Google Ads and Meta with AI agents: what to automate (and what not to)

An AI agent on Google Ads and Meta is a layer on top of the platforms that monitors performance, kills losing creatives, rebalances budget across campaigns, generates new copy/image assets, and alerts a human buyer when a strategic call is needed. It doesn't replace Performance Max or Advantage+ — it governs them from the outside, because neither platform is optimized for YOUR business but for its own.

What we automate with AI (safe): creative rotation based on fatigue and CTR, generating 20 copy variants from a client brief, auto-pause of ad groups with CPA > threshold for 3 days, budget rebalancing across campaigns on marginal ROAS, anomaly monitoring with Slack alerts.

What we NEVER automate: opening new markets, catalog feed choice for Shopping, conversion goal changes, monthly total budget. Those are strategic decisions with month-long consequences — an LLM taking them autonomously is an existential risk for a mid-market client.

Architecture: n8n or Temporal as orchestrator, official Google Ads API + Meta Marketing API connectors (no scraping tools), an LLM for asset generation and to explain decisions in human language, a database for full audit log.

The real value: AI agents do the boring work every hour, every day, without forgetting. A human buyer checks Meta twice a day; an agent checks 24 times and pings you only when needed. On a D2C beauty client this cut CPA by 19% at the same budget and freed 12 hrs/week of the media buyer's time.

How we schedule the agent's work through the week

A well-designed agent doesn't 'do everything always'. It has a cadence: anomaly checks every hour (CPC spike, CTR drop, budget burning too fast), campaign review every 4 hours (marginal ROAS, audience saturation), creative generation twice a week (Monday for weekend test, Thursday for weekend campaigns), morning digest at 8am with the top 3 actions to approve. This structure prevents 400 micro-decisions per day that no human can review.

The media buyer's role doesn't disappear — it shifts

The 'AI is stealing our jobs' crowd missed the point. The media buyer stops doing manual bid adjustments and starts doing strategic planning: which new market to open, which product deserves a dedicated campaign, which creative angle to test at higher risk. It's a level-up — those who don't make it get outperformed by those who do. Across our clients where the buyer embraced the model, ROAS grew on average +22%; where they resisted, it dropped 6% in 6 months.

Why we avoid 'all-in-one' commercial AI tools

The market is packed with platforms promising to run Meta/Google autonomously with AI. We steer clear for two reasons: 1) their models are optimized for the average client, not for yours (with your seasonality, per-product margins, specific goals); 2) they have no access to your real margin data, so they optimize for visible ROAS (which inflates revenue) rather than profit (which pays salaries). A custom agent with real data access costs 30% more per month but typically returns 3x.

Mistake to avoid: giving an LLM API write access with no approvals. Destructive actions (pause campaign, bid change > X%) must go through a policy layer with implicit or explicit human approval depending on threshold.