92% of people bought the flour anyway. A penny over a round-number limit is the kind of thing a person waves through without a second thought. The AI almost never did. Across thousands of runs it held the line at $10.00 and turned the sale away.

The researchers, Matthew DosSantos DiSorbo, Sinan Aral, and Harang Ju, then showed the models how people actually reason about a rule like that. The AI learned to bend it. And it carried that flexibility into decisions it had never seen before: hiring, lending, university admissions, customer service.

Here is why this matters if you are about to hand a workflow to an agent. Out of the box, an AI follows the rule you give it to the letter, including the moments a person would have known better. That works for a process where the rule really is the rule. It falls apart on anything that runs on judgment, where the exceptions are most of the job.

So we draw the line the way the research points. We let an agent handle the work where the rule genuinely is the rule, and we keep a person on anything where the right call depends on context the rule cannot see. In a regulated setting that is not a preference. A reviewer signs, an auditor can ask about it later, and the agent never gets to decide on its own what counts as close enough. It is the same reason clinical-ops work keeps a coordinator in the loop and why how the system is governed is the first thing we settle.

Source: MIT Initiative on the Digital Economy, four studies on agentic AI, 2026.