liveBook · Manning

Monitoring

As explained in AI Governance

Observation of a system or process to track its status, behavior, or performance.

A routine customer inquiry can become a security incident if an AI agent unexpectedly calls extra tools or accesses an unrelated database. Monitoring tracks system behavior during operation so teams can detect such deviations, investigate them, and adapt controls as the system changes. section 4.8.7 section 1.3

For generative AI, monitoring supplements point-in-time audits and stress testing. It supports probabilistic risk management: model behavior can change, so policies and controls must evolve rather than remain fixed after deployment. section 1.3

Build monitoring into the operational workflow:

  • Collect usage patterns, token consumption, and user feedback; connect these metrics to dashboards and alerts. section 2.3.4
  • Cross-check input and output call logs against vendor logs. Track system-prompt updates and vendor configuration changes so investigators can account for changes across the integration. section 2.3.4
  • Capture agent-specific signals, including unexpected tool calls, unusual reasoning patterns, and objectives that shift across sessions. Reasoning traces can help reconstruct incidents. section 4.8.7
  • Define compliance events for data access, external calls, autonomous decisions affecting individuals, disallowed data access, errors, and human overrides. Use the audit trail to review cases and track regressions. section 5.6.4

Investigating an agent deviation

A claims-status inquiry normally triggers one tool call, but an agent makes three or attempts to access the adjuster assignment database during a routine question. Investigate the tool calls and captured reasoning traces. While investigating, operators can narrow the agent’s allowed actions, remove tools, or require more human approvals; a kill switch provides a way to halt and isolate it immediately. section 4.8.7

Monitoring also helps detect compliance regressions after updates. A sudden rise in attempts to use disallowed data sources is a signal to investigate, not merely a performance metric. section 5.6.4

How AI Governance covers it

How deep it goes
Specifies telemetry, review, investigation, and containment practices without providing a full implementation.
Examples
Uses a claims-status agent making unexpected tool calls and an update that increases disallowed data-source attempts.
What you can do after
Operate/configure monitoring by selecting telemetry, connecting alerts, reviewing traces, and tightening agent restrictions during investigation.
How it connects
Connects observability with governance, audit trails, privacy compliance, and incident containment.

Also explained in

Compare how 51 books explain monitoring →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.