A night shift worker clocked in at 10pm and never clocked out.
Midnight passed. Sunrise passed. Noon the next day and they were still working. Reliable, fast, no complaints, and still at it, because nobody had told them when the shift ends.
The overtime bill arrived on Friday and it was enormous. Every task had been done correctly. Quality was never the problem. The missing piece was one line nobody had written down: when to stop.
AI agents carry exactly that risk. An agent never gets tired, never feels unsure, and never glances at the clock. It keeps going until something tells it not to.
Three lists, before it goes live
What the agent does alone. (opens in a new tab) The routine actions. Send a standard reply. File a document. Update a record. Low risk, high volume, no judgment needed.
What needs a human. The actions where judgment matters. Anything touching a customer directly (opens in a new tab). Anything involving money above a set amount. Anything unusual or new. The agent stops here and waits.
When it stops entirely. A time limit. A spending limit. An error count. Some condition that says enough, shut down and wait. This is the clock-out rule. Without it, your agent is the night shift worker who never goes home.
What that looked like in practice
One team launched an agent to handle internal support tickets. Before it went live they wrote the three lists: what it could close alone, what it had to hand over, and a hard stop after fifty tickets in any twenty-four hours.
The agent handled forty percent of tickets cleanly. The ones it passed on genuinely needed a person. The fifty-ticket cap was never reached, but it was the thing that let the team hand over the work without watching it all day.
Those three lists are your governance. They cost nothing to write and everything to skip.