AI PPC management: what it does and what it cannot do

Three layers of PPC automation

Bid-layer tools — Smart Bidding, target CPA, target ROAS — have been part of Google Ads for years. They adjust individual auctions in real time using signals no human could process. This layer is mature, well-documented, and mostly safe to leave on.

Rule-based automation sits one level up: IF daily spend exceeds $X, pause this campaign. IF CTR drops below threshold, send an alert. These rules are explicit and auditable but brittle — they only fire on conditions you thought to define.

AI-driven management is the third layer, and it is the one people argue about. An AI client reads the account, reasons about patterns, and proposes structural changes: reallocate budget from Campaign A to Campaign B, add a negative keyword list, create a new ad group. The quality of that layer depends entirely on how the tool handles the boundary between proposing and executing.

Why the propose/approve split is not optional

An AI that can execute structural changes without confirmation is an AI that can double your daily budget on a mistaken inference. The model does not know it made a mistake — that is the definition of a mistake. Human review is not a nice-to-have; it is the error-correction loop the whole system requires.

SpendSignoff gives the model two scopes: mcp.read to see the account and mcp.draft to stage a change. There is no mcp.approve scope issued to the model. Approval happens in the dashboard, under your credentials, with a before→after diff you must confirm.

What good AI PPC management looks like in practice

You open Claude and ask "Which campaigns are pacing above budget and why?" SpendSignoff reads the account and answers with spend, impression share, and a ranked list. No dashboard tab-switching.

You say "Draft a bid increase for the top-three converting ad groups in the search campaign." The model stages three before→after diffs. You review each one and approve or reject. The approved changes push live. The rejected ones are discarded with a log entry.

That workflow is faster than doing it manually and safer than handing the AI full write access, because you stay in the decision chain on every dollar moved.

What AI PPC management cannot do yet

• Create net-new ad creatives end-to-end without human review (SpendSignoff V1 stages draft copy for approval).
• Reason about offline conversion data unless you explicitly sync it via the reporting API.
• Catch policy violations before they trigger a Google or Meta disapproval — the model can flag known risk patterns but does not simulate the ad review queue.
• Make attributions decisions across platforms with perfect accuracy — cross-platform attribution in V1 is normalized but not ML-modeled.

Choosing a tool

Ask one question: when the AI decides to change a budget or add a keyword, is there a human step before the change goes live? If the answer is "it depends" or "you can configure that," the default is probably no. Defaults are what happen when you are not watching.