A few weeks ago, a man at church asked me whether AI could do anything about his meetings. They wander, he said. Ten minutes on the agenda, ten on somebody’s hunting trip. I told him what I’d tell anyone: record it, tell everybody you’re recording, then hand the transcript to the tool with the meeting’s purpose and one instruction — leave out anything that didn’t stay in topic. The hunting trip stays in the raw file. The notes come back clean, with the decisions and follow-ups pulled out.
He said he’d try it. My wife, driving away, said I’d just run a live case study for the AI education material I keep saying I’m building.
I’ve been thinking about it since, because the trick isn’t the transcript. It’s the sentence he has to write before the tool can help him: what the meeting was for.
The tool who notices
Here’s the version of AI that interests me, and it’s a long way from how most people use it. Imagine a tool that has sat in on every meeting at a company for two years, read every ticket and every thread, and one morning says, before anyone asks: the reason your reports keep disagreeing is that three departments use the same word for three different things. Here’s where it breaks. Here’s a recommended fix.
That’s a useful tool, not a vending machine. And it’s the same technology we have today, with memory and context attached.
The hole
Now the hole. Who told it which definition was right?
Nobody, and it doesn’t need to know to see an issue. It can see one word with three shapes and things breaking at the seams. What it can’t do is decide which shape wins. That’s judgment, and judgment belongs to a person.
There’s a bigger hole under that one. Two years of raw transcripts isn’t context. It’s a pile. Nobody said which conversations mattered or why, and a tool fed a pile will cheerfully find patterns in the noise and sound confident about them. Garbage in, eloquent garbage out.
Curation is the work
So the work is curating, and it’s small. After every meeting, somebody writes down what it was for, what got decided, why, and what’s still open. Five lines. Let the AI draft them from the transcript; a human corrects them. The correction is where the judgment lives. It’s what I told the man at church, done every week instead of once.
Most organizations I’ve seen are at step zero, feeding transcripts to a tool and hoping.
Folklore
Some won’t record at all. They’ve been told by legal that less on paper means less to discover. I understand the instinct. I it is a strange gray area. A written decision is discoverable and defensible. An unwritten one is only indefensible, and you find out which you needed after the fact.
Standards with no record of who set them and why aren’t standards. They’re folklore. You can’t audit folklore, defend it, or fix it. Every governance program worth the name runs on a decision log — what was decided, by whom, for what reason, and what it changed. That’s lineage. Nobody who works with data would accept a field with no lineage. We accept decisions with none every day.
The AI tool who notices what nobody asked is only as smart as the decision log it reads. And the decision log is what good governance asked for long before any of this.
We haven’t scratched the surface of AI. We also haven’t scratched the surface of writing down why we do what we do, and the second one is the bottleneck.
