Exception handling and escalation
Detecting when a situation falls outside the AI system's confidence or authority, and automatically routing it to a qualified human with the right context — rather than guessing or failing silently.
Why it matters
No AI system handles every case. The difference between a trustworthy system and a dangerous one is what happens at the edges. Good exception handling means the AI knows what it doesn't know, escalates cleanly, and gives the human enough context to act quickly.
Where it shows up
finance
When an invoice doesn't match any known policy category or when a variance exceeds the AI's confidence threshold, the system flags it for human review with context rather than forcing a classification.
hr
When a manager's question touches protected characteristics, pending litigation, or situations the system hasn't been trained on, it immediately routes to HR with a summary of why escalation was triggered.
procurement
When a vendor request doesn't fit established categories, when a sole-source justification is needed, or when the request crosses departmental boundaries, the system routes to procurement leadership with context.
Common mistakes
- Setting escalation thresholds too high (system guesses when it should escalate)
- Setting escalation thresholds too low (everything gets escalated, defeating the purpose)
- Escalating without context — the human gets the case but not the AI's analysis
- Not having a defined person or team to receive escalations
Signals that a workflow needs this pattern
- The workflow handles edge cases that are hard to define in advance
- Incorrect handling of edge cases could cause legal, financial, or reputational harm
- The domain has well-defined 'normal' cases but poorly defined exception procedures
- Current exception handling is informal (ask someone, figure it out)