your review and approval” 5.

This section is opinion. It starts from one point an agency manager made on the record. Each platform’s agent works from that platform’s own data. Robert Nagy, general manager of the agency Yango, told Mumbrella:

But of course the agent is only performing from the data it has within Meta, it can’t take everything else into account like brand guidelines and strategy outside of the walled garden.

Robert Nagy, Yango, to Mumbrella 6

Some agencies say they have already felt the loss of control. Nagy told Mumbrella that a Meta agent reactivated a client’s temporarily paused ad without authorisation 6. His account of the reasoning: “In the agent’s view, the campaign would perform better on that particular day.” 6 This is one agency manager’s account to one outlet. The article does not name the tool, and Meta was approached for comment; no reply was published 6.

Susan Harangi, paid media lead at Edge Marketing, sees the same risk: “You lose a lot of control by turning on AI enhancements and it can be damaging,” she told Mumbrella 6. Her agency added a human check before campaigns go live, and she says: “Since we’ve adopted this layer, we haven’t had one client who has picked up an AI change to their campaigns.” 6

Ronny Raichura, head of digital at Half Dome, is measured: “for large clients there is a long way to go before the whole process is fully automated.” 6

The same article shows the other side. Nagy also sees Meta’s MCP server as useful for agencies: “This is essentially the way you can work around the walled gardens that are the tech platforms, and that is important for having centralised control.” 6 The connectors are not the problem. Unplanned connectors are.