Why Our AI Never Does the Math

Ask a language model to add up a column of numbers and it will usually give you an answer. It will sound confident. Sometimes it will be wrong, and nothing in the answer will tell you which time that was.

For most uses that is a minor problem. For a finance team, it is the whole problem. A number in a board report has to be right, and it has to be right the same way every time someone runs it.

So we follow one rule in every finance system we build: code computes, the AI explains.

What that rule means in practice

Every figure that matters (revenue, expenses, balances, variances, forecasts) is calculated by deterministic code. The same file goes in, the same numbers come out, every run. Headline totals are summed from the detail underneath them, never copied from text a model wrote.

The AI model gets the finished numbers and does what it is good at: it writes the summary, points out what changed, and puts the results in plain language for the person reading the report.

If the model's explanation and the computed numbers ever disagree, the computed numbers win, and the report says so.

A real example of why this matters

While building our own finance reporting tool, we found a bug in how one kind of row was being read. Cost rows that were entered as positive numbers were being counted as revenue. On the affected data, reported revenue was overstated by about 20 percent.

It had gone unnoticed because the sample data used in testing happened to follow a convention that hid it. The fix was simple once it was found. What found it was not a model reading the output and deciding it looked off. It was deterministic checks comparing totals that should have agreed and did not.

That bug is now covered by tests that run on every change. That is the kind of safeguard you can only build when the math lives in code you can inspect.

Why this matters when you choose a vendor

Plenty of tools now promise AI for finance. Some of them let the model produce the numbers directly. That can look impressive in a demo and still fail quietly on your data.

Before you trust any AI finance product, ask three questions:

1. Where are the numbers calculated? In code you can audit, or by the model?

2. If I run the same file twice, do I get the same totals? If the answer is "usually," keep looking.

3. What happens when data is missing? A good system tells you what it could not compute. A bad one fills the gap and moves on.

The short version

AI is very good at explaining financial results. It should not be the thing producing them. Keep the math in code, keep the explanation in the model, and you get reports that are both readable and correct.

If you are evaluating AI for your finance team and want a second opinion on how a tool handles its numbers, email [email protected].