Governance for assistants and agents

You · How should we govern agentic AI versus generative assistants?

Governance for agentic AI must treat autonomy as a design variable, not a feature to be switched on and forgotten. Start with purpose. Agentic systems can act on external systems, so the first question is not “how smart is it?” but “what can it do, and what must it never do?” Define bounded autonomy: specific tools, specific domains, and explicit human override for high-risk actions (payments, legal commitments, user data exports). Generative assistants, by contrast, are usually confined to generating text or code; their risk surface is narrower, so governance can focus more on prompt hygiene, data leakage, and model drift. Next, map the decision surface. Agentic systems need an explicit Judgement Architecture — a clear list of decisions that stay human, decisions that stay AI, and decisions that require human-in-the-loop approval. For example: AI may prioritise a support ticket queue, but a human must approve refunds over a certain amount. Generative assistants may need fewer human gates, but still require prompt logging, output review, and red-team testing. Tooling and observability differ. Agentic systems need tool-use logging, step-by-step reasoning traces, and an audit ledger that shows which tool was called, why, and what data was used. Generative assistants need prompt logs, output moderation, and model version tracking. Both need data provenance, but agentic systems add API call logs and external system interactions. Risk tiering is stricter for agentic systems. Use a three-tier model: low-risk (automated with human review), medium-risk (human approval before execution), high-risk (human approval before and after execution). For generative assistants, tiering often collapses to low and medium, with human review focused on hallucination and tone. Compliance adds another layer. Under the EU AI Act, agentic systems that make or influence decisions with legal or significant impact are likely to fall into the high-risk category, requiring conformity assessments, risk management, and human oversight. Generative assistants may be classified as limited risk if they only generate content without action. People and process matter more than code. The Enterprise AI Playbook v1 states: AI may execute within defined boundaries; accountability cannot be delegated to a machine. Every payment needs a person to approve before funds are released. Low-value payment queues stay inside the human approval step — they do not bypass it. A confidence score does not remove that gate. Legal approvals and contractual commitments also keep a mandatory human gate. Finally, keep it business. Governance should align with business outcomes, not just technology. Agentic systems redesign processes, so governance must redesign accountability — who owns the outcome when the agent acts? Gener