Flagship 01 · AI systems
Repetitive knowledge work rebuilt around specialist AI agents - with orchestration, controlled autonomy and governance that keeps judgment human.
Leverage
Metadata across four languages. Human baseline ~7 min per localized entity.
Estimated average across the combined Google-platform workload.
Estimate, ~30+ hours - research, content, design, code, localization.
Context
Three categories of work kept consuming disproportionate human capacity: repeated performance analysis across disconnected Google platforms, metadata production across a four-language catalog, and semantic content architecture that required research, design, code and localization for every page.
None of these are judgment problems. They are production problems with judgment checkpoints - which makes them candidates for a specific kind of AI architecture, not for generic chatbot use.
The centerpiece
The mistake would have been one automation pattern applied three times. Each system earned a different level of autonomy, because each fails differently.
Read-only autonomous analysis
Three specialist agents - Ads, GA4, Search Console - report to an orchestrator over MCP + OAuth. Cron-scheduled, fully autonomous, and physically unable to touch an ad account: permissions are read-only by design.
Human checkpoint: the external specialist executes. System →
Calibrated production automation
Every field passed human Excel review until output survived 3-4 calibration runs per language unchanged. Only then was approval removed. Autonomy here was granted by evidence, not by optimism.
Human checkpoint: removed after calibration. System →
Human selection among alternatives
Roughly three competing model prototypes per page - the models genuinely interpret the brief differently. The human selects, critiques and directs; agents execute and iterate. Judgment never leaves the loop.
Human checkpoint: every page. System →
Ownership
Outcome & reflection
The measured leverage is real but deliberately bounded: 220+ hours of avoided metadata production, an estimated ~9.5 analytical hours saved weekly, an estimated four-plus working days saved per four-language hub. No financial attribution is invented - the output informs budget reallocations, performance investigations and optimization priorities, and that is the honest claim.
The less obvious result is organizational: attention moved from production to judgment. The interesting design question was never "what can AI do here?" but "where exactly should the human checkpoint sit?"
Systems & thinking
Marko Tomic