What I do — and how I do it differently.

Regulated institutions are deploying agentic AI faster than the teams governing it can interpret what the control failures actually look like. This workshop gives risk and compliance teams a working framework they can apply to systems already in production — not principles, but a structured method. The curriculum covers four areas: the AI Control Failure Taxonomy and its five failure modes, examined through incident-based case analysis; why existing penetration testing and compliance review methodologies do not detect agentic control failure; the EU AI Act and FINMA regulatory obligation map for deployed agents; and a structured assessment method participants leave with and can apply to their own deployments. Two booking models: institutional session for one organisation, or open enrolment with participants from multiple institutions.

The EU AI Act places deployer obligations on regulated institutions from August 2026. FINMA Guidance 08/2024 applies now. Most institutions subject to both do not have a dedicated AI Risk Manager in place — and building the function from scratch while navigating active compliance timelines requires someone who understands both the regulatory architecture and the operational layer of deployed AI systems. This engagement embeds that capability directly into the institution's risk and compliance function for a defined period — typically three to six months at 20–30% FTE — producing a functioning AI risk framework, documented control architecture, and a governance structure the institution owns after the engagement ends.

Most regulated institutions have an AI governance policy. Few have the operational architecture that makes it enforceable at the deployment layer. This mandate closes that gap — translating governance commitments into the control structures a regulator can audit and a risk committee can act on. Work covers AI control framework design, Agentic Control Plane architecture, EU AI Act and FINMA readiness, behavioural governance implementation, and board-level reporting structure. The Agentic Control Plane is the live operational layer giving every department real-time visibility into what an agent is doing, under whose authority, and within what boundaries — the difference between a governance document and a functioning control system. Each mandate is scoped in a discovery call, with a written proposal returned within one week.

This assessment exists for an operational decision. Traditional penetration testing examines whether someone can break into a system. An Agentic Control Assessment examines what the deployed system does when it pursues its objective by a route no one authorised. The output is not a risk score for a third party — it is a behavioural capability map for the team responsible for the deployment. Four structured assessment categories are applied to live production systems: goal displacement through context manipulation and crafted inputs, permission escalation by chaining individually authorised actions, audit completeness to determine whether unauthorised actions leave a traceable record, and cascading failure to assess whether one agent's output can function as malicious input to a second agent in the same environment.

This assessment exists for a risk transfer decision. Agentic AI liability is being underwritten without a methodology built for systems that pursue objectives. Traditional cyber questionnaires were designed for systems that wait for instructions — they do not reach the behavioural control layer. This assessment evaluates a deployed agentic system's control architecture against five failure modes and produces output structured for an insurance decision: a per-failure-mode risk score, an aggregated underwriting rating, and a one-page carrier summary. Two delivery models: carrier-commissioned as pre-bind due diligence, or insured-commissioned as a condition-of-coverage requirement.

Governance frameworks delivered as advisory documents typically name the right principles but do not reach the control architecture of the deployed agents. This review identifies exactly what the prior assessment missed at the agentic execution layer — and documents it with the specificity a regulator or an incident investigation would require. The scope is fixed to that gap: not a repeat of the prior engagement, but a forensic review of what it did not reach.

Available for keynote and panel engagements on agentic AI control failure, behavioural governance, and the gap between current compliance frameworks and the execution layer of deployed AI systems. Sessions draw directly from primary research, regulatory analysis, and the AI Control Failure Taxonomy. Delivered in English and German. Current topics include: AI Agents Don't Break the Rules — They Follow the Wrong Ones; The Governance Gap EU AI Act Does Not Close; and What Forensic Investigation Reveals About Agentic AI Control Failure. Upcoming: WIC Day 2026, Kongresshaus Zürich, September 17.