I watched a chief executive do something for five consecutive quarters that most of her peers would have called a waste of an hour. At the top of every quarterly operating review, before a single result was discussed, she had the original approval conditions for each AI-assisted workflow read aloud, exactly as the board had recorded them: who reviewed the output, how many people signed off, what triggered an escalation, and who held the authority to stop the process. Then she asked the room one question. Is this still true?
For four quarters the answer was yes, and I could see the executive team's patience thinning. In the fifth quarter, an operations vice president said something careful. A weekend accommodation on one workflow had let a single on-call reviewer stand in for the two-person sign-off the board had approved. It had started as a one-weekend fix eleven weeks earlier. Nothing had gone wrong. It had become the default. No one had decided that. It had simply kept being the way the weekend ran.
THE LAW OF THE STANDING CONDITION
Every approval a board grants is conditional on a configuration that existed the day it was granted, and any condition nobody re-verifies has already begun to expire.
I did not learn this law from an AI system. I learned it years earlier, watching boards approve acquisitions, plant expansions, and lending programs, then never ask again whether the assumptions they approved were still the assumptions the company was operating under. The board takes a photograph of the proposal, votes on the photograph, and files it. The company runs a film. Every frame after the vote is a small, locally reasonable adjustment made by a person solving a real problem with no reason to know what the photograph looked like. AI did not create the gap between the still image and the moving one. It shortened the time in which that gap becomes consequential, because these systems act on their own recommendations at a pace no quarterly meeting was built to follow.
The mechanism is not negligence. It is the ordinary physics of an organization under staffing pressure. A reviewer calls in sick and a supervisor authorizes a one-person sign-off for the day. The day passes without incident, and the absence of an incident gets read, locally, as evidence the change was safe. The next week inherits the change. The next budget cycle encounters a workflow that has already proven it can run on less, and the case for running it on less still is easier to make because the first reduction already happened and nothing broke. No single person in that sequence has done anything a reasonable manager would not do. That is exactly why nobody reports it upward. There is nothing, at any single step, that looks like a decision worth reporting.
The cost of ignoring this law does not show up as a malfunction. It shows up as a question the board cannot answer. When a regulator, an auditor, or a plaintiff's counsel asks what conditions the board approved and whether those conditions were in force on the day something went wrong, a board that approved a two-person sign-off and was actually running an automated overnight pass-through has no good answer. The harm in that moment is not the drift itself. It is the elapsed time between when the organization could have known and when the board learned. The National Association of Corporate Directors' 2025 Board Practices and Oversight Survey found that six percent of boards maintain AI-specific management reporting metrics, which means, mechanically, that for most boards there is no standing channel through which a weekend accommodation on a review step would ever arrive. The information does not fail to reach the board because someone hid it. It fails to reach the board because no one built the road.
What that chief executive understood, and what I have seen very few others understand, is that an approval is not a permission the company received. It is a promise the company made. She treated the board's recorded conditions as a contract her organization had signed, and she treated the quarterly question as the moment the contract came due. There is a name I use for the effect that had on the people around her. I call it the Expectation Elevation Model, and its work here was specific: her executives stopped believing that the board's approval was something to be obtained and started believing it was something to be kept. The vice president who spoke up in the fifth quarter did not report the drift because a policy required it. He reported it because the question had been asked four times before, in that room, and by then everyone understood what it was for.
The application does not require a new system. It requires that someone re-read the approval before the results, and that the reading be treated as the beginning of the meeting rather than a formality ahead of it. Name who owns each condition. State what happens, and to whom, and within how many days, when a condition stops being true. Then ask the question every cycle, including the cycles where the answer is yes, because the cycles where the answer is yes are what teach the room to give the honest answer on the cycle where it is not. A board can run the same discipline from its side of the table with one line on every AI oversight agenda: which of the conditions we approved are no longer true?
She retired the following year. The question did not. I checked, some time later, and the operating review still opened with the conditions read aloud, run by people she had developed and asked by executives who had never met her. That is the only measure of leadership I have ever trusted. Not whether the discipline held while she was in the room, but whether the people she produced kept asking the question after she left, and whether they built it from conviction rather than from the audit that would otherwise have found it for them.
The board approved a photograph. Your company is running a film. Someone has to keep asking whether they still match.
Glenn E. Daniels II, Touch Stone Publishers