Your team is reviewing four candidate tasks for AI agent deployment. For each, you need to decide: fully automate (agent acts), augment human (agent recommends, human acts), or stay human (no agent involvement).
(1) Drafting personalized customer thank-you notes after purchases (low stakes, reversible, high volume) (2) Approving credit applications above $50K (high stakes, reversible, low volume) (3) Routing customer support tickets to the right team (low stakes, easily corrected, very high volume) (4) Deciding when to terminate an employee (high stakes, hard to reverse, requires accountability)
What's the most defensible split?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
The Automation Matrix: stakes × reversibility × accountability. Low-stakes high-volume tasks (thank-you notes, ticket routing) are where agents add the most value with the least risk — automate. High-stakes but reversible decisions (credit approvals) benefit from agent analysis but need human judgment for accountability — augment. High-stakes hard-to-reverse decisions with required human accountability (employment termination) shouldn't be agent-driven at all — even with high accuracy, the accountability surface is wrong. Match the automation level to the combined axes, not to any single one.
"Agents are valuable when they handle complex decisions" treats automation as the goal. But automating an irreversible high-stakes decision with no human in the loop transfers accountability to a system that can't carry it. Capability isn't the only consideration.
"Always humans" forfeits the agent's value on the high-volume safe tasks (thank-you notes, ticket routing) where it adds the most leverage with no real risk. Blanket refusal is as un-nuanced as blanket automation.
Training data tells you whether an agent can do a task. The Automation Matrix tells you whether it should. These are different questions. You might have great data on employment termination decisions and still shouldn't automate them — accountability isn't a data problem.