Adding Zeroes

By · · 4 min read

Why Monday's number is different on Wednesday

The number you reported on Monday is different on Wednesday. Nobody touched the data. The definition moved.

That is the usual story behind a drifting report, and it is easy to miss because every version looks right on its own. Monday's "net sales" counted orders by the day they were placed. Wednesday's counted them by the day they shipped. Both are defensible. Neither is the same number.

Three things make a number move

First, a definition changed without anyone deciding it. "Active customer" meant bought in the last 90 days when someone built the report, and someone later tidied it to 120. A returned order was netted out in one tab and left in another.

Second, late data arrived. Refunds, returns, chargebacks and orders from a marketplace that settles a few days behind all land after your first pull. The number was right when you read it and wrong a day later, through no fault of the person who pulled it.

Third, something broke. A connection between two systems dropped a day, or doubled one. This is the real data error, and the first two are easy to mistake for it.

The reason to separate them is that the fix is different for each. A moved definition needs a written definition. Late data needs a stated cut-off. A broken connection needs a repair. If you freeze your definitions and the drift continues, you were chasing the wrong one of the three, and you will not know which until you check.

The old way

In the older days you would keep last week's report, pull this week's, and put the two side by side in a spreadsheet. Then you would go line by line, ask who changed what, open the export behind each number that moved, and guess. Four hours later you have your answer, if the person who knows the answer was in that day.

Today, you hand both versions to Claude, an AI tool you type to the way you would brief a fast assistant, and have your answer in about thirty seconds. (A prompt is just the instructions you type to the AI. You can paste this one as it is, then paste the two reports under it.)

Below are two versions of the same report for the same dates. The first was sent on Monday. The second was pulled today.

For every number that differs, tell me:
1. The number in each version and the size of the change.
2. The most likely cause, chosen from these four:
   (a) a definition changed, such as what counts as a sale, a customer or a return
   (b) late data arrived, such as refunds, returns or payments that settle days after the order
   (c) a data feed broke or doubled a day
   (d) I cannot tell from what you gave me
3. The evidence for your choice, in one sentence.

For every (d), give me the one question I should ask, and who is most likely to know.

Put the result in a table, biggest change first. After the table, list every definition you had to assume to compare the two versions, one line each, so I can confirm or correct them.

First version:

Second version:

The table does most of the work. A row marked (c) goes to whoever looks after your systems that day. A row marked (a) goes into a one-page sheet of definitions, and you can ask Claude to write that sheet from its own list of assumptions.

Make it run before the meeting

The check is worth most when it runs before anyone reads the report. Whoever looks after your computers can set it to run at 6 AM every morning and put the table in your inbox, so you see the drift before your team does. Ask for it plainly: "Run this check each morning on the newest report against the last one sent, and email me the table."

Where this does not hold

If the drift comes from a broken connection, a definitions sheet will not stop it. That is why the check sorts the causes instead of assuming the first one. It is also why you should read the (c) and (d) rows yourself the first few weeks. Claude is working from two documents and cannot see the system underneath, so a call marked (c) is a lead to check, not a verdict.

The other limit is that Claude will tell you a definition probably changed. It cannot tell you which one you meant. That decision stays with you, and it takes a minute once the question is on the table.

CPG AI Assessment

Fix the CPG workflow that keeps landing back on your desk.

I map the recurring reporting or operating workflow first, then build only when the first useful fix is clear.