What triggers a human review of an automated AI decision?

Every AI system needs a clear answer to the question: when does the machine stop and the human start? If you cannot answer that before deployment, you are not ready to deploy. The answer is not 'when something goes wrong.' The answer is a defined set of triggers that route specific cases to human review automatically.

Confidence-based triggers: most AI models produce a confidence score alongside their output. Set a threshold. Below 85% confidence? Human review. The threshold depends on the stakes. For a customer service response, 70% might be fine. For a clinical decision, 95% might not be enough.

Impact-based triggers: any decision above a defined dollar amount, a defined risk level, or affecting a defined number of people goes to a human. A $50 refund can be automated. A $50,000 contract renegotiation cannot. Draw the line before the first transaction.

Regulatory triggers: certain decisions require human oversight by law. The EU AI Act mandates human oversight for high-risk AI systems. Know which of your workflows fall into that category and build the oversight in from the start. Claude Guard enforces these triggers at the architecture level so they cannot be accidentally bypassed.