What is the AIS inventory expected by the AMF’s AI guideline?

What the guideline says, in its own words

The AMF’s guideline on the use of artificial intelligence — final since April 2026, in force May 1, 2027 — devotes a section to the “AIS inventory” (« inventaire des SIA »). The expectation is plain: the institution should keep a regular record of all its models and AIS, and AIS whose risk is determined to be non-negligible should be listed in a centralized inventory, each with a risk rating that is updated periodically.

The guideline’s other expectations orbit that inventory: institutional governance in which senior management designates an accountable person for all AIS, controls at every stage of the AIS lifecycle, and ongoing monitoring. It is a guideline — the language is supervisory expectation (“should”), not statute — but it is the text the regulator will use when it examines the institutions it oversees. The binding text is French.

Why doesn’t a spreadsheet hold up?

An inventory only has value in an examination if it reflects reality as things happen. Three difficulties come up constantly:

• The perimeter moves weekly. AIS arrive through vendors — a software update that switches on an agent, a copilot enabled by default — without passing through the internal process. A hand-kept inventory documents what you knew, not what is running.
• The risk rating has to live. A rating assigned once and never revisited contradicts the expectation of periodic updates and ongoing monitoring.
• The evidence has to be contemporaneous. An inventory reconstructed in the weeks before an examination shows one day’s state; it does not demonstrate diligence over time. That is the whole difference between documenting and demonstrating.

How does it read alongside OSFI E-23’s model inventory?

OSFI Guideline E-23 — effective the same day, May 1, 2027 — expects federally regulated financial institutions to maintain a comprehensive, accurate, evergreen model inventory under robust controls. The two texts do not merge: E-23 says “model,” the AMF says “AIS.” But one underlying base of information — which systems, what access, what rating, what history — can feed both inventories, each in its own regulator’s vocabulary.

That is precisely what a system of record is for: Agentica continuously maps every AI agent in your business environment and records it in a tamper-proof history, from which the inventory the AMF expects is an extract generated on demand — never a document to reconstruct. Explainability and bias-monitoring obligations remain the institution’s own; an honest inventory records that they exist, without standing in for them.