I spent this morning reading the Companies Act because I wanted to write about AI meeting notes, and I did not want to write the post that already exists forty times over.
Section 73(7) stopped me. Board resolutions, it says, must be dated and sequentially numbered.
Dated is fine. The date is in the room. Everything else the Act asks for is in the room too: what was discussed, who declared an interest, what was decided, who chaired it. Give a recording to any competent transcription tool and you will get all of it back, in order, with names attached.
The number is not in the room.
Where the number actually comes from
A sequence number is not a fact about the meeting. It is a fact about every meeting before it. Resolution fourteen is only resolution fourteen because thirteen others were numbered first, and that information lives in a register somewhere, not in anything anybody said out loud.
So the most complete recording of a meeting that has ever been made cannot produce it. Not because the model is not good enough. Because the value is not there to extract. You could put a microphone in front of every director for a decade and the number would still have to come from the file.
I found that genuinely clarifying, and then slightly annoying, because I have hit this before and never had a name for it.
The same field, everywhere
Every automation I have built has one of these. One field that looks like the others and behaves nothing like them.
Invoice numbers are the obvious case. Everything else on an invoice is derived from the job: the client, the work, the dates, the totals. The number is derived from the last invoice, which is a completely different kind of question, and it is the field that goes wrong. I have watched a numbering scheme collide with itself because the number was being chosen by whoever was typing rather than read from the register. Nothing else on the document could fail that way.
File naming is the same. Version numbers are the same. Anything that answers "which one is this" rather than "what is this" depends on state held outside the event, and that dependency is invisible until two things are created at once.
The pattern is: content comes from the event, identity comes from the register. Automation is very good at the first and structurally incapable of the second, and the reason it looks like one job is that they arrive on the same piece of paper.
Why this changed the post I was writing
I had planned to write a piece about turning a call into actions nobody has to re-type, which is a real problem and a good post. What I actually shipped is that post with a spine through it, because the Act draws the line for me in a way I could not have drawn myself.
Section 73(6) says the minutes must contain every resolution adopted. Section 73(7) says the resolution is dated and numbered. Section 73(8) says the thing becomes evidence when the chair signs it. Read those three in a row and the boundary is drawn precisely where I would want to draw it as a builder: automate the transcript, automate the draft, automate the actions and their owners and their dates, and then stop, deliberately, before the step that assigns identity and the step that assigns responsibility.
That is not a compromise between automation and caution. It is the actual seam.
What I keep getting wrong about this
My instinct, every time, is to try to close the gap. To have the system read the register, take the next number, write it in, and hand back a finished document. And you can build that. It is not hard.
The reason I keep not doing it is that the failure mode is silent. If the transcript is wrong, somebody reads it and notices. If the number is wrong, nothing looks wrong at all. You get a perfectly formatted resolution with a number that already belongs to something else, and you find out years later when somebody goes looking for a decision and finds two.
Content errors announce themselves. Identity errors wait.
So the number stays a human step, or at minimum a step that fails loudly rather than proceeding on a guess. It is thirty seconds of work and it is the only part of the job that cannot be reconstructed later from the recording.
The part I did not expect
I went to the Act to find something nobody else writing about AI meeting notes had bothered to look up. I came back with a rule about my own build work that I have needed for two years.
That has happened enough times now that I should probably stop calling it a coincidence. The compliance text is not a source of trivia to sprinkle over a post. It is a specification, written by people who had to think very carefully about what a record has to be in order to be worth anything, and it keeps describing the systems I am building more precisely than my own notes do.
Today it told me where to stop automating. I did not have a principled answer to that question yesterday.