SOP · Automated Ads Management
How our ads manage themselves
An AI agent runs every Monday, Wednesday, and Friday. It pulls live data from Meta, Google, and Triple Whale, analyzes performance, makes safe changes, and emails a report. No manual exports, no spreadsheets, no forgotten optimizations.
Stage 01 · Check
Preflight
Loading...
Why it's built this way
  • AI agents, not dashboards. A cron job spawns a Claude Sonnet agent three times a week. It pulls live data, analyzes it, executes safe changes, and emails a report. The work happens whether anyone is looking or not.
  • Artifact-driven pipeline. Each stage writes a structured JSON artifact before the next stage reads it. Every decision is inspectable and reversible. Re-run a single stage without redoing the whole pipeline.
  • Analysis separate from execution. Recommends first, executes second. The agent can be aggressive in analysis without being dangerous in action. Risky changes always require human approval.
  • Self-reviewing feedback loop. Every action gets a scheduled review date. The system checks its own past decisions before making new ones. Most ad tools skip this entirely.
  • Cross-platform truth. Meta and Google both over-claim. Triple Whale showed only 5 of 97 orders had a Meta touchpoint. Google touched 26. Without blended attribution, budget decisions would be based on inflated numbers.
  • Institutional memory. Insights and Creative DNA files carry forward across every run. The system remembers what works and what doesn't without being told each time.
  • Locked guardrails. Playbook thresholds can't change until 4+ weeks of data and 3+ reviewed actions support it. Max 10% shift per update. One fluky week can't rewrite the rules.
  • Git as the audit trail. Every run commits snapshots, artifacts, logs, and reports. Version control for decisions, not code. Trace any change back to the exact run, data, and reasoning.
The bigger picture
What this actually enables
The 8-stage pipeline isn't the end goal. It's the foundation for something bigger: a unified e-commerce brain that merges every major data source — Google Ads, Meta Ads, Triple Whale, Shopify — into one structured, queryable knowledge base that updates itself three times a week.

Because every run writes clean snapshots, structured artifacts, and rich memory files, you can talk to the entire system in plain English. Ask "what's our real CAC in South America after refunds?" and it pulls from Triple Whale attribution, Google Ads geo data, and Shopify order history in one answer. Say "lower the tROAS on Europe PMax to 2.0" and it executes the API call, logs the change, and schedules the impact review automatically.

This isn't a dashboard you stare at. It's a conversational interface to your entire ad operation. The data is already unified. The audit trail is already built. The feedback loops are already running. You just talk to it.

What used to take an analyst a full day — pulling exports from three platforms, normalizing the data, comparing attribution models, writing up recommendations — happens in 5 minutes, three times a week, with no human involvement. And when you do step in, you're not starting from a blank spreadsheet. You're picking up a conversation with a system that already knows your account history, your creative patterns, and what worked last time.
FAQ
How often does it run?
Three times a week — Monday, Wednesday, and Friday. Google Ads at 10:30 AM PT, Meta Ads at 11:00 AM PT. Each run takes 5–10 minutes.
What can it change without asking?
Only low-risk, reversible actions: pausing ads that burn money, adding negative keywords, decreasing budgets. On Google, it auto-executes negative keywords. On Meta, no changes at all — everything waits for human review.
What requires human approval?
Budget increases, new keywords or creatives, bid strategy changes, and audience modifications. Queued as recommendations and wait for sign-off.
Why not just trust platform-reported numbers?
Meta and Google both over-claim. Triple Whale showed only 5 of 97 orders had any Meta touchpoint. Google touched 26. Without an independent source of truth, budget decisions would be based on inflated numbers.
What happens if something goes wrong?
Preflight catches broken state. Every action is logged with exact metrics and a 7-day review date. Every run is committed to git. All auto-executed changes are reversible.
Can it change its own rules?
Only with 4+ weeks of history, 3+ reviewed actions, and evidence from both platform and blended metrics. Max 10% shift per update.
How is this different from other ad tools?
Most tools show dashboards and wait, or auto-optimize without transparency. This does both — autonomous safe changes with a complete, inspectable audit trail. It also reviews its own past actions before making new ones.