📄 Your German customer's renewal contract carries an extra annex this year.
"List every AI system in the platform, give each one's AI Act risk tier, and warrant it's correct." Sign by month-end, or the agreement lapses.
😬 Your own sales deck calls it "AI-powered" — so does the GINOP grant application you won three years ago. Now a customer wants a signed warranty behind that word.
🔍 So you open the product to see what actually runs.
🔹 Most of it is the reporting layer you've built for a decade. It aggregates production data and turns red when a value crosses a customer-set limit. It never decides on its own.
🤖 Then there's the module you built for the production line. No thresholds: it learned from each machine's own history what normal looks like, and predicts when one is drifting toward failure.
💡 That second one is what the AI Act calls inference: the system works out how to reach its output instead of executing a pre-written instruction. The reporting layer doesn't — so it isn't an AI system. The predictive module is.
🖊 The warranty you can sign in good conscience is a narrow one. Go through the platform module by module and split what infers from what runs rules. The rule-based majority you can warrant as outside the Act. The one learning part, classify on its own merits — not on what a sales deck called it.