Four regimes in the public inventory now let us say this without waving at unpublished rows: supervisors are converging on how model risk is managed and diverging on where the perimeter is drawn. That is a better briefing than “AI regulation is fragmented.”
Mechanism: inventory, challenge, lifecycle
Read OSFI Guideline E-23 (effective 1 May 2027), AMF (Québec)’s Model Risk Management Guideline (issued June 2025, no commencement date in the instrument), SR 26-2 (17 April 2026) and Korea’s Framework Act with its Enforcement Decree (co-commencing 22 January 2026) side by side and the operating shape is familiar:
- A model inventory, with materiality or risk rating deciding the intensity of the rest
- Independent review or effective challenge, not self-certification
- Lifecycle evidence: development, use, monitoring, exit
- Vendor and third-party models inside the same perimeter, which is why OSFI B-10 has to sit next to E-23. Fourth-party feeder models reached through third-party review have nowhere to live without it.
Three inventories, three objects. Do not collapse them:
| Instrument | Object | Fields |
|---|---|---|
| AMF (Québec) AI Guideline | centralized inventory of AISs | Appendix 2 |
| AMF (Québec) TPRM §10 | inventory of arrangements plus a centralized register for critical ones | §10 bullet list |
| OSFI E-23 §C.1 | model inventory (non-negligible inherent risk only) | Appendix 1 |
AMF (Québec)’s AIS-inventory duty is not a section of the MRM Guideline. It is in the Guideline for the Use of Artificial Intelligence. The instrument’s word is “inventory,” not directory. Footnote 13 points at Appendix 2 for the field list. Footnote 14 cross-refers to the MRM Guideline in a different context (model risk rating precedence). The MRM Guideline still wants a centralised inventory of important models. Two instruments, one inventory problem, a third object under TPRM.
The English AI PDF carries a March 2026 footer. That is a version stamp, not Issued. Issued is first publication in any official language (April 2026). Issued (EN) is 27 April 2026. Do not gold-tick 7 April 2026 from a bulletin. The instrument takes effect on 1 May 2027, the same day as E-23. That April 1 2027 date circulating against E-23 is a law-firm reading. The OSFI instrument page says 1 May 2027. First-party governs. The lesson is not that the bulletin is generally unreliable. It is that a date may not carry a gold tick from a secondary source.
Boundary: what is a model and who is in
The definitions do not agree.
- AMF (Québec) MRM §2.1 is the broadest of the three financial-services texts: “a formalized representation of a concept, theory, process or system”, which may include sub-models, with no output-type limit. §1 keeps reputational damage inside model risk. §3.4 is model validation and internal audit, directed inward: internal audit “should assess the adequacy of processes and controls” with respect to “elements such as” the listed heads. Material deficiencies go to relevant stakeholders, particularly risk management. §4.2 gives the validation team and internal audit “an opportunity, on a periodic basis or as required, to present their observations to the board of directors.” §4.4 is a separate limb: the institution “must submit to the AMF, at the frequency determined by the AMF or when requested, a report by the auditor” covering audit scope on models and the effectiveness of model controls. That is not the S-20 financial-statement opinion. Treat any 1 May or 1 April on an AMF (Québec) row as unverified until it is traced to a named instrument. Search lautorite.qc.ca, not amf-france.org.
- OSFI E-23 covers all models at all FRFIs, including foreign bank and foreign insurance branches per Guideline E-4. The 2017 guideline was deposit-taking institutions only. FRPPs were in the draft and dropped. Foreign insurance branches are in §A.2 even when metadata omits them. Working rule: read the guideline, not the letter. Footnote 1 adopts the OECD definition of an AI system for the purposes of the guideline. Vendor and third-party models route to B-10.
- SR 26-2 uses a tighter quantitative-estimate definition, then carves generative and agentic models out in footnote 3 while telling firms their existing risk-management practices should still govern tools not covered. The $30bn screen generally excludes community banks unless their activities sit outside traditional community banking.
- Korea does not use a model definition at all. It uses a closed list of high-impact AI domains. Do not use the EU risk-tier label here. That collides with MSIT’s own English and with Korea’s closed ten-domain list. Sectoral financial regulation can displace the Framework Act: Article 5 cuts both ways, other statutes’ special provisions prevail and new or amended statutes must conform to this Act’s purpose.
