Customer data glossary

Kill switch (AI agents)

A kill switch is the tested way a named owner stops an AI agent rail from outside the model, so that new work stops and the job returns to a stated manual path.

By Art Kukharenko · Published 2026-11-02 · Updated 2026-11-02

All glossary terms

Part of the Learn hub and the customer data glossary.

What it means in practice

A kill switch answers four questions on the rail card: who may press it, what stops, what happens to work in progress and what keeps running. The control sits in a gateway or in the Orchestrator, never in a prompt or among the agent's own tools.

A good switch removes the action path and keeps the safety net. Disabling a pause rule, for example, leaves alerts flowing to a person. Stopping an agent should not stop the rule-based checks around it, such as consent enforcement.

Some stops should fire without a person. A rail can pause on repeated rejections or a breached limit. Amazon SES shows the idea outside AI: sending can pause automatically, per configuration set or account, using CloudWatch alarms 1.

For high-risk AI systems, EU AI Act Article 14(4)(e) expects overseers to be able to interrupt a system "through a 'stop' button or a similar procedure" 2. Annex III obligations apply from 2 December 2027. Marketing segmentation is not high-risk in itself, so take advice on your own jobs. The design is sound either way.

A stop nobody has tried may fail when needed, so drill it and time it. The roadmap's checklist suggests a drill in the last quarter. PrivyCDP's audit ledger records each AI-assisted action and the actor that authorised it, so a drill's review has a record to read.

Example (health insurance, illustrative): a member-communications team sees false pauses from the Anomaly Watcher. The owner disables its pause rule. Alerts continue to a person, sends continue under normal controls, and the change goes on record.

Rail card: the Owner and kill switch field.
Blast radius: what a stop contains.
Orchestrator agent: broadcasts a pause.
Approval gate: stops work before it runs.

Where it matters

Agent rails, guardrails, policies and prompts: what actually enforces a stop.
Marketing anomaly detection: an automatic pause and a person who resumes.

Back to the customer data glossary or the Learn hub.

Browse use cases Function playbooks Agent roster

FAQ

Who should be able to press the kill switch?

The rail's named owner, plus an on-call role for out-of-hours incidents. The agent must never be able to press it, reverse it or remove it.

What happens to work in progress?

The rail card should say. Typically nothing new starts, runs waiting at a gate expire, and a person picks up the queue by hand.

Sources

This entry is general information, not legal advice.