AI Agents at Work: Where Automation Ends and Judgment Begins

The wrong question

Most conversations about AI agents start with capability: can the model do this task? Increasingly the answer is yes, at least most of the time. That makes capability a poor guide to what should be automated.

A better starting point is consequence. What happens when the agent gets it wrong, who notices, and how easily can it be undone?

Two questions, four zones

Ask of every task: how costly is an error, and how reversible is the action? Low cost and reversible, such as tagging a support ticket or drafting a summary, is where agents can act alone, with sampling to check quality.

Low cost but hard to reverse, such as sending an external message, deserves a light confirmation. High cost but reversible, such as reprioritizing a backlog, suits agent proposals that a person accepts. High cost and irreversible, such as pricing, hiring or commitments to customers, stays a human decision, with the agent preparing evidence and options.

Design the handover

Where a person decides, the quality of the handover decides the quality of the decision. The agent should present what it found, where the evidence came from, which options it considered and what it is unsure about.

The decision and its reasons should be recorded alongside the agent's input. That record is what makes the system auditable, and what lets the team move a task into a more automated zone later, once its track record justifies it.

Start small

Map one workflow step by step and place each step in a zone. Automate the low-risk steps first, measure error rates, and only then widen the agent's authority. Trust, like any other capability, is built on evidence.