Explainability is the cross-jurisdiction hook that actually operationalises. Korea’s Act Article 3(2) gives impacted persons a right to explanation “to the extent technically and reasonably possible.” Article 34(1)2 and Decree Article 28’s seven assessment heads turn that into an assessment. Pair it with E-23 explainability and AMF (Québec) §2.1, not with a slogan.
Reputation risk is the cleanest divergence
On 7 April 2026 the OCC and FDIC issued a final rule, Prohibition on the Use of Reputation Risk by Regulators (OCC 2026-12; FIL-13-2026), effective 9 June 2026. The Fed did not join. Reputation risk is prohibited as a basis for US supervision by those two agencies. It is retained in Canadian model-risk definitions: E-23 §A.4 (“operational, or reputational consequences”) and AMF (Québec) MRM §1 (“could damage a financial institution’s reputation”). Do not pair the US rule with OSFI’s housekeeping of a 2005 reputation-risk discussion paper. Different mechanism: a constraint on supervisors versus retirement of an outdated paper.
Korea’s enforcement map, because the press framing is wrong
Three tracks, all with a ceiling of KRW 30 million. Never a turnover percentage.
| Track | Provisions | Route | Ceiling |
|---|---|---|---|
| Direct fine | Art. 31(1) prior notice. Art. 36(1) domestic agent | Art. 43(1) 1-2 | KRW 30m |
| Investigate, then order, then fine | Art. 31(2) generative labelling. Art. 31(3) synthetic/deepfake. Art. 32(1)-(2) safety. Art. 34(1) high-impact duties | Art. 40(1) then Art. 40(3) cease/correct order. Art. 43(1)3 only if the order is ignored | KRW 30m |
| Unenforceable | Art. 30(3) pre-certification. Art. 35(1) impact assessment | “shall endeavour.” Lever is procurement preference: Arts. 30(4), 35(2) | none |
Two publishable consequences. Deepfake labelling carries no first-instance penalty in Korea. The only criminal provision, Article 42, binds Committee members who breach Article 7(9) confidentiality, not operators. Anyone citing criminal liability for a business under the Korean Act has misread it.
The Decree is where the Act becomes operable. Co-commencement 22 January 2026, with Appendix 1’s digital-medical-device limb on 24 January 2026 matching Act Article 2(4)(d). Presidential Decree No. 36053 was amended by Presidential Decree No. 36506 on 20 July 2026 (in force 21 July 2026) to operationalise Act No. 21311. Article 24 sets a single cumulative-compute threshold for the safety duty. Article 29’s domestic-agent tests are disjunctive: KRW 1tn revenue or KRW 10bn AI-service revenue or 1 million average daily domestic users over the preceding three months at year-end or having been fined under Act Article 43(1)3. The fourth is a ratchet. Ignore a corrective order once and the agent duty attaches permanently. Article 2 implements Act Article 4(2): three designating authorities over enumerated functions, not a blanket carve-out. Article 7(13) sunsets the National Artificial Intelligence Committee five years after commencement, not the Act. Act No. 21311, on law.go.kr, later renamed that body. A BSA letter dated 15 October 2025 already said “National AI Strategic Committee (NAISC)”, so the rename was in administrative use before passage. Official English (Strategy vs Strategic) is unsettled. Pin MSIT’s own usage, not trackers. The 21311 limb is AI research institutes, not AI safety institutes. The Korea AI Safety Institute was already grounded by Art. 12 of the original Act. Act No. 21311 also adds a statutory “training data” definition, a legal basis for those research institutes and stronger protections for AI-vulnerable groups in impact assessments. It renames the National AI Committee to the National AI Strategy Committee (국가인공지능전략위원회). A BSA letter dated 15 October 2025 already said “National AI Strategic Committee (NAISC)”, so the rename was in administrative use before passage. Official English (Strategy vs Strategic) is unsettled. Pin MSIT’s own usage, not trackers. KLRI English is an unofficial translation and still carries the defects we do not propagate. Identifiers 21311 and 36506 are first-party on law.go.kr. English names and any MSIT one-year grace before administrative fines remain secondary.
