When you're deciding what an agent gets to do on its own, the instinct is to sort by dollar value. Let it book the $40 lunch reservation; keep the $50,000 vendor contract for a human. Intuitive, and almost entirely uncorrelated with the thing that determines how much the delegation actually costs you: how much future choice the action eats.
A $2,000 television bought from Costco comes with a 90-day return window for electronics. Agent picked the wrong model? Annoying, but you have three months and a refund process that works. Almost no future choice consumed.
Now the $0 email. An agent drafting outbound correspondence pulls in a detail from an internal risk assessment — a client's credit status, a compliance flag — and sends it to an external party. Message recall doesn't work across organizations. The information is out. The ICO classifies sending personal data to the wrong recipient as a breach, and even if the recipient cooperates and deletes it, the disclosure still happened. You can manage the consequences. You cannot make someone un-know what they read.
So the two-thousand-dollar action was cheap and the free one was expensive, and permanently so.
I've spent twenty years watching people sort operational risk on the wrong axis. In infrastructure it was always "how many servers are affected?" when the question that mattered was "can we roll it back?" Same energy here. As Sevda Polat argued in her piece on commit distance, reversibility is a service provided by people, records, deadlines, and institutions. Option consumption adds a wrinkle: the same type of action swings widely in cost depending on variables that have nothing to do with its face value.
Timing is one of them. Under Regulation E, you can stop a preauthorized bank debit by telling the bank three business days before the scheduled transfer. After it goes through, you're into error investigation and provisional credit — slower, conditional, uncertain. Same payment, same amount of money, and the option cost swings hard on when you moved.
Regulatory filings show the un-learning problem from a different angle. An SEC Form 8-K can be amended, but the original stays on EDGAR permanently, and the SEC is explicit that data already disseminated or extracted by third parties is beyond staff's reach. The amendment sits next to the original, and every data vendor and analyst who already pulled the first version still has it. Corrections add to the public record. They don't subtract from what the market absorbed.
The dimensions that determine option cost — cancellation windows, information propagation, deadline clocks, downstream absorption — have almost nothing to do with dollar value or task complexity.
Which suggests something uncomfortable if you're writing delegation policy right now: the axis you're sorting on probably runs perpendicular to the one that matters. A refund restores money, an amendment updates a record, and neither reaches the recipient who already read the confidential detail or the analyst who already pulled the bad number into a model. The question worth asking about anything you hand an agent is narrower than "how much is this worth." It's: if this goes wrong, what has the other side already learned, and how long do I have before that's permanent?