The most useful question to ask about an AI agent is not “how clever is it?” but “what is it allowed to do without asking?”
Action classes
We sort every action an agent might take into classes, each with its own rule:
| Action class | Examples | Rule |
|---|---|---|
| Read and analyse | Summarize documents, prepare a report | Allowed |
| Internal write | Create a task, draft a reply, update an internal record | Allowed and logged |
| Outward-facing, financial or production-changing | Send a message to a customer, make a payment, change a live system | Owner approval every time |
| Legal | Sign, commit the organization | Never delegated |
Decision cards
When an agent reaches an action that needs approval, it does not just stop. It prepares a decision card: what it proposes, why, what it based that on, and what happens if the answer is no. A person approves, edits or rejects it in one place.
Every action leaves a record
Agents log what they did and explain why. That record is what lets a manager trust the system, an auditor review it, and an engineer improve it.
Why this matters
Organizations do not adopt agents because they are impressive. They adopt them when they are predictable. Permissions, approval gates and an honest record are what turn an impressive demonstration into something people can rely on at work.