Motorola Solutions CTO Mahesh Saptharishi published an analysis in Forbes Tech Council arguing that AI agents deployed in emergency services need a capability he calls “escalation intelligence”: the ability to recognize when a situation exceeds the agent’s safe operating parameters and hand control back to a human operator without delay.
The Silent Failure Problem
The core risk, according to Saptharishi, is that AI failures in mission-critical environments look different from hardware failures. When a computer-aided dispatch system crashes, the failure is obvious. When an LLM hallucinates during a 911 call, it delivers a “plausible lie with the same confidence as a life-saving truth,” he wrote in the Forbes piece.
The pressure to automate is real. 911 call centers are operating with 25% vacancy rates, according to an IAED/NASNA staffing survey. Non-emergency calls consume dispatcher bandwidth that could go to active emergencies. AI agents can handle that intake, but the design constraints are different from enterprise deployments where the worst outcome is a bad email.
Four Design Pillars
Saptharishi outlines four requirements for AI agents in mission-critical environments, per his Forbes analysis:
Architectural accountability. AI agents handling non-emergency calls must detect linguistic cues of violence, suicidal ideation, or immediate danger and hand the interaction back to a dispatcher. The system optimizes for reliable human handoff, not call duration.
Contextual intelligence. General-purpose LLMs see accuracy decline on domain-specific tasks, according to Gartner research. Emergency AI needs to dynamically pivot as incidents unfold, surfacing critical information while filtering out noise.
Cognitive engagement. Saptharishi argues AI should be “mind-extending rather than mind-replacing,” citing behavioral scientist Dr. Harriet Rowthorn. Deliberate friction points prevent operators from deferring judgment to the system.
Continuous validation. Mission-critical AI requires ongoing evaluation against evolving threats, not one-time safety checks before deployment.
The Constraint on Agent Autonomy
The analysis presents a direct tension with the enterprise push toward fully autonomous agents. In emergency services, healthcare triage, and financial operations, the measure of success for an AI agent is not how many decisions it makes independently. It is whether the human operators working alongside it trust its judgment enough to act on its recommendations, and whether the system catches the moments when it should stop acting entirely. For teams building agents in regulated industries, escalation intelligence is becoming a design requirement that sits alongside runtime authorization and identity governance in the emerging agent infrastructure stack.