An executable doctrine to govern 2,000+ agents

01

Successful deployment created the governance need

I created the AI charter, set the contractual and operating guardrails for the bank’s first internal agent-building platform, and helped train business users and produce the learning content.

Adoption spread beyond technical teams: 2,000+ agents were created within months. Keeping a file-by-file governance model would have cancelled out the benefit of the deployment.

02

EU AI Act and internal requirements became an executable framework

I started with the first complete diagrams, dimensions and rule sets, then brought Digital, IT, Compliance, Security, Legal and Regulatory teams around those artifacts. The group tested and changed them until a common model was validated.

The framework organizes company knowledge in one centralized, versioned form that AI can use. It now assesses any AI system, not only agents. The LLM structures facts; deterministic rules make the decision and preserve its rationale.

03

Control and administration now focus on the exceptions

AIR produces an assessment file in about ten minutes, including the classification, applied rules, rationale and obligations. Under the active framework, 96–97% of the portfolio follows the green path; the high-risk share is smaller still.

Experts can focus on the few cases that require deeper control and review samples periodically. Questionnaires are shorter, duplicate requests disappear and the time saved compounds across every function previously involved.

04

The doctrine supports a portfolio of agents and decision engines

AIR is the internal service built to apply and evolve the doctrine. It imports systems from existing sources and keeps their configuration, assessment, triggered rules, obligations and framework version in one registry. Its exports provide the reference for the EU AI Act register, AI Gateway and governance tools.

I also designed and deployed agents, reusable skills and several decision engines for investment-file analysis, legal review and contract analysis. One compares contracts against the doctrine and internal clause library, produces the analysis and prepares a proposed Word redline with tracked changes.

The standard assistant on the first agent platform was replicated across environments and is now its most-used agent. Backed by a shared skill library, it helps employees produce the prompt and operating guide for an agent, checks whether the use is permitted and adds specialized guardrails. An MCP extension is now being designed to connect it directly to the rules engine, registry and similar agents, skills, connectors and prior decisions.