Bain published the definitive enterprise agentic architecture this week, and it settles the vocabulary. Three layers: orchestration, observability, and governed data access. Policy as runtime enforcement. Non-human identity as first-class infrastructure. Gartner’s parallel finding is blunt: 40% of agentic AI projects will be canceled by 2027, and the failing layer is infrastructure, not models. NatWest is the exemplar. Campaigns that took 60+ days now take one day, targeting 80-90% autonomous execution.
The question for any bank is no longer which models to choose. It is whether the orchestration and policy layer exists at all.
OSFI privately warned Canada’s Big Six banks about a specific cyber risk this week. A confidential April 29 alert to CTOs, CISOs, and CROs flagged that Anthropic’s Claude Mythos “significantly compresses the timeframe for effective risk mitigation.” The model finds and exploits vulnerabilities faster than banks can patch. The public bulletin stays technology-neutral. The private posture is model-specific. This is precedent-setting. Banks need model-specific threat assessment capability, not just category-level governance.
ISO 20022 arrives in November and it is the largest payment deadline of the year. Fedwire, CHIPS, and SWIFT CBPR+ all go ISO 20022-exclusive in the same month. No phased migration. All three converge at once. Structured addresses are mandatory; unstructured means rejection. The operational challenge is real, but this is also an AI data quality event: richer structured data enables better AI fraud detection.
The UK is legislating for AI agent payments and stablecoin integration. The FCA gains new Open Banking powers; the PSR has been merged into the FCA. Mills calls it “an arms race.”
FIS launched an Embedded Banking Platform that inverts the BaaS pattern: accounts stay on the bank’s own ledger, not a virtual one. The bank keeps customer ownership and regulatory control. M&T Bank is in the pilot, Q4 2026. This is the post-BaaS-consent-order answer to the dumb-pipe threat.
The IMF published a canonical three-layer payments framework (intent, authorization, settlement) with the core tension being probabilistic AI versus deterministic payment requirements.
The signal pattern is clear: architecture is converging. Bain, BCG, IMF, and BIAN are settling on the same three-layer vocabulary. The blueprint is done. The execution gap is the differentiator. Regulatory bifurcation is widening. The UK legislates proactively while OSFI reacts to model-specific risks. Canada’s E-23 in May 2027 is the hard deadline.