Why the Analysis Was the Easy Part
Sprint 1: fourteen rounds of calibration, and why fast output is built slowly.
Sprint 0 had cleared the ground. The product was defined, the principles were set, and the questions of what we were building and how to build it were already answered. Sprint 1 was where we finally started to build.
———
There's a story people like to tell about AI: you ask, and it answers. You type the prompt, wait a moment, and the machine does the work. That is not the story we’re living. What we learned was almost the reverse: the speed people would eventually see was paid for in advance, with a great deal of unglamorous human judgment applied early and often.
Our first proof came the first time Trial 1 ran end to end, in late April. The output was technically perfect, yet completely lifeless. Every number checked out and every claim traced back to the data, and still it read like a form letter, the kind that uses your name twice and tells you nothing. Accurate to the decimal, but nothing we would ever put in front of a client.
———
Getting the analysis right turned out to be the easy part. Teaching the system to produce something a person would actually want to read, something that sounded like a sharp colleague rather than a template filling in blanks, something that provided real insight, took fourteen rounds of calibration.
Fourteen.
Each round was a small, specific correction: the difference between stating a finding and explaining why it mattered, the phrasing that gave the machine away, the place where confidence tipped into overclaiming. None of it was work you could prompt your way through in an afternoon.
Molly called it a detour. It took longer than the analysis ever did, but the lesson stuck: producing mathematically accurate analysis was easier than producing real, useful insights. The intelligence was close to a commodity; the strategic voice was the craft. That gap between correct and good is where most of the real work lives, and the part no one sees in the finished product.
———
Calibration was one half of that effort. The other half happened in the first days of Sprint 1, before a single output existed, in a decision we could have skipped. The question was simple: when something underneath the system changes, and it always does, how much breaks?
If every skill reached straight into the data, the answer was everything at once. One renamed column on a Tuesday morning, and every part that depended on it would fail together. So we did the slower thing and built a stable layer in between, so that when the ground shifted, only that layer had to move and the skills above it kept working. The same principle shaped the architecture: keep the system flexible enough to evolve without rebuilding everything underneath.
The faster path was right there. Writing each skill directly against the data as it existed in April would have shipped sooner, but the real cost of speed isn’t the first release. It’s paying the same tax every time the ground shifts. That, too, was craft, just the kind that doesn't show up in a demo.
———
Both decisions came from the same place, and it's the thing a "we adopted AI" announcement always leaves out. Quality at speed isn't summoned. It is manufactured, front-loaded with judgment and refined by hand until it holds. The fourteen rounds of calibration and the architectural choice were the same act of building, aimed at a system that could move fast without falling apart or sounding like a machine.
By the end of Sprint 1, the system was real. Every layer was running on real data and producing output that was correct, even if not yet client-ready. We had real parts and no whole.
———
The visible product would come later.
That's the strange thing about building this way: it can feel like progress without a product, right up until it doesn't.
NEXT IN THE SERIES: Twelve Weeks in Three Days: What acceleration actually felt like, and what it didn't.