In 2026 the AI tools arrived embedded in the systems banks already run, and the regulators rewrote model-risk guidance for the first time in fifteen years, then deliberately left generative and agentic AI out of it. The carve-out is not a reprieve. The real work now sits with the bank: govern what you already have and evaluate what you buy.

For three years, "AI in banking" mostly meant a slide in a strategy deck. In 2026 that changed. Agentic systems — software that does not just answer questions but takes actions, moves across systems, and executes multi-step work with limited human intervention — crossed from demonstration to deployment. And for most banks they did not arrive through an innovation lab. They arrived through the vendors the bank already runs. That single fact reframes the entire conversation, and it is the fact I keep coming back to in conversations with bank executives.

The Tools Are Already Inside the Building

The core providers have made AI a platform feature rather than a project. In the span of one spring, Fiserv launched agentOS, an operating system for agentic AI in banking built with OpenAI and AWS; FIS brought an agentic financial-crimes agent to market with Anthropic, with institutions like Amalgamated Bank among the first to put it to work; and Oracle Financial Services unveiled an agentic platform with hundreds of prebuilt retail and corporate banking agents promised within twelve months and then extended it to corporate banking in April. The early use cases cluster where the work is high-volume, procedural, and auditable — financial-crime detection, regulatory-change triage, controls monitoring, onboarding, and loan analysis. And these are not demos. First Interstate Bank is piloting a commercial loan onboarding agent that moves data across systems to cut cycle times; Boulder Dam Credit Union cut report generation from ten minutes to seconds; Salem Five, City National, Bank OZK, and SouthState came online this summer.

There is a second arrival happening underneath the agents, and it is getting far less attention: the connector layer. Model Context Protocol, or MCP, is fast becoming the standard plumbing that lets an agent reach into core systems, document stores, and third-party data — which means an agent's real capability is defined by what it is connected to, not what it was built for. Every connector is an access grant. That is why the NSA issued security guidance on MCP in June, and why an unpatched MCP-related flaw was flagged as a banking-sector risk the same month. A bank that inventories its agents but not its connectors has inventoried half the exposure.

Adoption is real but uneven. A Temenos-commissioned survey of more than 400 banks found a majority with generative AI live or in development, but only 11 percent actually implemented, and the share drops to roughly 40 percent for institutions under $10 billion in assets. The agentic numbers tell the same story from the other direction. Accenture found 57 percent of banking executives expect AI agents to be fully embedded in risk, compliance, and fraud functions within three years, while Deloitte puts the share of organizations with agentic AI actually in production at just 11 percent. The distance between those two numbers is where both advantage and risk are being built.

For community and regional institutions, the more important fact is quieter: the AI footprint is arriving embedded in third-party platforms. With three core providers serving more than 70 percent of U.S. banks, the provider's AI roadmap is becoming the bank's AI strategy by default — unless the bank decides otherwise. That is a governance and evaluation question, not a technology one.

The Rulebook Changed by Leaving AI Out

On April 17, 2026, the Federal Reserve, OCC, and FDIC jointly issued revised model-risk guidance — SR 26-2, OCC Bulletin 2026-13, and FDIC FIL-15-2026 — replacing SR 11-7, the standard that had governed model risk for fifteen years. The new guidance is principles-based and risk-based, aimed most directly at banks above $30 billion in assets, a dramatic lift from the old $1 billion threshold, but scaling by actual model-risk exposure rather than asset size alone.

The headline most institutions will take away is that generative and agentic AI sit explicitly outside its scope, described as "novel and rapidly evolving," with a request for information on banks' use of AI still to come. Read that boundary carefully, because it is the part most will misread. The carve-out is not a reprieve. Traditional statistical and machine-learning models remain fully in scope. The guidance reaffirms that the bank using a vendor's model — not the vendor — owns the obligation of understanding, validating, monitoring, and documenting it, even when the vendor won't open the box. And while the agencies took pains to say that noncompliance with guidance will not itself draw supervisory criticism, they kept every bit of their authority to act on unsafe or unsound practices, prescriptive AI standard or no.

The regulators did not lower the bar. They moved it — off the checklist and onto the bank's judgment. Governance now has to be built on principle rather than pointed at a rule; the burden was not reduced so much as relocated. Add the EU AI Act's high-risk obligations reaching credit scoring and other financial systems this August, and the direction is unmistakable: expectations are rising faster than rules.

This is the gap the industry keeps naming, and it shows up in the budget. By Deloitte's count, 93 percent of AI-related spending goes to technology and only 7 percent to the people, training, change management, and governance that make it safe to use. EY finds 60 to 70 percent of banks cannot replicate AI success beyond isolated programs. That is backwards, and it is precisely where the risk compounds.

Evaluation Becomes the Discipline That Matters

If the tools arrive embedded and the rulebook is deliberately silent, the decisive control moves earlier — to how a bank evaluates what it adopts. And evaluation itself has changed. When every vendor response says "AI," the task is no longer checking a feature box; it is separating real capability from a good demonstration. The object of evaluation has shifted from the model to the whole operating system around it: the orchestration layer, integrations, audit trail, data rights, and the model-risk posture of the AI embedded inside the product. Agents that dazzle in a sales demo routinely break in an audit.

That is also why the buy-versus-build calculus is unsettled. Generic AI teams stumble not on the technology but on the regulatory, integration, and model-risk realities of a bank — buying looks faster right up until an examiner asks who validated the vendor's model. Real evaluation in 2026 is vendor-agnostic and conflict-free. It prices the true total cost and model risk of AI features, and it asks a blunt question before anyone signs: what, exactly, will this institution be governing once the ink is dry? Rigor is what separates transformation from pilot-chasing.

Your Assignment

None of this is a technology story. It never was. AI in banking is an operating-model change, not a technology purchase. It is a redesign of how an institution decides, controls, and operates, with governance as the foundation the rest of it stands on. The regulators just made that explicit in the most understated way possible — by handing the pen back to you, the bank.

The rulebook's blank pages are not a loophole to exploit. They are the assignment you must complete. The institutions that treat governance and disciplined evaluation as competitive advantages — and not compliance chores — are the ones that will reach production fastest, and safely. Everyone else will be improvising.

In my next article, releasing Tuesday July 21, I look at the obvious next question — if a bank does not pick up the pen, what actually protects it, and what does standing still really cost?

CCG Catalyst advises community and regional banks, credit unions, and fintech companies on AI governance, core and payments modernization, and vendor evaluation. If your institution is weighing what embedded AI and the revised model-risk guidance mean for its governance and its vendor decisions, reach out to our team at www.ccgcatalyst.com, or see the full library at CCG Insights.

See our latest announcement: CCG Catalyst's Paul Schaus Named a 2026 Top Consultant by Consulting Magazine

By: Paul Schaus | Founder & Managing Partner, CCG Catalyst Consulting

Disclaimer: The views expressed in this article represent the perspective of CCG Catalyst Consulting based on our direct experience advising financial institutions. This commentary is intended to stimulate industry discussion and does not constitute legal, accounting, or regulatory advice.