Not Quite Right
The last several weeks ate my headspace. School, kids’ activities, birthday, a couple of holidays, work running hot. Stack enough of that and you start to feel like you’ve got nothing worth saying. This was forcefully pulled out over a series of reflections this weekend. Here’s what I keep relearning: consistency beats brilliance. I don’t need a breakthrough every week. I need to keep showing up… keep using the tools, keep trying things, keep thinking out loud. The good weeks come out of the boring ones. So consider this one a thinking-out-loud, no trophies to show here.
The Setup
I support an account where I’m not in the day-to-day. I don’t know the work the way the people running it do. I learn it through conversations. Calls, recaps, transcripts. I sit in, I ask questions, I listen for what matters. 8 months ago I wanted AI to help me hold all of that. So I tried to build the first knowledge system for the account. And my assumption, the one I didn’t recognize, was that AI could read the intent across everything the team produced, synthesize it, and surface the principles that mattered on its own. Feed it enough of what the team knows and the meaning would fall out. It was naive of me back then to “automate everything with AI”. (I know.)
My Approach + AI Role
My first instinct was to pour in the artifacts. I gathered a pile of tactical decks, loaded them up, and leaned on the AI to detect the intent behind them, synthesize it, and tell me what mattered. The logic felt right at the time. AI is genuinely good at intent and gist, so give it enough raw material and let it reason its way to the principles. On paper, that should have worked.
What Actually Happened
It failed early. It could cite the team’s work back to them. It could surface the right language at the right time. What it couldn’t do was reason about any of it. Team’s response said it all: “it’s just not quite right even though it sounds like us”. It got the gist. What it missed was the judgment. The deeper kind, where someone on the team knows something works because they’ve tested it against the real work. You can’t pull that out of decks alone, because what a team knows carries personal history. Why someone pushed back in one moment and let something go in the next. The meaning sits a layer underneath what gets written down. What dawned on me recently is that I’m already working around this somewhere else. When I take recap notes, my custom summary gets the gist but always misses the nuance. So I paste that summary back in and ask Copilot to QA it… where did this miss, what got flattened. Between the first summary and that second nuance pass, I can finally form a real read on how the conversation actually went. There are always misses. That’s the point. You need two anchor points before you can triangulate your own meaning. One pass is a guess. Two passes, read against each other, give me something I can actually judge from.
The Real Insight
For now here’s what I’m at. The real work is creating meaning, and the materials don’t hand it to you, you’ve got to build it. Which is why being in the meeting isn’t passive for me. I’m there to ask the questions and form my own read on what matters while the conversation is still live. The meaning starts forming in the room, before any transcript exists. Then I point AI at the transcript, but never empty-handed. I bring my own read of what mattered in that room, so the model focuses its attention the way I would and articulates the principles the way the team actually meant them. My meaning plus the transcript is what tells it where to look and how much each thing should carry. That’s the difference between feeding it documents and feeding it judgment. It’s slower. It’s less impressive than “AI doubled our output.” And there’s a harder layer underneath this… the trigger, knowing when a judgment is even needed. That’s what a tidy playbook quietly skips. (As I did before.) No result to report yet. We’ll find out soon whether my current second attempt at building the knowledge layer for the team holds.
Try This If…
If you’re building a knowledge system for work you don’t live in day-to-day, start with the conversations. Be in the meetings. Ask questions. Form your own read before you ever touch a deck. For recaps: don’t trust the first summary. Paste it back, ask the AI to QA its own work… where did it flatten something, what nuance got lost. The gap between those two passes is where your understanding forms. Then bring that meaning to the documents. You’re feeding it judgment, not just files. I’m still testing all of this. But it’s a better starting point than where I started.
Systems Lens
What I’m really building is a knowledge layer that sits between the team and the tools. A second brain for an account I don’t run. The point is that AI works with the team’s actual judgment instead of generic best practice. I’ve come to think you can’t skip the meaning-making. You can amplify judgment once it exists. You can’t manufacture it by pouring in documents and hoping the model figures out what matters. We have to make meaning in the work first. Then AI can amplify it. That’s the field note. If you’ve built something like this and it came back sounding right but not quite right, I’d like to know what you changed. DM me on LinkedIn.