Every owner has 3 to 5 spreadsheets that eat the night. Sales comparison, daily cash close, product list nobody updates.
Each one stalls a decision.
GPT-4 got file upload. Drop the XLS directly. Drop the PDF report. Drop the bank statement.
I run two pizzerias and have three owner-level tasks I do hands-on. Those three were the first to migrate.
Formulas and spreadsheet analysis
I’ve always used a well-structured CRM and Excel. The bottleneck wasn’t organization. The data was there, but extracting meaning depended on me.
When I can’t remember the formula to compare the same weekday against 4 weeks ago, I ask. The answer comes back as a ready-made Sheets formula, with the why behind it. What used to require research or an available analyst now arrives in seconds, inside the context of the spreadsheet I have open.
The data always existed. What changed was the speed of getting to the right question.
Sales comparison and projections
This is the one hitting me hardest.
I was already using GPT-4 to analyze numbers. But with file upload, the process changed. I drop in the weekly XLS and ask: “tell me what changed and what likely explains it.” Then I ask for a projection based on history.
The analysis isn’t perfect. But it’s 80% of what a junior analyst would deliver, in 30 seconds, with no meeting and no cost.
And this is where something I suspected is finally clicking:
The AI gain isn’t doing what I already do faster. It’s showing me what I wasn’t seeing.
Cash reconciliation
This was the heaviest process. An Excel manually consolidated data from three sources: iFood, direct delivery, and SAIPOS. On top of that, I’d pull bank statements and card sales results from each operator by hand.
Too many sources. Too much double-checking. Too much room for human error.
With GPT-4, I automated that crossover. I upload the files, describe each source’s structure, and ask for consolidation. It points out discrepancies before I have to go looking: “this card value is outside the pattern of the last 4 Fridays, worth checking” or “this category disappeared compared to last week.”
It doesn’t close the cash for me. It points where to look before I close it. Which is, in the end, what a good analyst does.
What I’ve learned so far
GPT-4 reads data, cross-references information, and returns hypotheses. Whoever uses it just to generate text is leaving the best part on the table.
What really changed with file upload is the zero-friction part: I don’t need to clean the data before asking. I drop the file as-is and start. Spreadsheet from three different systems, bank statement straight from the bank, card results per operator. It all goes in.
And the lesson that matters most: the ROI of these three uses together isn’t time savings.
It’s mental space. That effect doesn’t fit in a metric, but it’s what lets the owner sleep.
Three years later, the same principle became the foundation of the Digital Cortex: instead of uploading the same files every week, keep the context persistent, readable by any model. And an agent wired to my WhatsApp started generating most of the hypotheses on its own, without me having to open anything.