What an Escalation Matrix Actually Defines
An AI agent vendor escalation matrix documents the severity tiers, vendor ownership, and response protocols that apply when an agent incident occurs. Rather than relying on generic SaaS outage categories, it is grounded in runtime signals such as policy violations, unauthorized tool calls, and permission events, so each incident is routed to the vendor accountable for the system that produced the triggering log.
Why Generic Vendor SLAs Fall Short
Foundation model providers, tool and MCP providers, and orchestration platforms each own a different part of the agent stack, and each is accountable for a different category of failure. A single SLA written for SaaS uptime does not distinguish between a reasoning failure, a permission failure, and a containment failure, so incidents can stall while vendors dispute which boundary the failure crossed.
Constructing Severity Tiers from Runtime Signals
Severity tiers should be built from the runtime evidence a system actually produces: the log source, the policy or permission scope involved, and whether the action was blocked, logged only, or executed. Mapping