Insights

Insights · AI that decides

Should you hand the whole table to an AI and ask for the analysis?

The short answer is no, and the reason is not what the AI can do. A sales table doesn't hold "the October number": it holds several, all of them correct. If your company doesn't choose which one counts, the AI chooses, and doesn't tell you.

16 Nov 2026 · 5 min read

The same table, four numbers

The request sounds much the same in many AI projects: "let's plug the AI into the database and it will generate the month's analysis". It looks like the most natural use of the technology. The problem shows up at the first simple question.

Illustrative scenario

A shop has 13 orders on record between September and November. Management asks how much the shop sold in October. With the same table, the answer can be:

  • €4,200, adding up everything ordered in October;
  • €3,600, adding up what was delivered in October;
  • €2,800, by order, deducting what was returned;
  • €2,900, by delivery, deducting what was returned.

And against September? By one of these sums, October went up 71.43%. By another, it went down 3.45%. The data is made up and the table is small on purpose, so that anyone can check it by hand. None of the four sums contains a mistake.

The difference doesn't come from arithmetic. It comes from two choices every company makes without noticing: what counts as a sale (what went out, or what went out and didn't come back) and on which date the sale counts (when the customer ordered, or when they received it). People who know the business make those choices in their heads. The AI doesn't know your business.

What happens when the whole table goes to the AI

It answers, and that is the risk. The answer arrives in seconds, well written, with a number and an explanation. To get to the number, the AI had to pick one of the definitions, and the choice doesn't show in the text. At the next meeting, with the question worded differently, the choice may be another.

The report that goes up to management now has three problems at once: nobody knows which definition was used, nobody can redo the sum, and the text is convincing enough that nobody asks. "October grew 40%" and "October fell 3.45%" come out of the same table and lead to opposite decisions on stock, targets and campaigns.

There is also a technical reason. Researchers have shown that language models make worse use of information in the middle of a long text than of information at the beginning or the end. To the AI, a table with thousands of rows is a long text.

What usually goes wrong

  • Dropping the spreadsheet into the chat and asking for "insights". It is the fastest way to get good text with a number nobody owns.
  • Approving the project without asking who does the calculating. The question the sponsor needs to ask before signing is simple: who calculates the numbers, and who checks them?
  • Assuming an AI that runs code solves it. It adds correctly. But adds what? Without a written definition, the program the AI writes carries a guess, now looking like an exact number.
  • Leaving the definition in one person's head. As long as "sales" means one thing to the commercial team and another to finance, and that isn't written down, each report will use one of them.
  • Comparing without agreeing on the period. Against the previous month or against the same month last year? The AI chooses that on its own too.

How to decide, in four steps

  1. Write down the definition of each metric the AI will use: what goes in, what stays out and on which date it counts. One sentence per metric.
  2. Agree on the comparison period: this month against which one.
  3. Calculate first, outside the AI. The data team computes the metrics in code that runs the same way every month and can be tested. The AI receives a small table, with the numbers ready and the definition of each one.
  4. Name the person who checks. Someone signs off the numbers before they go up, and checks that every number in the AI's text is in the table it received.

Once that is done, the AI does what justifies the investment: comparing, explaining and writing on top of numbers someone calculated and checked. The test below asks three questions and shows what your company is missing to get there.

Teste rápido

A sua tabela está pronta para a IA?

Responda sim ou não às três perguntas. Cada não mostra o que fazer antes de entregar a tabela à IA.

  1. As métricas que a IA vai usar têm definição escrita?Se a resposta for não: escreva a definição de cada métrica: o que entra, o que sai e em que data conta.
  2. Existe um período de comparação combinado (este mês contra qual)?Se a resposta for não: combine o período de comparação: contra o mês anterior ou contra o mesmo mês do ano passado.
  3. Alguém confere os números antes de eles subirem para a diretoria?Se a resposta for não: escolha quem confere os números antes de eles subirem.

Como ler o resultado

  • Três sim: pronta. Dá para entregar a tabela de grandezas à IA.
  • Sim na primeira e não em alguma das outras: quase. Resolva o que falta antes de entregar.
  • Não na primeira: ainda não. Sem definição escrita, a IA vai inventar a definição, mesmo que o resto esteja em ordem.

Quick test

Is your table ready for AI?

Answer yes or no to the three questions. Each no shows what to do before handing the table to the AI.

  1. Do the metrics the AI will use have a written definition?If the answer is no: write down the definition of each metric: what goes in, what stays out and on which date it counts.
  2. Is there an agreed comparison period (this month against which one)?If the answer is no: agree on the comparison period: against the previous month or against the same month last year.
  3. Does someone check the numbers before they go up to management?If the answer is no: choose who checks the numbers before they go up.

How to read the result

  • Three yeses: ready. You can hand the table of quantities to the AI.
  • Yes to the first and no to either of the others: almost. Sort out what is missing before handing it over.
  • No to the first: not yet. Without a written definition, the AI will make up the definition, even if everything else is in order.

Test rapido

La vostra tabella è pronta per l'IA?

Rispondete sì o no alle tre domande. Ogni no indica che cosa fare prima di consegnare la tabella all'IA.

  1. Le metriche che l'IA userà hanno una definizione scritta?Se la risposta è no: scrivete la definizione di ogni metrica: che cosa entra, che cosa resta fuori e in quale data conta.
  2. Esiste un periodo di confronto concordato (questo mese rispetto a quale)?Se la risposta è no: concordate il periodo di confronto: rispetto al mese precedente o allo stesso mese dell'anno scorso.
  3. Qualcuno verifica i numeri prima che arrivino alla direzione?Se la risposta è no: scegliete chi verifica i numeri prima che salgano.

Come leggere il risultato

  • Tre sì: pronta. Potete consegnare all'IA la tabella delle grandezze.
  • Sì alla prima e no a una delle altre: quasi. Sistemate quello che manca prima di consegnare.
  • No alla prima: non ancora. Senza una definizione scritta, l'IA inventerà la definizione, anche se il resto è in ordine.
Next step

The test gives the direction. In a short diagnostic, we do it with your data: we list the metrics the AI will use, write down the definition of each one with the people who know the business, show where the same question gives different numbers today and say what to calculate before plugging in the AI.

How to use this in your company

The same order applies wherever there is a table and a business question:

  • Monthly sales report. Gross or net sales, by order date, delivery date or invoice date: choose, write it down and calculate first. The AI drafts the month's commentary.
  • Customer service indicators. What counts as a resolved ticket: one that was closed, or one that was closed and not reopened? The definition comes before the dashboard and before the AI.
  • Customer base. Did an "active customer" buy in the last 30 days or the last 90? The answer changes the size of the base and every analysis built on it.
  • Finance. Revenue in the month the sale was made or in the month the money came in. Both views exist for good reason; the report has to say which one it is showing.

In all of them, the deciding question is the same: does this number have a written definition, and does someone check it?

This applies to your company if…
  • the same question ("how much did we sell?") has already received different numbers from different departments;
  • the definitions of the metrics are in the head of whoever builds the report, not in a document;
  • the AI project plan says "the AI analyses the database" without saying who does the calculating;
  • reports compare periods without saying against which one;
  • nobody checks the numbers in AI-generated text before it circulates.

In the Lab NotebookWant the table of 13 orders, the full sum and the exercise to do with your own hands?First the quantities, then the AI

Read the technical version →

References

← all insights