Demo · Simulated data

The month is closed. Time to review the numbers.

Every month someone has to read through the close and check that the numbers add up for Kova, a made-up web shop. The machine has hidden a couple of errors in the numbers, the kind that eat the margin. You can look for them yourself first. Then you unleash an AI team that reviews right in front of you, with real model calls. You can check everything against the answer key afterward. The process is real even though the numbers are generated.

Comic panel: Torbjörn circles a gold-highlighted row in a giant spreadsheet, coffee in hand, while three small gnome helpers pitch in, one on a ladder with a magnifying glass, one on a stack of binders, and one stamping the report at a table.

The month is closed, but there are variances in the numbers.

How it starts

A made-up monthly close for the Kova web shop, with two or three errors hidden in the numbers

Click the button and a unique month is built from a random seed, with budget vs actual, supplier invoices, and six months of history. A couple of errors are hidden in the numbers. It is all made up, nothing is saved.

What you just saw

The code calculated, the AI read and judged, the human decides

01

The actual math was done by plain program code. Click a variance and the formula behind it appears. The AI reads the numbers and judges them, it never does the math itself.

02

The AI team read, judged, and wrote in real time. Three specialists each with their own focus area, one at a time so you can follow along, and an editor who weighs the findings and asks a follow-up question.

03

The human decides. The report is material to check and act on, not a decision in itself. That is why you got the answer key and the option to download a CSV yourself.

Under the hood · the same principle as in the real build

A simplified version of a way of working you can run on real numbers

  1. The only input

    A random seed

    The client sends only a random seed. The server rebuilds the exact same month itself, so there is no free text to manipulate.

  2. Code, not AI

    The math is calculated

    Budget, actuals, variances, and the answer key are calculated deterministically. The math is exact and checkable, every formula can be clicked open.

  3. Haiku 4.5 × 3

    Three specialists, one at a time

    One data block each, real calls. They could run in parallel, but they take turns here so the review is possible to follow. The answer key is never sent along, the team reviews blind.

  4. Sonnet 4.6 · two phases

    The Editor weighs in

    Reads the findings and writes the report with actions at both the row and routine level.

    ↺ asks a follow-up question first
  5. Your receipt

    Report, answer key, and CSV

    Honest grading even when the team misses, and a spreadsheet with live formulas so you can check the math yourself in Excel.

The scanning light and the spinning digits are presentation. The findings, the question, the answer, and the report are the models’.Nothing is stored. Your marks stay in your browser.Role names in the demo, model names here and in the printout. Nothing in between.Every run starts over from zero. A production build would keep the judgement calls, so next month starts smarter than this one.The calls go to Anthropic’s API, which is fine on made-up data. On real records the reading would run locally, and only de-identified questions of principle would go further.

This is a demo, not a production system. The numbers, the shop, and the suppliers are made up, but the review you saw was done for real while you waited. The point is the way of working. The code calculates, the AI reads and suggests, the human decides. The whole demo was built in a day by a marketer who works the same way with real numbers, ad costs, margins, and broken tracking.

Read how it was built

The whole build is documented, including the process review that tore up my first design.

Want someone on your team who knows how to do this?

Get in touch