AI agents are entering banks. The rulebook got there first.
51% of banks were piloting AI agents by mid-2026. The OCC has published a four-tier examiner framework, and the EU AI Act's high-risk deadline is 2 December 2027.
- 52% of financial services firms surveyed are in active adoption of agentic AI, but only 23% have reached the scaling or transforming stages. Self-reported m:…7898
- OCC Bulletin 2026-54 was issued on November 4, 2026, and the framework uses four tiers of risk, from Tier 1 (Critical) to Tier 4 (Low). m:…3717
- The EU AI Act's full obligations for high-risk AI systems come into force on 2 December 2027. m:…3969
- Plaid's Agent Detection identified 2.3x more agent sessions than competitors in side-by-side testing, with no increase in false positives compared to competitors. Self-reported m:…1907
The adoption gap
Banks are adopting agentic AI faster than they are scaling it. 52% of financial services firms surveyed are in active adoption of agentic AI, according to Mambu, but only 23% of financial services firms have reached the scaling or transforming stages of AI adoption. Fintechs lead traditional institutions in advanced AI adoption at 47% compared to 30%, and 56% of fintechs report increased profitability from AI compared to 34% of traditional institutions. m:…7898
The gap has a cause. Data fragmentation and quality are the leading obstacles to AI adoption in financial services, Mambu reports. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027 due to costs, unclear value, and inadequate risk controls. KPMG found that 51% of banks were piloting AI agents by mid-2026, which means most of the industry is still in the pilot phase while the cancellation clock runs. m:…7898 m:…3969
The US examiner framework
The US regulator has now written down how it will judge all of this. OCC Bulletin 2026-54 was issued on November 4, 2026, and it is a resource for examiners that does not establish supervisory expectations for banks. The framework uses four tiers of risk, from Tier 1 (Critical) to Tier 4 (Low), and risk classification in the framework is based on potential consequences, transfer of agency, and human oversight. m:…3717
The framework is deliberately light-touch, and deliberately broad. Banks are not expected to use the framework to assess their own use of AI or their AI risk management, and the OCC states that banks should not treat AI as a separate and distinct risk category. Examiners may use the framework to supplement existing model risk management, governance, technology, compliance, and operational-risk practices, and examiners are instructed to tailor examinations to a bank's size, complexity, and risk profile. The framework applies to national banks, federal savings associations, and federal branches and agencies, and the OCC intends to update the framework as technology, use cases, and supervisory understanding change. m:…3717
The OCC even renamed the thing it is regulating: the OCC uses the term 'Super Intelligence' (SI) to refer to technologies previously covered by the term artificial intelligence. The language is new. The supervisory approach is not: fold AI into the existing risk framework and judge it by consequences. m:…3717
The European deadline
Europe is less gentle. The EU AI Act's full obligations for high-risk AI systems come into force on 2 December 2027, and AI systems used to evaluate creditworthiness of natural persons are classified as high-risk under the EU AI Act. AI systems used for risk assessment and pricing in life and health insurance are classified as high-risk too. High-risk AI requirements include risk management, data governance, technical documentation, logging, human oversight, and quality management. m:…3969
The human-oversight bar is explicit. The EU AI Act requires that humans can understand, monitor, and override AI decisions in real time, and Article 14 of the EU AI Act requires high-risk AI systems to be designed for human oversight. Article 27 of the EU AI Act requires a fundamental rights impact assessment for high-risk systems before first use. EIOPA's August 2025 Opinion applies governance and oversight principles to insurers' AI systems outside the high-risk tier. m:…3969
The cost side is easy to underestimate. IDC predicts Global 1000 companies will underestimate AI infrastructure costs by 30% through 2027. Taktile advises teams to treat AI systems as products with dedicated budgets and cross-functional teams rather than one-off projects. The firm's own benchmark suggests the work is worth doing carefully: Taktile Labs' FinSpread-Bench AI agents achieve 96.5% accuracy in spreading financial statements compared to 89% for trained analysts. m:…3969
The tooling
The vendors are selling visibility first. Plaid has introduced Agent Detection to provide financial institutions with visibility into agent-driven interactions, and Agent Detection identifies whether an interaction is human or agent-driven and identifies the specific AI provider. Plaid's Agent Detection identified 2.3x more agent sessions than competitors in side-by-side testing, and testing showed no increase in false positives compared to competitors. Agent Detection is designed to help financial institutions shape how they support agent-initiated payments, and it is available today via a lightweight SDK. m:…1907
Plaid is also working on the permission layer: Plaid is developing a secure way for consumers to approve specific actions for AI agents, and Plaid already runs device, behavior, and network intelligence across hundreds of institutions. The pattern across the industry is the same. Tipalti reports that 46% of finance teams report broader AI implementation in AP, but only 6% of finance teams say their AP operations are fully automated with agentic AI. Mambu's advice is to start small: banks should start with read-only agent access to build trust and demonstrate value before automating actions, and successful agentic AI deployment requires a top-down governance mindset rather than just an engineering solution. m:…1907 m:…1457 m:…7898
One technical distinction matters for finance teams. MCP is an open standard that gives AI applications a consistent way to connect to external tools and data, according to Tipalti, but MCP does not reason, make decisions, or approve or execute finance workflows, and MCP does not guarantee security, compliance, or correct AI decisions on its own. Tipalti MCP is read-only at launch, and Tipalti's Finance AI agents operate inside the platform to support AP automation workflows. The plumbing is open. The accountability is not automatic. m:…1457
What is still unresolved
- Whether the OCC's examiner framework becomes a de facto standard even though it sets no supervisory expectations.
- Whether banks meet the EU AI Act's 2 December 2027 deadline for high-risk systems.
- Whether Gartner's prediction holds that over 40% of agentic AI projects will be cancelled by the end of 2027.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.