I had the AI read 8 years of books in one day — the structure of the business showed up in numbers

2026-06-11 ／ Vol 22 ／ draft at the time of publishing

In Vol 11 I wrote that month-end close had quietly become "just verifying". That was about putting entries into the books every month. This is the sequel: reading the books that have piled up.

It started with a simple question.
"Since founding in 2019, what has this business actually run on? Can I say it as a structure?"
The monthly input works. The monthly close closes itself. And yet I couldn't answer that question on the spot.

The material was all there. I just hadn't read it

As a by-product of the monthly automation, the annual reports from my accounting software (the balance-transition tables) for 2019–2026 — 8 full years — were already sitting in a designated folder. Balance sheet and income statement, with monthly columns on the P&L side.
In other words, the material had been there all along. I had simply never done the work of lining it up side by side and reading it.

I handed the whole set to the AI (Claude) and analyzed it from multiple angles in a single day. Annual trends, by channel, by product, break-even point. This is a one-person EC business in the tens-of-millions-of-yen range, and getting to where I could describe its inner workings as a structure took exactly one day.

Even a Power Query person couldn't do this in a day

As I confessed in Vol 11, I'm reasonably confident pushing CSVs around with Power Query and pivot tables.
Even so, 8 years × multiple angles in one day was out of reach. A query for annual trends, another to split by channel, another for per-product profit, then the break-even math. Every new angle means rebuilding the query, and the work keeps stacking up. It would have eaten several weekends. Which is exactly why I hadn't done it for 8 years.

Analysis with the AI felt different. Each time one aggregation landed, the next angle came as a proposal from the other side. See the annual trend, and it suggests splitting by channel. See the channels, and it asks about per-product profit. See the products, and it offers to compute the break-even point.
My job was only to accept or reject the direction, and to verify the numbers that came out. The cost of the digging itself had dropped to almost zero.

Four structural findings

I'm deliberately keeping the exact numbers vague here — the structure is the subject.

1) The bulk of unit volume is thin-margin items. Most of what ships out the door is razor blade refills, with a very thin per-unit margin. I carry higher-margin products too, but in unit terms they barely register. Half of the answer to "what does this business run on" was right here.

2) The break-even point became a number for the first time. From the business-wide contribution margin ratio and the fixed costs, the break-even revenue fell out of a single formula. Eight years in, it was the first time I saw my own break-even point as an actual number.

3) Fee structures differ by channel. Selling the same product, the fee rates differ widely from mall to mall, and the contribution margin ratio ends up clearly apart. "Where you sell" turned out to matter about as much as "what you sell."

4) Income outside of sales is part of the structure. Income that lands under "miscellaneous" in the books turned out to be a piece holding up the thin retail margins. My gut had filed it under "nice bonus"; structurally, it was bigger than I thought.

The experience of "sort of knew" becoming a confirmed number

Honestly, all four were things I "sort of knew." That the business runs on thin-margin razor blades, that some channels carry heavy fees — I knew it as a feeling.
But a feeling and a confirmed number are different things.

With only a feeling, the next move becomes "I'm vaguely anxious, so let's do something." With confirmed numbers, priorities line up. In a structure like this, what moves the needle is shifting toward higher-margin products and rethinking the channel mix. Conversely, the ratios also tell you that shaving fixed costs here and there won't change the structure.
It's still at the level of direction, but knowing "which one to think about first" is a big deal. Only once your gut feeling has been answer-checked can you pick the next move.

The premise: AI aggregation gets it wrong once

One important premise, stated honestly. As I wrote in Vol 11, I proceeded on the assumption that the AI's aggregation will be wrong at least once.

So this time, too, I set things up so the answers could be checked. The annual revenue and profit figures are already fixed as "correct answers" in the accounting software's reports, so the first step was to reconcile the AI's aggregation against those known values. Only after they matched did we move on to the deeper angles (by channel, by product).
If you hold the correct answers while you delegate, AI aggregation becomes a powerful first-draft generator — Vol 11's conclusion held up unchanged at the 8-year scale.

The lesson: books become "structure" only when you line them up

Keep the books every month. Close the month. I had been doing that for 8 straight years. But that was a record, not a structure.

The record was there every month. The structure only appeared once 8 years were lined up side by side.
What changed wasn't the books — it's that the cost of "lining them up" dropped to almost zero.

I suspect a lot of people are in the same place: monthly records quietly piling up. The material may already be in your hands. If the only thing stopping you from reading it was the weekends it would cost to line it all up — that barrier is mostly gone now.