Marko Tomic

System 01 · AI systems

Google Performance Intelligence

Repeated performance analysis across disconnected Google platforms, rebuilt as an autonomous agent workflow that saves an estimated ~9.5 analytical hours per week - while every account change remains a human decision.

StatusAutonomous · cron Mon + Thu IntegrationMCP + OAuth + API PermissionsRead-only by design

Problem

Analysis was scattered across three platforms

Performance questions required repeated manual analysis across Google Ads, GA4 and Search Console - three disconnected interfaces, each with its own logic. The work was repetitive and consumed specialist analytical capacity that should have been spent on decisions, not data assembly.

This system is part of the broader AI Operations architecture. The Performance Intelligence system extends the earlier Performance & Growth transformation.

Architecture

Specialist agents, one orchestrator

Three specialist agents - one each for Google Ads, GA4 and Search Console - connect via MCP + OAuth + API access with strictly read-only permissions. A main orchestrator reviews their reports, condenses them, highlights the most important changes, saves a local archive and sends a summary to a shared Google Drive for the external advertising specialist.

  • Input
  • Agent
  • Orchestrator
  • Governance gate
  • Output
  • Human judgment
Fully autonomous via cron, running every Monday and Thursday. Subagents write local reports; the orchestrator condenses and distributes. The permission boundary is physical, not a policy: the agents cannot write to an ad account.

Governance

Read-only is a decision, not a limitation

The AI does not directly alter advertising accounts. The external specialist retains execution authority - and that is intentional. Analysis is a repeatable production task; budget and account changes are judgment calls with financial consequences. The permission boundary sits exactly between the two.

The tradeoff that shaped the architecture: write access would have made the system faster and the analysis directly actionable - at the cost of wrong decisions executed confidently, at scale, twice a week. Read-only gives up that speed and keeps the failure mode harmless.

Impact

What it changes

~9.5 h/wk Analytical hours saved

Estimated average across the combined Google-platform analytical workload.

The output regularly informs budget reallocations, performance investigations and optimization priorities. No financial attribution is invented for this system - its value is that better analysis arrives twice a week without anyone producing it.

The design question this system answers is where the human checkpoint should sit when the analysis is cheap but the actions are expensive.

Related