Three Regulators, One Month, the Same Six Demands
In July 2026, the Reserve Bank of India published draft AI governance guidelines for banks and regulated lenders. The comment period closes July 24. On its own, an RBI consultation might not make your reading list — but put it next to Canada's OSFI Guideline E-23 (model risk management, effective May 2027) and the EU AI Act's high-risk obligations (enforcement from August 2, 2026), and a pattern emerges that should concern every board, not just banks.
Three regulators, on three continents, with three very different legal traditions, have independently converged on the same core requirements for organizations that deploy AI in consequential decisions. When that happens, it stops being a compliance question for one jurisdiction and becomes the de facto global standard of care.
The Six Pillars of AI Model Risk
Strip away the jurisdiction-specific language and the three regimes demand the same six things:
| Requirement | RBI Draft (2026) | OSFI E-23 (2027) | EU AI Act |
|---|---|---|---|
| Board-approved AI/model governance | ✓ Explicit | ✓ Senior management accountability | ✓ Provider/deployer accountability (Art. 17) |
| Enterprise-wide model inventory | ✓ Explicit | ✓ With materiality tiering | ✓ High-risk system registration |
| Independent validation (incl. vendor models) | ✓ Explicit | ✓ Core requirement | ✓ Conformity assessment (Art. 43) |
| Human oversight of automated decisions | ✓ Explicit | ✓ Model risk controls | ✓ Article 14 |
| Controls for customer-facing generative AI | ✓ Explicit | ✓ In scope as models | ✓ Transparency (Art. 50) |
| Restrictions & decommissioning ("kill switch") | ✓ Explicit | ✓ Model lifecycle end-stage | ✓ Post-market monitoring & withdrawal |
The US already got there first: the Federal Reserve's SR 11-7 guidance has required board-owned model risk management, inventories, and independent validation since 2011. What's new is the explicit extension to generative and agentic AI — and the spread beyond banking supervision into horizontal law like the EU AI Act.
Testing Is Not Validation
In our assessment data, the weakest pillar by far is independent validation. Most organizations conflate two very different activities:
- Testing — the team that built the model checks that it works as intended.
- Validation — an independent function, with no stake in the model shipping, challenges its assumptions, data, performance, and limitations.
Regulators care about the second one. A model validated only by its own developers is, in supervisory terms, not validated at all. And the requirement explicitly extends to vendor-supplied models: "we bought it from a reputable vendor" is not a validation strategy under SR 11-7, E-23, or the RBI draft. If your organization runs a third-party credit scoring model, fraud engine, or LLM-powered decision tool and has never independently challenged it, that is a documented gap the day a supervisor asks.
The question a regulator will ask is not "did you test it?" — it's "who, independent of the build team, challenged this model, when, and where is the report?"
OSFI E-23: The North American Deadline Nobody Is Talking About
While attention focuses on Brussels, Canadian federally regulated financial institutions face a concrete domestic deadline: OSFI Guideline E-23 takes effect May 1, 2027. It expands model risk management well beyond traditional credit and market models to explicitly cover AI and machine learning — including generative AI. It demands an enterprise model inventory with materiality-based tiering, lifecycle controls from design through decommissioning, and independent review proportionate to model risk.
For banks, insurers, and trust companies, E-23 readiness work needs to start now — building an inventory, tiering it by materiality, and standing up an independent validation function takes quarters, not weeks. For everyone else, E-23 is the clearest preview of where AI regulation in North America is heading.
The Kill Switch Goes Mainstream
The provision drawing the most commentary in the RBI draft is the requirement for defined restrictions and decommissioning capability — the ability to constrain or shut down an AI system that misbehaves. For anyone deploying agentic AI, this is not theoretical. An autonomous agent with tool access and no tested kill switch is an incident report waiting for a timestamp.
A credible kill-switch capability means: per-agent and global shutdown mechanisms, a tested playbook for invoking them, and evidence that autonomy boundaries are enforced programmatically — not just documented in a policy. If you want to see what happens when those controls are missing, our Agent Behavior Simulation walks an executive audience through a realistic attack on their own agent configuration, step by step.
A 90-Day Readiness Plan
- Weeks 1–2: Baseline. Score your governance maturity against all six pillars. Identify which regimes apply to you (EU AI Act nexus? Canadian FRFI? Indian banking exposure?).
- Weeks 3–6: Inventory. Build or complete an enterprise AI/model inventory. Tier every entry by materiality and record its validation status — most organizations discover systems they didn't know existed.
- Weeks 7–10: Independent validation. Stand up an independent challenge process — internal second line or external — starting with your highest-materiality models, vendor models included.
- Weeks 11–13: Controls & kill switch. Verify human oversight points for automated decisions, transparency controls on customer-facing generative AI, and tested decommissioning/shutdown procedures for autonomous systems.
The convergence is the story: whether the letterhead says RBI, OSFI, or the European Commission, the expectations are the same. Organizations that treat these six pillars as a single global program — rather than three separate compliance projects — will spend less and be ready first.