Microsoft Azure Moves 45% of AI Incidents Off LLM Reasoning, Cuts Per-Incident Cost 70%

Azure Networking engineer Arun Malik detailed a four-layer framework that moves an AI incident-response agent from read-only assistant to production actor. The approach removed 45% of incidents from LLM reasoning and cut per-incident cost by 70%.

• 45% of AI incidents no longer require LLM reasoning
• Per-incident cost dropped 70%
• The model only parses telemetry; deterministic code decides
• Any production mutation requires human authorization