For bank technology leaders the conversation about AI governance has moved beyond the question of whether staff should use generative AI. The more immediate questions now are where it is already being used, what information it can access, how to evaluate new capabilities from existing vendors, and when an AI system should be allowed to take action rather than simply make a suggestion.
To help reduce some of the ambiguity around these topics and separate theory versus practice, National Leader Forum brought together a group of senior technology and security executives from mid-sized banks around the United States for a virtual knowledge exchange and invited AI governance subject matter expert Gerald Kierce to moderate the discussion.
All of the participating banks were at different stages of adoption but their practical concerns overlapped and the discussion offers a useful picture of what governance looks like inside many regional and community banking organizations today.
Panelists: Rob Beachy, CIO, Sturgis. Rusty Haferkamp, EVP & CISO, Central National. Bradley Enneking, CIO, Mutual Bancorp. Ryan Dopp, SVP & CIO, Sunwest. Session moderator: Gerald Kierce, Co-Founder & CEO, Trustible.
The views expressed during the discussion are those of the individual participants and do not necessarily represent the views of their employers or other participants.
Finding the AI that is already in use
Several panelists described the difficulty of establishing an accurate inventory of AI usage. It is no longer sufficient to count licensed ChatGPT or Copilot users. Employees may reach AI through websites, browser-based services, development tools, and features embedded in software the bank already licenses.
One CISO from a community bank in the south-central U.S. approaching $2 billion in assets described how a short, board-approved AI policy from 2024 evolved into a more comprehensive approach that includes monitoring which AI tools employees are using, restricting unauthorized applications, and formally reviewing new tools before approving them. He illustrated the rapid growth of AI activity, describing how one of the bank’s monitoring reports went from 178 AI-related entries to more than 900 within just a few months. Currently his bank permits Microsoft Copilot as the only AI client approved for installation on bank PCs and uses security tools to detect other AI activity needing review. Others in the discussion echoed that their banks’ technology and security teams likewise monitor endpoints and favor tenant-controlled AI tools to keep bank data within established protections.
A bank CIO at a Midwestern community institution emphasized the benefit of giving employees a clear path to propose AI uses because a blanket refusal can push experimentation out of sight. He also mentioned his team provides training on the organization’s AI policies for the whole workforce, not just employees with dedicated AI licenses, because AI is increasingly built into an otherwise approved vendor application.
“You start telling your end users no and all that’s really doing is pushing them to the background. You’re just not going to know what they’re doing.” – Rob Beachy, Bank CIO, ~200 employees, ~$1.0B in assets
Approve use cases, not just platforms
The session moderator emphasized the distinction between evaluating a whole platform versus evaluating individual AI use cases offered by the platform. “You could adopt Copilot and you can use it for 15 different purposes and each purpose potentially presents a different risk.” Using AI to draft internal correspondence, summarize loan applications, assess job candidates or interact with customers calls for distinct scrutiny, even when the platform itself is approved.
A CIO of a community bank located in the Midwest agreed this distinction is important because approving an AI platform does not necessarily mean employees are authorized to use every feature or capability it offers. At that organization, when an employee or department wants to introduce a new AI application, the request goes before a cross-functional, board-level governance committee that includes executive management, IT, information security, compliance, risk and HR. The bank maintains both a list of approved AI systems and a separate register of approved use cases specifying which systems can be used for each activity.
Another CIO from a New England-area mutual banking organization with approximately $10 billion in assets said the bank relies on a cross-functional AI working group to review new requests. The focus has shifted from employee requests to use tools such as ChatGPT toward evaluating AI capabilities that existing and prospective software vendors are introducing into their products. This has led the bank to develop an extensive AI-specific vendor questionnaire and strengthen contractual requirements around how vendors can use the bank’s data.
Vendor due diligence needs AI-specific questions and contract language
Participants said that while existing third-party risk management processes remain useful, they do not always address the new types of questions AI introduces: Is customer or bank data being used for model training? What data is transmitted to another provider? Where does it go? What monitoring is available? Who is responsible if an output is wrong or a breach occurs?
One CIO described developing an extensive AI-specific vendor questionnaire to address gaps in the bank’s existing third-party risk assessments. The team has also established graduated positions for contract negotiations: vendors may use AI without touching bank data; an alternative permits use of bank data solely to deliver contracted services; and a more permissive option allows model training only with aggregated or anonymized information. This translates abstract data concerns into specific contractual boundaries.
Another CIO from a business-focused bank in the Western U.S. raised a software supply-chain concern, describing how a team had requested approval to use a library in its custom development work. As Ryan Dopp, SVP & CIO of Sunwest Bank, explained, “It’s really, really important we understand what libraries are being used in any kind of custom development. How are those getting updated? How are we ensuring that they’re getting updated?” He added that this scrutiny extends to vendors, including their security controls and contractual liability if something goes wrong.
But the group acknowledged that eliminating all vendor-related risk is unrealistic. One CISO drew a comparison to lending, where loan officers routinely assess a borrower’s financial condition and other factors before deciding whether the risk is acceptable.
“You have to do your due diligence, but unless you’re going to go down a rabbit hole and never do anything, you have to get to a point where you trust somebody.” – Bank CIO, ~200 employees, ~$1.0B in assets
The takeaway was to treat AI disclosures, data rights and responsibilities as part of procurement and contract review, including when a familiar vendor quietly adds AI features to an established product.
