Find out that an expense spiked before you pay the bill

Who it is for

Whoever pays the bills at a small business: the owner, the administrator or the person who handles the finances, with an expense spreadsheet by supplier or category and no time to review it line by line.

The problem

Expenses get reviewed after they are already paid. A service that doubled its price, a subscription nobody uses or a supplier that invoiced twice slip through among hundreds of lines identical to last month's.

Ingredients

• Your expense spreadsheet for the last six months, with columns for supplier or category, month and amount (exported from your bank, your accounting software or built by hand)
• An account on claude.ai or ChatGPT that accepts file attachments
• Ten minutes to decide which threshold matters to you: for example, flag anything that rises more than 40%

Step by step

1. 1

   Arrange the spreadsheet in long format

   One row per expense: supplier or category, month, amount. If your bank gives you a statement, first ask the model to convert it to that format. Without this, every comparison comes out skewed.

2. 2

   Define the expected value and the threshold

   The expected value is the average of the previous months for that same line. The threshold is how much it can move before it gets flagged. According to the source, BMW uses 40% as a general rule and 60% for services that naturally fluctuate more. Start with 40% and adjust.

3. 3

   Run the prompt with the spreadsheet attached

   Attach the file and paste this recipe's prompt with your data. Ask for the anomaly table and the likely explanation for each one. The first time, check that the calculations match two or three lines you know well.

4. 4

   Decide and record

   Mark each anomaly as "normal", "review" or "dispute". Normal items (a seasonal purchase, an advance payment) go in as known exceptions for the following month: that is how the list shrinks over time.

5. 5

   Repeat every month with the same prompt

   Save the prompt and the spreadsheet in a Project. The following month you only add the new rows and the exceptions you already know. Fifteen minutes.

Replace what is in brackets with your company details, or use the guided mode above: it asks you and fills it in for you.

# Role
You are the cost analyst at [COMPANY NAME], a [INDUSTRY] company in Chile. You work with spreadsheets, you are skeptical and you do not invent explanations the data does not support.

# Context
Attached is a spreadsheet with the expenses of the last [NUMBER OF MONTHS] months, one row per expense, with the columns: supplier or category, month and amount in Chilean pesos. We already know these lines fluctuate and they should not be reported except for huge changes: [LINES THAT ALWAYS FLUCTUATE, OR "NONE"].

# Task
1. For each supplier or category, calculate the expected spend for the last month as the average of the previous months.
2. Flag as an anomaly every line whose last month deviates more than [THRESHOLD, FOR EXAMPLE 40]% from the expected value, upward or downward.
3. For each anomaly, propose the most likely explanation given the available data and the question an administrator would ask to confirm it.

# Output format
A table with: supplier or category · expected · last month · variation in % · likely explanation · confirmation question. Sort from largest to smallest variation. Below it, one line with the total number of lines reviewed and how many were flagged. If something cannot be calculated for lack of data, say so instead of estimating it.

# Examples
Input: "Internet, March 45,000; April 45,000; May 89,000".
Expected output: "Internet · 45,000 · 89,000 · +98% · likely plan change or two months billed at once · was there a plan change or a late invoice in May?".

What usually goes wrong

• ✕Comparing monthly totals instead of line by line. A total can look normal while one line doubled and another disappeared.
• ✕Asking the model to "review the expenses" without a threshold. You get general comments back; with a numeric threshold you get a short, actionable list.
• ✕Not recording the known exceptions. Every month it flags the same seasonal purchase again and you end up ignoring the whole list.