Challenge: the month’s number lived scattered across six sources (dining room, iFood, bank, card, payroll, reconciliations), each in its own format. Pulling it together and categorizing line by line cost a full day, and the close was only ready around the 10th.
Solution: one collector per source, fetching on its own and dropping into the right spot of the P&L, with AI categorizing the statement and card on my own criteria, anchored on history. Before writing anything, it shows me the result to check.
Results: the lost day became run and check, and the close for two companies is ready in the very start of the next month, early enough to correct course while there’s still time.
The month’s number doesn’t live in one place
A question for any owner: how much was left last month, and why? Most can’t answer on the spot. For me, for a long time, that number was only ready around the 10th, after a full day pulling everything together by hand. A close that’s only ready once the month is over is useless for deciding anything. If it went badly, you find out when there’s nothing left to react to.
And there’s a second trap: pulling it together isn’t even the worst part. The real work is categorizing. Every expense on the statement and every line on the card has to become a category, or the total tells you nothing. “This much went out” is useless. “This much on ingredients, this much on staff, this much on marketplace fees” is a decision. It’s that classification, line by line, that eats the afternoon.
I put AI to repeat my criteria, not to guess
The obvious route would be hiring someone to type this in every month, or subscribing to a management system that promises to integrate everything. But the first is paying for manual work that fails silently, and the second hits the same wall: my six sources don’t speak the same language, and no off-the-shelf system knows my categorization criteria.
So each source got a collector that fetches on its own and drops into the right spot of the P&L. What sold, it pulls straight from the POS and iFood. What I spent, it reads from the statement and the card invoices. Nobody types anything.
And the part that was hell, categorizing, is where AI comes in: it looks at my history from previous months and my reference table, and classifies each entry the way I would classify it. It isn’t guessing, it’s repeating my criteria, fast.
I stopped driving by the rearview mirror
In practice, it works like this: I run it, the AI categorizes everything and shows me the result before writing anything. I trust, but I verify. On a number that becomes a decision, AI that writes on its own without me seeing it is an invitation to a costly mistake. The lost day became run and check.
But time isn’t what matters. What changed was when I see the month: the close for two companies is ready in the very start of the next month, early enough for me to correct course while there’s still time. Three years ago this started small, with me uploading a loose spreadsheet to AI and asking it to find what I wasn’t seeing. I told that beginning here.
A number that’s ready on the 10th is history. A number that’s ready on the 2nd is a decision.
Building a controllership that pulls the sources together and categorizes on your own criteria, without you losing control of the number, is exactly the kind of project I help companies build.
Stack: one collector per source (POS and iFood pulled via browser, statement and card read straight from the bank’s files), AI categorization anchored on history and a reference table, everything written into a single P&L spreadsheet, with a mandatory review step before any write.