Asking AI providers for monitoring the bank can actually use
One blind spot participants brought up was that although some AI platforms provide administrative reports showing usage volumes, they offer limited visibility into what employees are actually doing with the tools or what sensitive information they may be sharing. Without sufficient activity logs, one team has resorted to manually reviewing selected users’ interactions to verify compliance with the bank’s policies which is a labor-intensive process that is difficult to scale.
Some in the group felt that vendors providing AI platforms should give institutional customers access to more detailed activity logs, ideally in a format that can be integrated into the bank’s existing security information and event management (SIEM) system. Another CIO described an alternative approach already implemented at the bank using a separate security solution that monitors AI inputs and outputs and can actively block employees from submitting prohibited information.
“We’ve got to know where our data is. Examiners, regulators, and customers, they’re going to want to know who’s touching it.” – Bank CISO, ~110 employees, ~$1.5B in assets
The discussion highlighted an important consideration when evaluating vendors that provide AI platforms or introduce AI capabilities into existing software: will the bank have sufficient visibility into how the technology is being used, and can it monitor and enforce its own policies after deployment? These capabilities may be just as important as the security assurances a vendor provides during the initial approval process.
Another real-world example involved contact-center software providers that are introducing extensive AI capabilities while their governance practices remain limited or poorly defined. The concern was not simply whether those capabilities worked, but whether the bank could establish appropriate safeguards for how customer data would be handled.
AI agents are not a single risk category
The panel drew a crucial distinction between assisted work and delegated action. Along with the vendor and intended purpose, a bank needs to assess the agent’s access to tools and data, level of autonomy, ability to operate continuously, and the points where a human can intervene.
One CIO separated bounded assistants such as writing coaches and internal review helpers from continuously operating development agents and systems empowered to make customer-facing decisions. Risk depends on what an agent is permitted to access and do, not just whether a product is labeled “agentic.” At that bank, certain internal assistants can operate within established Microsoft controls, but continuously running development agents and AI-driven customer lending decisions are not authorized.
Participants also stressed there are special considerations when it comes to wire transfers, ACH and payment authorization, where speed and automation cannot replace customer authentication and human accountability. As one participant put it, “It’s still going to take a human looking at it.”
Of particular concern was the lack of an effective kill switch for certain externally hosted AI workflows. Once data have been sent through an API to an outside provider, the bank may have limited ability to interrupt subsequent processing.
For some banks in the panel, the decision has been to prohibit autonomous AI workflows altogether for now. One CIO explained that the bank is not deploying independently operating agents because it cannot yet demonstrate sufficient control over their behavior. A possible path forward may be small, single-purpose agents coordinated by an orchestrator that records every handoff and lets the bank stop a workflow before the next action. Locally run models or hardware could also offer greater control over sensitive data and a more direct way to interrupt an agent.
Policy, process and evidence: what examiners expect
While the regulatory playbook for AI is still evolving, panelists noted that examiners are applying familiar expectations: banks must be able to explain their AI policies, demonstrate how those policies are implemented, and provide documentation showing that appropriate oversight and risk assessments have taken place.
“What’s your AI policy? How are you governing it? Now show us evidence that you’re doing that.” – Bradley Enneking, Bank CIO, ~870 employees, ~$8.4B in assets
Examiners have requested evidence of how institutions vet AI platforms and use cases, and one CIO noted that documented intake forms and risk assessments serve as proof of oversight. The panel also emphasized educating business stakeholders about the controls they must enforce, with one participant noting, “You’ve got to tell the CEO, you’ve got to tell operations, you’ve got to tell HR.”
One CIO drew a line at automated lending decisions. If AI influenced an approval or denial, the bank would need to explain that decision to regulators. Until the technology can meet that explainability requirement, the team is limiting AI to assistance roles such as correspondence drafting and internal budget analysis.
Practical actions for bank technology leaders
Across the discussion, several actions stood out for institutions looking to strengthen governance without waiting for a perfect framework:
- Inventory actual use, not just licenses. Include web services, endpoint activity, developer tooling and AI embedded in existing vendors.
- Create a documented intake path. Require a business owner, purpose, data categories, risk review and explicit operating restrictions for each proposed use case.
- Keep systems and uses linked. Maintain both an approved-tool inventory and a use-case register indicating which tools and data are permitted for each activity.
- Update third-party review and contracting. Ask about training rights, data handling, controls, incident responsibility and meaningful access to logs.
- Gate automation by consequence. Distinguish drafting assistance from persistent agents, financial transactions and customer-facing decisions; preserve human review where needed.
- Retain evidence as the program operates. Keep approval records, exceptions, assessments and ongoing monitoring available for management and examiners.
About Trustible
Trustible is an AI governance platform for banks and other regulated organizations that helps teams identify and inventory AI systems, assess risks, document reviews and approvals, and monitor systems after deployment. Its platform brings AI use cases, vendors, and governance requirements together to support responsible adoption. Contact Gerald Kierce to talk more about your goals and how Trustible can simplify AI governance for your organization.
Gerald Kierce, Co-Founder & CEO
www.trustible.ai