Governing Northwold Assurance Group's AI agents: every tool call an agent proposes, such as a refund prompted by untrusted text in a customer document, passes a policy engine outside the model, a human approval above threshold and a tool that holds only a per-task token; nine agents each cut their unapproved high-consequence action types to zero with almost every call policy-checked; and unapproved consequential actions fall from 1,840 a week to 4, inside the Board's appetite of 10, after 37 refunds worth £318,740 went to the wrong accounts.

CY-PRO-038Pro Level · LLM Security

LLM Tool Use Governance

Govern what AI agents can do — tools, credentials, approvals, policy, trust boundaries, monitoring, assurance and incident response — and defend it to a Board.

40 hours12 chapters60 lessons433 activities12 labs

What you'll learn

Measure agent risk by what agents can do and did — tool grants, high-consequence capability, unapproved consequential actions — rather than by claims about model guardrails.
Design and evidence controls outside the model: least-privilege tools, scoped credentials, human approvals, policy as code, parameter binding and trust boundaries.
Build metadata-only monitoring, assure controls with an approved benign test library and respond to agent misuse within minutes.
Cost governance options, set a measurable risk appetite with key risk indicators, name residual risks and defend a single resolution to a Board.

Price:৳15,000