Claude Code × CData stack cuts ad ops ~70%, finds ¥500,000 waste

The setup is described not as a replacement for judgment but as a way to make previously impractical work feasible. The author draws the line clearly: the stack handles data retrieval, analysis and repetitive tasks, while decisions like ad strategy and which search intents to target stay with the human. That framing matters because the headline numbers — the roughly 70% time cut and the ¥500,000 in waste — are presented as the result of acting on problems the stack made visible, not as outputs the tool produced on its own.

The data-access design reflects that caution. CData Connect AI answers read queries, including a natural-language request to pull ad groups with zero conversions and more than ¥40,000 in spend over the last 30 days. Writes — RSA headline and description updates, pausing keywords, budget changes — are deliberately routed through separate Google Ads API scripts, with a dry run and one human review before anything reaches the production account. The author argues AI does not need write permission at all, and this split is presented as the safety design for using AI on a live ad account.

The author also explains the choice of Connect AI over Google's official Ads MCP server by pointing to cross-platform needs: ads data alone does not capture the picture when Meta, Bing, GA4, Search Console and BigQuery are also in play. Joining ad data with Search Console, for example, lets the marketer spot keywords already ranking well organically but still drawing expensive bids — a use case the single-platform official server cannot cover.