The Part You Can't Hand Off

Cory spent last week walking three teams through the inventory system we're building for Wine to Water, and the thing he reported back was not “they loved it” or “they hated it.” It was that they kept getting stuck on the number 50.

Somewhere on a screen there was a value that said 50, and finance wanted to know why it said 50, and once that question was in the room nobody could talk about anything else. Not the workflow. Not whether the sequence of screens matched the way anyone actually moves a pallet. Just 50. Cory kept trying to pull them back up to the level of “does this match how you work,” and they kept diving back down to the level of “that number is wrong.”

I want to be clear that this is not a story about users being difficult. It's the opposite. People will not tell you what they need in the abstract, because in the abstract they don't know. Ask a warehouse team what their ideal inventory system looks like and you'll get a shrug and maybe something about barcodes. Show them a half-finished screen with a wrong number on it and suddenly they have opinions, detailed ones, going back years. The wrong number is the invitation. It gives them something to push against, and what they say while pushing is the actual requirements document.

One guy on the manufacturing side was the one who really engaged, which is why he's getting the beta. Somebody in finance wants to be able to adjust values directly in the system, and Cory is holding that off, correctly I think, until we've settled what the workflow is, because the fastest way to end up with a system nobody trusts is to let people start hand-editing it before it's right.

Here's the part where I get to admit I got something wrong.

I had built the Pack Kits workflow so that you log what you finished. You packed the kits, you come to the screen, you tell it how many. Perfectly sensible, and it maps almost exactly to nothing that happens in that building. What actually happens is that somebody schedules a run of 1,500, and then that run takes several days, and people need to see how far along it is, and at the end there's a QC step that somebody wants on paper with a checkbox. I built a receipt. They needed a plan. I would not have figured that out by thinking harder about it at my desk, and no model was going to tell me, because the information simply wasn't anywhere except in the heads of people who had never been asked.

Same with the little stuff, which turns out not to be little. Nobody counts in units. They count in boxes plus however many loose ones are sitting on top, so the input needs to accept both. A full pallet is 1,500, so that should be a button and not something you type. Unassembled filter builds move from L2 to L3, always, so the location field should already know that. Every one of those is a five minute change and every one of them is the difference between a system people use and a system people work around.

When Cory and I got on the call today, he fed the whole feedback doc into Claude while we were talking and it came back with 27 distinct changes. Which is great, and it's genuinely the reason a two person effort can rebuild a real inventory system in weeks instead of quarters. But notice the order of operations. The model was extraordinary at turning a messy pile of human reactions into a clean, prioritized, buildable list. It contributed absolutely nothing to generating that pile. That came from Cory sitting in a room with three teams and letting them be annoyed at him for an hour.

This is what I keep meaning when I say the interesting work is the unglamorous work. There's an enormous amount of noise right now about AI-generated everything, and most of what people are calling slop is not slop because a machine made it. It's slop because nobody checked it against reality. A system built from a generic template for a generic nonprofit would look completely finished and would be useless to Wine to Water specifically, and the gap between those two things is entirely made of conversations. Somebody has to know that finance wants a ledger view showing cost at each movement, that transfers get entered by manufacturing and verified by the warehouse like a purchase order, that the eventual dream is serializing every individual filter so you can trace it from build event to the family that received it. None of that is in a template. All of it came from talking.

The building got faster. The knowing did not.

You'll build it three or four times. That's fine. Each round costs an afternoon now, which is the genuinely new thing about this moment, and much less new than everyone is making it sound. What hasn't changed at all is that you have to go ask.

Cory's spending this week logging everything that looks wrong in the next version. He's told me to expect a lot. I'm counting on it.