There is a task in finance that nobody celebrates but almost every company struggles with: matching what came in against what was expected. Cross-checking the bank statement, the card processor report, the sales spreadsheet, and the billing system until everything lines up. When it does not, someone spends hours hunting down the discrepancy. That is the bottleneck. And it is exactly what one operation brought to Steply to solve with AI.
This article tells that story without the technical jargon: what the real pain was, why hiring more people did not fix it, what AI actually does here (and what it does not), and how the architecture behind it shifted the game from "close that takes days" to "discrepancy flagged same day." If your finance team gets stuck every month-end, the problem is probably the same.
1. Where the finance team was actually stuck
The operation did not have a shortage of competent people. It had a problem of work that does not scale. Every week, someone from finance exported four different sources: the bank statement, the card processor report, the sales spreadsheet, and the billing system. Then they cross-checked line by line to confirm that every expected payment actually came in, at the right amount and on the right date.
When everything matched, it was just wasted time. When it did not match, and it often did not, the hunt began. An extra fee here, a payment that landed two days late there, a canceled order that nobody wrote off. Each of those discrepancies cost anywhere from fifteen minutes to a full afternoon to trace. At month-end, that turned into two to three days of one person doing nothing else.
The side effect was the worst part: because the close was delayed, decisions were delayed too. The owner only knew the real cash position well after the month had turned. Making purchasing, hiring, or discount decisions based on numbers from three weeks ago means deciding in the dark.
2. Why hiring more people did not fix it
The first reaction of almost every company at this point is the same: bring in another person to help with the reconciliation. It seems logical, and it is the most expensive and least effective solution available.
Financial reconciliation is repetitive work that demands sustained attention, and humans are naturally bad at that. After the hundredth line, the eyes get tired and the discrepancy slips through. Two people checking the same spreadsheet do not double the accuracy. They create rework and conflicting criteria ("did you account for that fee? I did not"). And the cost is permanent: the new hire does not fix the cause, they just absorb volume until the volume grows again.
The bottleneck was not number of hands. It was the fact that the reconciliation was manual and rebuilt from scratch every time. As long as the review depended on someone opening four files and comparing them by eye, the problem came back every month, bigger.
3. What AI actually does here (and what it does not)
Worth clearing up a common misconception right away. In this case, AI does not replace the finance team and does not make money decisions on its own. It does not approve payments, move funds, or decide anything that requires a human signature.
What it does is the tedious, mechanical part: read the four sources, understand that "Pix received R$ 1.240" in the bank is the same event as "order 8842 paid" in the system, and match the two automatically, even when the text, format, and date are not identical. This is where AI beats a traditional spreadsheet rule: a spreadsheet needs an exact match, and real life is never exact. Abbreviated names, amounts with fees deducted, installment payments, shifted dates. AI handles the mess that a rigid rule cannot.
The practical result: instead of one person cross-checking everything, the system does it automatically and delivers only the list of what did not match. The finance team stops searching for a needle in a haystack and only looks at the five or six needles left over. Human judgment stays in place. The digging became automatic.
4. The architecture behind it, without the technical jargon
The difference between "I pasted a ChatGPT into finance" and a solution that holds up day to day is in the architecture. Think of a busy restaurant kitchen: a good cook is not enough, you need flow, stations, and nobody bumping into each other. That is how we built it.
Each source comes in through its own door
Bank, card processor, sales, and billing are not handled manually. Each has a dedicated intake that receives the data as soon as it exists and puts it in a single unified format. Finance stopped exporting and importing files. The data arrives.
Reconciliation runs continuously, not at month-end
Instead of accumulating everything to review all at once, the matching happens all the time, in the background. A payment lands today, it gets reconciled today. When the month turns, the close is already 95% done because the work was spread across the whole month. The day-one crunch disappears.
The AI decides, but within clear limits
The intelligence doing the matching operates with safety rules around it. Above a certain value, or when confidence is low, it does not conclude on its own: it flags for human review. There is a log of everything it decided and why, so nothing is a black box. And there is an off switch: if something looks wrong, finance goes back to manual mode without depending on anyone.
What slips through becomes learning
Every discrepancy that a human resolves goes back into the system as an example. The unusual pattern from today (a new processor fee, a type of reversal) gets recognized automatically next time. The solution gets more accurate with use, unlike the spreadsheet, which stays exactly as big as the problem forever.
5. What changed at the end of the day
The number that matters to the owner is not "how many lines did AI match." It is when they started knowing the real cash position. Before: two to three days after the month turned. After: on the first business day, because the work was spread across the month and only exceptions were left to review.
The finance person was not laid off. They stopped digging and started doing what no machine does: investigating the root cause of the discrepancies that matter and following up with whoever needs to be followed up with. Their work moved up a level. The tedious work disappeared.
And the quiet win, the one nobody puts in the ROI spreadsheet: decisions at the right time. Knowing the real cash position on day 1 instead of day 20 changes how you negotiate with suppliers, when you hire, and how much of a discount you can offer without hurting yourself.
The reframe
The most common misreading is thinking this is "a robot doing math." It is not. The bottleneck was never the math. It was the manual work of matching sources that do not talk to each other. AI solved the part it does better than people (repetition and approximate matching at volume), and the architecture ensured it ran every day, safely and without becoming a black box.
If your close still takes days and decisions always arrive late, the problem is probably the same as this operation's. And the solution is not one more person checking a spreadsheet. It is stopping the spreadsheet-checking altogether.