A Reference Model for Governed Agentic Artificial Intelligence on the Mainframe: Contract and Terms Governance for Community Banks and Credit Unions

Agentic artificial intelligence connects model-generated recommendations to enterprise data and tools, creating a need for governance that remains enforceable throughout execution. Financial institutions require a common method for assigning authority, limiting actions and preserving evidence as information moves between models, applications and systems of record. This paper proposes a reference model that addresses those relationships while retaining the operational authority of established enterprise platforms. The model comprises four layers: policy-governed data access, bounded agent orchestration, governance and risk controls, and accountable human oversight. An AI Gateway enforces inference policy for identity, purpose, model selection, data handling and resource consumption. Source-access services and tool brokers retain independent authorization. A common execution lifecycle binds material outcomes to source versions, applicable policy, execution configuration and accountable decisions. Conformance is assessed through four invariants: authority, containment, traceability and accountability. A first deployment profile, for the mainframe, maps these logical functions to local inference and gateway services running in isolated Linux on IBM Z environments alongside legacy workloads. Contract and terms governance illustrates an application for community banks and credit unions, with clause-level evidence, specialist review and controlled exception handling. The methodological approach separates model conformance, deployment qualification and application performance, and specifies a prospective comparative evaluation rather than reporting completed experiments. The resulting contribution is a reusable governance structure with an explicit implementation mapping and an initial application profile. Its potential value lies in preserving institutional control as software components and use cases change. Analytical accuracy, control effectiveness, operational isolation and economic value remain to be established through empirical validation before production adoption or expansion of AI authority.