Who is actually accountable when an AI system makes a bad decision?

This is the question most executives ask too late. The vendor's terms of service make this very clear if you read them: the model provider is not liable for decisions made using their model. OpenAI, Anthropic, Google, all of them. The liability sits with the deployer. That is you.

Under the EU AI Act, the 'deployer' of a high-risk AI system bears specific obligations: human oversight, transparency to affected individuals, audit trails, and the ability to explain any automated decision. If you cannot meet those obligations, you are exposed. Not theoretically. Legally.

The practical protection is governance architecture. Every AI decision needs three things: a log of what data went in, what logic was applied, and what output was produced. A defined escalation path for when the AI's confidence is low. And a human who is named as accountable for that class of decision, with the authority to override.

Claude Guard provides the technical infrastructure for all three. But the organizational accountability, who owns the decision, who reviews the exceptions, who reports to the board, that is a leadership decision. SynthesisArc helps clients design both the technical and organizational governance in every engagement.