I have sat in board meetings where every director left believing the group had agreed on something, and none of them could later say, in the same words, what that something was. It happened again this year, in a session about an AI system nobody in the room fully understood, where three directors nodded at a proposal and one asked a single clarifying question that never got answered before the agenda moved on. Eight months later, the system made a decision that hurt someone. When the board tried to reconstruct what it had actually approved, it found four accounts and one shared feeling that everyone had been in agreement.
THE LAW OF THE DIVIDED ROOM
A decision that exists only in memory does not have one version. It has as many versions as there were people in the room, and the distance between those versions is invisible until something forces it into the open.
Why the Room Is Not Enough
This is not a new failure. I watched it long before anyone asked a board to oversee an algorithm. A strategy gets discussed at length, everyone nods, the minutes record “approved” with no further detail, and eighteen months later two executives are arguing about what was actually decided, each one certain of a different scope, a different budget, a different owner. AI has not created this pattern. It has raised the cost of it, because the systems now in question move faster than the quarterly meeting that was supposed to be governing them, and because the question a regulator or a plaintiff eventually asks is never “did the board discuss this.” It is “what did the board decide, and where is that decision written down.”
The mechanism is not complicated, and it is not a character flaw. Human memory is not a recording. It is a reconstruction, rebuilt every time it is retrieved, shaped by what each person already believed walking into the room. Five directors hearing the same fifteen minute discussion will store five slightly different summaries, and none of them will notice the divergence until they are asked to produce a single, specific account under pressure. A board that relies on shared memory instead of a shared record is not aligned. It only feels aligned, right up until the moment alignment is tested.
The cost of ignoring this law does not arrive on the day of the vague meeting. It arrives later, at the worst possible time, when a regulator, an auditor, or an attorney asks the board to produce what it actually decided, and the board discovers it has four people willing to describe four different answers with equal conviction. At that point the argument is no longer about whether the board’s judgment was sound. It is about whether the board can prove it exercised judgment at all. A good decision with no record loses to a mediocre decision with one, because the record is what the process is judged on once memory stops being sufficient evidence.
I do not tell boards to write everything down because paperwork is virtuous. I tell them because the record is the only version of the decision that outlives the room. The directors who sat in that AI discussion will eventually leave the board, retire, or simply lose the nuance they carried out the door. What replaces them is not their memory. It is whatever the organization actually wrote down. If nothing was written down, what replaces them is nothing.
The application is not complicated either. Before the meeting ends, someone states out loud what was decided, in one sentence specific enough that a person who was not in the room could act on it. That sentence gets written down, with a name attached to who owns it and what happens if the threshold it describes gets crossed. This takes ninety seconds. Most boards skip it because the room feels aligned and ninety seconds feels unnecessary. The boards that survive contact with a hard question are the ones that did not skip it.
The room you were in will not be there to defend you later. The record will.
Glenn E. Daniels II, Touch Stone Publishers