What is an AI inventory?

What does a real AI inventory contain?

Four columns, for every entry: what (the agent, copilot, or integration — named and dated), who (the tool’s owner and the responsible person), what access (the data and systems its permissions cover), what activity (what it actually did, recorded as it happened). The fourth column is the one almost always missing — and the one that turns a list into evidence.

Two traps to avoid:

• An AI inventory is not a data inventory. Data mapping answers “where is our information?”; the AI inventory answers “who — which agent — touches it, and what is it doing?”. Data-mapping tools do not see agents; the two exercises complement each other without merging.
• A declared inventory goes stale. If the inventory only contains what teams declared, it documents intentions. Shadow AI — the vendor-enabled agent, the tool adopted without announcement — never adds itself: it has to be discovered through the traces it leaves.

Why do regulators ask for it first?

Because everything else depends on it. The AMF’s AI guideline expects an AIS inventory with risk ratings in a centralized register; OSFI’s E-23 expects a comprehensive, accurate, evergreen model inventory — both in force May 1, 2027. The logic is the same for a municipal council or an auditor: you cannot rate, monitor, or explain what you have not counted.

How do you keep it without staffing a full-time role?

By wiring it to traces rather than declarations: audit logs, permissions, connected applications — collected continuously and written into a history that does not fade. That is how Agentica works: continuous discovery and mapping of every AI agent in your business environment, read-only, metadata and signals only, recorded in a tamper-proof history from which the inventory — general, or in a given regulator’s vocabulary — is extracted on demand.