“If a machine is expected to be infallible, it cannot also be intelligent.”

therestisai
Elevating human intelligence
Governance for assistants and agents
Compass
You · How should we govern agentic AI versus generative assistants?
Governance is not a single switch but a layered system that treats agentic AI and generative assistants differently because their risk profiles and autonomy levels differ. Agentic AI — systems that perceive, decide, and act within a bounded domain — demand stronger governance because they can trigger real-world effects (e.g., API calls, data writes, payment approvals). The core principle is accountability cannot be delegated to a machine. Generative assistants — systems that generate content or suggestions — require governance too, but the focus is more on accuracy, hallucination, and data leakage rather than autonomous execution. Below is a concise, practical framework you can adopt today without waiting for a full policy rewrite. 1. Define the Goal and Scope - Agentic AI: Start with a narrow, well-scoped job description. Example: “Auto-respond to low-risk customer queries using our CRM and knowledge base; escalate anything uncertain to a human.” - Generative Assistant: Define the task as “suggesting drafts or summarising” rather than “approving or executing.” - Boundary rule: Never allow an agentic system to approve payments, sign contracts, or override human legal sign-off. Keep a mandatory human gate before any irreversible action. 2. Risk Tiering (Playbook v1) Apply a three-tier risk model to every capability: | Tier | Capability | Governance | |------|------------|