Analysing numbers and tables with AI without ending up with nonsense

At some point you have a file in front of you that you really ought to analyse. The revenue export from your accounting tool, the customer list, last quarter's timesheets. Four hundred rows, and all you want to know is which five customers bring in the most and whether revenue really does collapse over the summer.

That's exactly why almost everyone now uploads their table to an AI. It works, too, but there's a catch that hardly anyone explains: an AI can process a table in two completely different ways. Either it reads the numbers as text and estimates the result, or it writes a small program in the background that genuinely computes the numbers. The first is convenient and produces sums that are off by a few percent. The second is reliable.

The difference isn't visible from the outside. Both answers look equally assured. These ten steps make sure you get the second kind.

Step 1, formulate the question before you upload the file

The reflex is to drag the file in and write "analyse this for me". Then you get a general description of the table, which tells you nothing, because you already know your own table.

Write down beforehand which two or three questions you really want answered. "Which ten customers together make up what percentage of revenue", "in which months is revenue below average", "on which line items is the margin under twenty percent". Concrete questions get concrete numbers, vague briefs get prose.

Step 2, tidy the file up first

Most errors don't happen during the calculation, they happen while the file is being read. Four things reliably throw an analysis off.

Several header rows stacked on top of each other, merged cells, subtotals in the middle of the data and empty separator rows between blocks. The subtotals in particular are nasty, because they get counted twice, and the grand total then looks clean and is still wrong. So pull a copy, throw the subtotals out, make sure there's exactly one header row and let every row be one record. Five minutes of tidying save you half the checking.

Step 3, take out what doesn't belong in an upload

Before the file goes anywhere: go through the columns and delete everything that isn't needed for your question. Names, email addresses, phone numbers, addresses, dates of birth.

For the question "which customers bring in the most revenue" you don't need real names. A customer number is entirely enough, and you can map it back yourself at the end. This isn't an over-the-top data protection reflex, it's the simplest way to sidestep the topic completely. If you work with customer data regularly, Data protection with AI tools is the groundwork for it.

Step 4, explain the columns

A column called "Betr_N" is the net amount to you and a riddle to the AI. It then guesses, and usually it guesses right, but not always.

Write two or three lines about it. Which column is the revenue, is it gross or net, in which currency, in what format is the date, what do the status codes mean. Especially important with German exports: say explicitly that the comma is the decimal separator and the full stop separates the thousands. If that gets read wrong, 1.250,00 euros quickly turns into a completely different value, and then the total doesn't add up at either end.

Step 5, force calculating instead of estimating

This is the heart of the matter. Write into the brief explicitly that the analysis is to be calculated and not estimated. One phrasing that reliably works: "Analyse the file by actually computing it, not by estimating the values from the text. Show me the calculation."

With tools that have an execution environment for code, this triggers a small analysis program being written and run. The result is then a real calculation, not an estimate. You can tell because you get shown code or a computation step and not just a result. If plain prose with round numbers comes back, it was estimated. Then you ask again.

Step 6, demand a sample you can recompute yourself

Do you trust the analysis? Not yet. Have the underlying rows shown to you for one result. So not just "customer 4711 has 18,400 euros in revenue", but the list of individual items that were added up to get there.

You recompute that one item by hand, with a spreadsheet or a calculator. If it matches, you can trust the rest quite far, because the same calculation ran for every row. If it doesn't match, you've found out in two minutes that the whole analysis is useless. Those two minutes are the best investment in the entire process.

Step 7, ask about what's missing

The most interesting things are often not in the analysis but in the gaps. So actively ask the counter-question: "Which rows have empty mandatory fields, which entries are duplicates, which values are outliers and which rows could you not assign?"

An export almost always has a few broken rows. If twelve of them got silently skipped, your grand total is too low and nobody said so. So always ask how many rows actually went into the calculation, and compare that with the row count of your file.

Step 8, ask for tables explicitly

A known pattern with these tools is that they'll narrate a question whose answer is really a table. You then get three paragraphs of prose about revenue development, out of which you have to dig the numbers again.

Just say so. "Result as a table with the columns month, revenue, change from previous month in percent." If you want to carry on working with it, ask for the result as CSV on top, then you can tip it straight back into your spreadsheet instead of typing it out.

Step 9, get the formula, not just the result

If you need an analysis regularly, the one-off answer is the worse outcome. Ask for the method instead: "Give me the formula for Excel or Google Sheets that lets me calculate this myself."

Then you have a column in your own table that is correct again immediately with the next export, without you having to upload anything. The AI took the thinking off your hands once, and after that your table does the calculating again. That's almost always the better solution than an analysis you have to beg for anew every month.

Step 10, decide what you repeat

When an analysis has worked well, save the complete brief, so the column explanation from step 4, the compute-don't-estimate sentence from step 5 and the desired output format from step 8. In a notes app, as a text snippet, anywhere.

With the next export the whole process then comes down to uploading the file, pasting in the brief, recomputing one sample. Three minutes instead of half an hour. And because the brief is the same every time, the results are comparable across the months as well, which in my experience is not the case with questions you phrase freshly each time.

What's next

Which of the big tools can even read files and run code is in Tools in AI in Level 2. If an analysis becomes the basis for a decision or a quote, run it through Fact-checking an AI answer as well. And when you notice that you're doing the same analysis every month, that's the point where automation pays off, and for that My first n8n flow for non-coders is the simplest way in.