On 4 November I’ll be speaking at KI KONNEX 2026, the AI summit for the insurance industry that V.E.R.S. Leipzig is holding in Cologne. The talk is called ‘When agents act: AI governance as a design principle’. It picks up where my talk in the V.E.R.S. AI network left off in May — with the argument that governance belongs at the start of an AI project, not at its end — and asks what changes once AI no longer suggests, but acts.
An assistant drafts. An agent acts.
As long as AI only makes suggestions, a human decides in the end. That assumption sits quietly under most of today’s AI governance: case-by-case review before go-live, a human in the loop afterwards. It works for systems that produce outputs for people to act on.
Agents break it. Take one agent and one action: sending an email. If the email is a draft rejection that lands with a claims handler, a person still decides. If the same email goes straight to the applicant, nobody does. Technically the two are identical. For governance they are different worlds. What separates an agent from an assistant is not intelligence. It is whether a human carries out the consequential action.
Autonomy is a design decision
The EU AI Act has contained the relevant word since 2024. Article 3 defines an AI system as one ‘designed to operate with varying levels of autonomy’. When the Commission consulted on its guidelines for that definition, respondents asked for exactly this term to be clarified. With agents, the question moves from interpretation to operation: the level of autonomy is no longer something to classify after the fact, but something to design.
The core idea of the talk follows from that. Autonomy is granted per action, not per agent. The same agent may research freely and still need a human sign-off before it sends anything that cannot be recalled. What decides is the target, the reversibility and the potential impact of the action. No agent runs without an entry in the register, and the boundaries are enforced by the platform the agent runs on, not by the self-assessment of whoever built it.
Why governance sets the pace
Case-by-case review does not scale to agents. Once agents can be built in self-service, they arrive in numbers no review board can process one at a time. Organisations are then left choosing between a bottleneck and shadow agents, and neither is fast.
A framework that decides in advance which actions an agent may take on its own changes that arithmetic. Agents that stay within those bounds go live without individual review, because the platform enforces the limits. Review capacity is concentrated on the few that act externally or come close to high-risk decisions. Governance sets the pace not because it reviews faster, but because it makes most reviews unnecessary.
In Cologne I’ll show how the building blocks from May — from the definition of an AI system to its anchoring in the organisation — extend to agents, and I’ll be explicit about where the open questions remain. If you are attending KI KONNEX on 3 and 4 November, I look forward to the discussion.
Note & sources
This preview is based on my forthcoming talk ‘Wenn Agenten selbst handeln: KI-Governance als Designprinzip’ at KI KONNEX 2026, the AI summit for insurers of V.E.R.S. Leipzig, at the smartvillage in Cologne-Mülheim on 3 and 4 November 2026. It builds on my talk ‘Effektive KI-Governance’ in the V.E.R.S. Leipzig AI network for the insurance industry on 18 May 2026.
- KI KONNEX 2026 — the AI summit for insurers, V.E.R.S. Leipzig.
- Regulation (EU) 2024/1689 (AI Act), Article 3(1) on the definition of an AI system.
- European Commission, analysis of stakeholder feedback on AI definitions and prohibited practices, 12 May 2025.