Answer — AI Governance
What is agentic risk?
Agentic risk is the risk created by AI agents that take actions — accessing data, modifying files, sending communications — faster and more broadly than humans can track. It is managed with a living inventory and a record of what each agent actually did.
How is agentic risk different from “AI risk” in general?
A model that errs produces a wrong answer. An agent that errs takes a wrong action: it reads a folder nobody thought about, modifies a file, sends an email, triggers a transaction. The difference is one of kind, not degree — the error does not stay on a screen; it enters your systems and your business history.
Three properties make this risk hard to track by hand:
- Speed. An agent chains together in seconds what an employee would spread over days.
- Reach. An agent connected to a work environment potentially touches everything its permissions cover — often far more than anyone pictured when switching it on.
- Origin. Many agents arrive through vendors — an update that enables a feature, a copilot deployed by default — without passing through any internal approval process.
Why is this becoming a regulatory question?
Because the applicable frameworks look precisely at what systems do, not just what they are. Law 25 requires transparency when decisions are based exclusively on automated processing — which presupposes knowing which systems make such decisions. The AMF’s AI guideline expects an AIS inventory with risk ratings and ongoing monitoring — and an agent enabled without your knowledge is, by definition, outside the inventory.
How is agentic risk managed?
With two complementary instruments: a living inventory — which agents, what access, what activity — and a tamper-proof history of what each agent actually did, recorded as it happens. The first lets you act; the second lets you demonstrate, months later, that you knew and that you acted. That is what Agentica produces: continuous mapping of every AI agent in your business environment, recorded in a tamper-proof history under independent custody — read-only, metadata and signals only.