Exactly. The label test misses the risk. If a system infers from inputs and routes outcomes for people, the governance question starts there — regardless of whether the organization calls it “AI.”
This definition question seems like the real front door to AI Act compliance.
A lot of organizations will undercount AI systems because they see them as ordinary software, embedded tools, copilots, plugins, or workflow features. But once a system moves from data input → inference → recommendation/content/decision → real-world or institutional effect, the governance question changes.
The next hard step is operational: not only “is this an AI system?” but “what controls govern its outputs when they influence high-stakes action?”
The 'front door' framing is great. I think companies don't get through it because they're applying the wrong test. They ask “did we build something we call AI” rather than “does this system infer from inputs to produce outputs that affect decisions about people”. The Article 3(1) definition follows the function, not the label. A workflow tool that routes cases based on pattern recognition is an AI system. I can imagine companies that have several of those and haven't counted them.
Exactly. The label test misses the risk. If a system infers from inputs and routes outcomes for people, the governance question starts there — regardless of whether the organization calls it “AI.”
This definition question seems like the real front door to AI Act compliance.
A lot of organizations will undercount AI systems because they see them as ordinary software, embedded tools, copilots, plugins, or workflow features. But once a system moves from data input → inference → recommendation/content/decision → real-world or institutional effect, the governance question changes.
The next hard step is operational: not only “is this an AI system?” but “what controls govern its outputs when they influence high-stakes action?”
The 'front door' framing is great. I think companies don't get through it because they're applying the wrong test. They ask “did we build something we call AI” rather than “does this system infer from inputs to produce outputs that affect decisions about people”. The Article 3(1) definition follows the function, not the label. A workflow tool that routes cases based on pattern recognition is an AI system. I can imagine companies that have several of those and haven't counted them.