Designing AI Governance When Systems No Longer Fit Old Categories

Designing AI Governance When Systems No Longer Fit Old Categories

Product and consulting teams like clear categories. A system is a tool, a platform, a service, an API, or an agent. But AI and robotics increasingly blur these categories. David J. Gunkel’s Person, Thing, Robot, based on the title, table of contents, and excerpt, is useful for technology leaders because it shows why classification is not a philosophical luxury. It is a governance requirement.

The book examines the moral and legal ontology of robots. The excerpt explains the central problem: robots are technological artifacts, but they are not quite like ordinary things. They interact socially, talk, simulate intelligence, and encourage people to treat them as more than objects. At the same time, they do not fit easily into the category of persons. The result is a challenge to the traditional person/thing distinction that has shaped law and ethics for centuries.

In AI productization, the same tension appears in practical form. A generative AI assistant is sold as a tool, branded as a collaborator, experienced as a conversational partner, and implemented as a semi-autonomous workflow component. A robotic system may be a machine, a user interface, a safety risk, a data collector, and a decision node at the same time. If governance assumes only one category, the product will develop blind spots.

This matters most when AI systems act across boundaries. A customer-service bot may make statements that create contractual expectations. A coding agent may introduce security vulnerabilities. A procurement agent may trigger purchases. A warehouse robot may affect worker safety. A recommendation system may shape access to services. In each case, the technical design must be paired with a role definition.

The book’s table of contents signals a useful structure for consultants: things, persons, natural persons, artificial/legal persons, both/and, and deconstructing things. These are not implementation categories, but they help teams ask better questions. Is the system merely an instrument? Does it represent the organization? Does it create obligations? Does it need partial capabilities under law or policy? Does its social interface cause users to overtrust it? Does its design encourage the organization to hide human responsibility behind machine autonomy?

For ozycore.de’s audience, the productization lesson is clear: AI governance should be embedded into system architecture. A product should include permission boundaries, audit trails, disclosure mechanisms, escalation paths, human approval points, incident ownership, and model-behavior documentation. These are not add-ons. They express what kind of “entity” the system is allowed to be in the organization.

One practical framework is to define every AI system through five role dimensions. First, agency: can it act or only recommend? Second, authority: whose authority does it use? Third, accountability: who owns outcomes and failures? Fourth, visibility: do users know they are interacting with AI? Fifth, reversibility: can decisions or actions be stopped, reviewed, or corrected? These dimensions turn abstract ontology into product requirements.

Gunkel does not offer a ready-made operational framework, and the excerpt explicitly says the book is more an invitation than a final destination. That is precisely why it is valuable. Technology is moving faster than our inherited categories. Consulting teams that help clients define those categories will build safer and more durable AI systems.

The future of AI governance will not be solved only by better models. It will be solved by better institutional design around models. Before deploying an AI agent, ask a simple question: what is this system allowed to be?