Your AI Problem Is a Bookkeeping Problem

The Spreadsheet Came First

I connected my P&L to an AI model and it worked on the first attempt. No troubleshooting, no clever prompting, no second go at it. I have been asked how, and the honest answer disappoints people. It worked because I had spent months building the file by hand before I ever pointed anything clever at it.

That is not the story anyone wants. The story people want is about the tool.

What the AI Actually Asks Me

Here is what happens when I record a new client now. I type one line. A new customer, with a name.

Then it asks me questions. Which product is this. How many sessions. What is the contract value. What is the payment plan.

Those are not questions the model invented. They are the columns of my Sales sheet. It is asking because those fields exist and have to be filled, and it knows they exist because the file told it so. The structure is the instruction set. I did not write a clever prompt. I built a file that could only be filled in one way, and the prompt became simple as a result.

When it finishes, it writes the row and the dashboard updates. That last part looks like the impressive bit. It is the least impressive bit.

Three Sheets That Never Talk Over Each Other

My business keeps sales, revenue and payments in three separate places, and this is the decision I would defend hardest.

A sale is a contract value. Somebody agreed to buy six sessions. That number exists the day they sign and it does not change.

Revenue is what I have actually earned. It lands session by session, on the date the session happened. Six sessions sold is not six sessions delivered, and the gap between those two things is most of what people get wrong about their own business.

A payment is cash arriving in a bank account. It has its own date, and that date has nothing to do with either of the other two.

Most people collapse all three into one column called income and then wonder why the number never matches their bank balance. It never matches because it was never the same number. Keeping them apart is not pedantry. It is the difference between knowing what you sold, what you have earned, and what you have been paid, which are three genuinely different questions a business owner has to answer every month.

Every Sale Has an ID

Each sale in my file gets an identifier. S001, S002, and so on. When a payment comes in, it is logged against that identifier rather than against a name.

This sounds like a trivial housekeeping choice. It is the reason the file can be read by anything other than me.

Names are unreliable. People spell them differently, use a maiden name on the bank transfer, pay through a company, or share a first name with another client. An identifier does not have those problems. When a payment references a sale, there is exactly one sale it can mean.

Nothing about that is an AI idea. It is how any finance system has worked for decades, and I built it in because I did not know another way to work.

The Category List Is Four Lines Long

There is a sheet in my file that does almost nothing. It contains four product names and that is all.

Every other sheet pulls from it. Which means I cannot invent a new product category in a hurry and then spend months afterwards trying to work out whether two differently worded lines were ever the same thing.

Four options, chosen once. If I want a fifth, I add it deliberately and it becomes available everywhere at once.

This is the least glamorous decision in the entire file and it does more work than anything else. Categories that drift are the single most common reason a set of numbers cannot be summarised, by a person or by a machine. You cannot group what was never consistently named.

Why This Is Finance Work, Not Technology Work

None of what I have just described is technical. There is no code in it. A person who has never written a line of anything could build all of it.

But it is finance work, and it is the part of finance that never gets discussed because it is dull. Consistent categories. One place where each figure lives. A separation between what you sold, what you earned and what you were paid. A process that produces the same answer regardless of who is doing it.

I spent over a decade in corporate finance and most of that job was this. Not analysis, not modelling, not the interesting parts. Building numbers other people could trust without having to ask me what they meant. It is the same instinct behind the spreadsheets I used to build that AI now assembles in minutes, except the structure underneath turned out to be the part that lasted.

At the time it felt like the boring foundation under the real work. It turns out it was the real work, and I only found that out when I handed my file to something that could not ask me clarifying questions.

What I See When I Look at Other People's Files

When somebody tells me AI did nothing for their business, I want to see the file before I discuss the tool.

Usually what I find is a business that has never been defined. Categories that changed twice in a year because somebody was in a hurry. The same figure living in three places, each slightly different, none of them agreed. Income and cash treated as the same thing. A process that exists, genuinely and completely, but only inside one person's head.

There is a version of this that looks healthy from the outside. A large top line that seems like proof of a strong business, sitting above costs that were never mapped against it. The gross number reassures. The net number, when it finally gets calculated, is a different conversation entirely. It is the business version of what happens to individuals when the income arrives without the system underneath it.

None of that is an artificial intelligence failing. It was there before, quietly, and it will still be there afterwards.

The Tidying Was the Job

Connecting my file to a model took an afternoon. Building the file took considerably longer than that, and I did it before I had any idea it would be useful for this.

If you are hoping AI will bring order to a business that does not have any, it will not. It will move faster through the same confusion and hand it back to you sooner.

But if you have already done the unglamorous work, if your categories hold and your numbers agree with each other, then the technology has something to stand on. It can see how your business works because you decided how your business works, and you wrote it down.

That decision is available to anyone. It has been available for a long time. It just never had a reason to feel urgent before.

Software cannot organise a business that was never organised. It can only show you, faster than you would like, that it wasn't.