DESIGN NOTE · YOUNGJUN PARK
Connecting AI alerts to action
A design note on AI-assisted operations. Benefits require validation against a baseline; they are not assumed here.
Start with the decision
Define who must act, what action is possible, and when the opportunity expires before choosing a model. An alert with no owner or action creates more monitoring work.
Show evidence and uncertainty
Include supporting events, freshness, missing inputs, and the reason for the priority. Keep generated explanations separate from authoritative system facts.
Keep consequential actions reviewable
Use explicit human review for consequential operational actions. Log the proposed action, approval, override, and outcome so decisions can be examined.
Measure useful action
Compare false positives, response time, missed exceptions, and operator workload against the existing process. Model output alone is not evidence of operational improvement.