Same report from AI twice, different numbers. How to get consistent results from AI

Last month the AI made a sales summary and everything added up. This month you paste in similar exports and the same request, and the total comes out different. This time returns were taken off the sales figure, while last time they sat in a separate column. That is usually the moment people start looking for a way to get consistent results from AI.

Nobody made a mistake in the request. The AI worked out the method again from scratch. If you want it to count the same way every month, you have to give it the method instead of hoping it lands on the same one.

Why AI counts differently every time

A request like "make the sales report for September" leaves the AI a lot of decisions. Net or gross. What to do with returns, and where an order placed in September and cancelled in October belongs.

The person who has made this report for years has the answers in their head. The AI does not.

So it makes those decisions anew each time. Sensibly, but not always the same way. On top of that, a language model does not give word for word identical answers, even to an identical question. Two runs of the same report can end up counting different things.

The worst part is that both results look convincing. The table is tidy and the totals match within the sheet. You only see the difference when someone compares one month with the next.

Consistent results from AI start with a written method

The fix is boring. Write the method down once and give it to the AI every time. Net or gross, what happens to returns, what the columns are called, in what order the salespeople appear. That description can live in a few places:

• General instructions in settings. ChatGPT and Claude both let you save instructions that apply to all your conversations. They work well for tone and language. For the method of one report they are too general.
• A project with instructions and files. Both tools let you create a project and add instructions and files to it, and every conversation in the project starts with that context. Put the method there, together with one past report you know is correct.
• Skills in Claude. These are folders of instructions, sometimes with scripts and files, that Claude loads on its own when it decides they fit the task. They are available on the free plan too, once code execution is switched on in settings. Claude can help you write one by asking how you work. ChatGPT has similar skills on its business plans.

For people who write code, a rules file in the project does the same job. In Claude Code it is CLAUDE.md, a file in the project's root folder that Claude Code reads at the start of every session. Coding standards and important decisions go there.

Read the instruction before you trust it

The easiest way is to ask the AI to write the method down after a report that came out right. That is a sensible idea, as long as you read what it wrote. Such an instruction can contain a sentence nobody on your side ever agreed to. For example, that returns are counted separately when your company deducts them.

The AI wrote down what it did, not what it should do. Fix that straight away, before the instruction starts running every month.

An instruction is not a lock either. In the Claude Code documentation Anthropic says plainly that Claude treats these files as context, not as enforced settings. The more specific and concise the instructions, the more consistently Claude follows them. A few clear sentences beat a long description where the important rule gets lost somewhere in the middle.

Test it on last month

Before you stop counting by hand, run a test. Take data from a month where you know the right result, because that report has already gone out and been checked. Run the instruction on it in a new conversation.

If the total and the split by salesperson match, the instruction is ready. If not, find the sentence that causes the difference, fix it and test again. Usually the gap comes from a decision nobody wrote down.

Repeat the same test after every change to the instruction and after a model update, because a new model can read the same words differently.

Start with one report you make every month. Write its method down and check it against last month's data before you hand it over to the AI. If the report should build itself without anyone pasting in exports, that is automation on the side of your system, not a conversation with AI.