What a programme should do with this
Encode the shared mechanism once: inventory, materiality, challenge, lifecycle, vendor reach-through. Then price the differentials. US programmes cannot use reputation risk as a supervisory residual and cannot pretend footnote 3 is a free pass. Canadian programmes keep reputational consequences inside model risk and have to reconcile OSFI enterprise-wide scope with AMF (Québec)’s broader definition plus a separate AI guideline. Korean programmes map to an enumerated high-impact list and a two-track enforcement map that no other jurisdiction in the pack has.
Internal audit diverges on placement versus findings. SR 26-2 §VI addresses where internal audit sits. AMF (Québec) MRM §3.4.3 addresses what audit does with what it finds: inform relevant stakeholders, particularly those responsible for risk management. §4.4 is a separate AMF-facing auditor report on model-audit scope and control effectiveness, at the frequency the AMF sets or on request. That report is not the S-20 financial-statement opinion.
Draft-to-final on the English MRM PDF (consultation December 2024 versus final June 2025) softens second-line and third-line allocation in the same section:
| Draft | Final |
|---|---|
| at least once a year | on a periodic basis or as required (§4.2) |
| internal audit should ensure | internal audit should assess (§3.4.2) |
| with respect to: | with respect to elements such as: (§3.4.2) |
| may, at its discretion, perform certain technical validations | not in the English final |
The last row is O-28. It may be unanswerable from the English final alone. The clause is absent. Recovering whether it was dropped or never made the English text needs the French draft PDF.
Section numbers also move between draft and final. A citation taken from a draft is suspect until it is re-pinned. The English final does contain §4.4. A bare §6.1 is the AI guideline, not MRM.
AMF (Québec) third-party pointer that goes stale:
| Date | Rows | Transition |
|---|---|---|
| 2027-04-01 | ca-amf-tprm, ca-amf-mrm | Outsourcing Guideline (April 2009) superseded. MRM footnote 12 goes stale. |
Do not fold 1 April 2027 into the AI guideline’s 1 May 2027. They are different instruments. Do not fold 1 April 2027 into E-23 either. Both E-23 and the AMF (Québec) AI Guideline commence on 1 May 2027.
Canada’s two regulators have converged, deliberately and asymmetrically
OSFI E-23, issued 11 September 2025 and effective 1 May 2027, folds AI into the definition of “model,” adopts the OECD definition of an AI system in footnote 1 and requires an enterprise model inventory of every model whose inherent risk is non-negligible. It routes vendor and third-party models to B-10.
The AMF (Québec) Guideline for the Use of AI takes effect on the same day, 1 May 2027. Annex 1 states that the expectations are based on the OECD’s principles for trustworthy AI. Footnote 2 cites the OECD Council Recommendation. Footnote 1 cites the IAIS 2024 application paper on the supervision of AI. The guideline requires a centralized inventory of AI systems. The AMF files it on its own site under /guidelines/use-of-models/, beside MRM. Both regulators locate AI inside the model-risk taxonomy rather than beside it.
The AMF’s response-to-comments records that commenters sought “alignment with OSFI’s expectations, particularly with regard to artificial intelligence systems (AISs) within its Guideline E-23, Model Risk Management (2027).” That is a first-party admission of alignment intent, naming E-23. AMF TPRM §1 draws on BCBS, IAIS and FSB.
The stagger is deliberate. Quebec’s Third-Party Risk Management Guideline commences a month earlier, 1 April 2027, revoking a 2009 outsourcing guideline and reproducing an architecture OSFI already built federally. B-10 has been in force since 1 May 2024 (OSFI consultation-response). On models and AI, Quebec has caught up to the day. On third-party risk, it is still closing a three-year gap.
Both same-day commencement limbs and the shared OECD anchor are first-party on both sides. The French-language publication day for the AMF AI and TPRM pair is still a month-precision Issued field. The English day is in the note.
See the Jurisdiction Packs for the rows. See What SR 26-2 actually changed and OSFI E-23 for insurers for the single-market notes this comparison sits on.