Govern agents like privileged staff, not like software features. Before an agent touches production, a board should require an inventory, a named owner, a distinct identity with least privilege, human approval for consequential actions, the ability to roll back, and an audit trail that shows every action and who authorised it. Report on those controls, not on adoption rates.
A chatbot gives a wrong answer; an agent takes a wrong action. It can change a configuration, send an email, move money or delete data, at machine speed and with whatever permissions it was given. The failure modes are new: goal hijacking through prompt injection, misuse of the tools it can call, privilege abuse, poisoned memory, and errors that cascade across systems before anyone notices.
The frameworks are converging on the same answer. Singapore's IMDA published a governance framework for agentic AI in January 2026 built on bounding risk upfront and keeping humans accountable. OWASP's Top 10 for agentic applications catalogues the attack surface, and NIST launched an AI agent standards initiative in February 2026.
| Control | What the board should see evidence of |
|---|---|
| Inventory and owner | Every agent listed, with one accountable business owner |
| Identity and least privilege | Each agent has its own identity and only the permissions its task needs |
| Autonomy tiers | Which actions run alone, which need human approval, which are forbidden |
| Rollback and kill switch | Tested ability to stop an agent and undo what it did |
| Audit trail | Every action logged with its trigger, its inputs and who approved it |
| Third-party agents | The same controls applied to agents inside vendor products |
I built an agentic tool for Microsoft 365 and Azure operations. The model took a weekend. The approval workflows, permission boundaries, rollback and audit trail took months. That ratio is normal, and it is the point: the controls are what let you approve the next ten use cases without re-litigating risk each time.
Board reporting, AI governance, DPDP readiness or CMMC. Tell me which and what the constraint is.
me@sarfarazchougule.com →