How to analyze product profitability with AI

TLDR: QuickBooks Online wins on US ecosystem depth, Xero provides balanced reporting and Digits pushes AI-led analysis furthest. Data structure and traceability matter more than generated prose.

The logo on your homepage

They are the account everybody names when asked who you work with. They also negotiated hard on price three years ago, they send more support requests than any other client, and their work always lands the week something else is due. Nobody has ever counted the hours. When somebody finally does, the flagship account turns out to earn less per hour than the small client nobody mentions.

The argument that follows runs for three meetings and settles nothing, and it is not about the numbers. It is that nobody ever wrote down how shared costs get allocated, so everyone defends the split that flatters their side and all of them are right under a different rule.

That is the job here. Not producing margin by line, which any product does. Producing it in a way that lets the conversation afterward actually close.

Pick the unit of profitability and stay with it

Product, service, project or customer. Exactly one, the one that matches the decision you are about to make. If you are deciding prices, the unit is the product. If you are deciding who to stop serving, the unit is the customer. Mixing both in one analysis produces a report that answers neither question.

Tag revenue and direct costs against that dimension from the first day of the pilot. One well-tagged month is worth more than a year of approximate data, because tags cannot be reconstructed backwards with any confidence. Nobody remembers in November which project the materials bought in March belonged to.

Declare the allocation rule before you look at the result

QuickBooks Online starts at $85, Xero at $25 and Digits at $65, with reporting depth varying by tier. When testing, check whether the product carries your tagging dimension into reports, whether it keeps filters, and whether you can go from a margin figure to the entries that make it up. A margin nobody can open is a number nobody believes by the second meeting.

The four costs that never show up

The owner’s time. Usually the largest cost in the business and almost never on any line, because it does not arrive as payroll. Until you charge it somewhere, every line looks profitable and you know nothing. This does not need accounting precision: an honest estimate of hours per line, made once and revisited yearly, moves the result more than any other adjustment.

Post-sale hours on the big account. Large revenue, a hard-won discount and a support request every week. Per invoice they are your best customer. Per hour they can be your worst, and they are the one nobody wants to put in the table. If you measure them, measure them under the same rule as everybody else.

The line that carries the others. The entry-level product with thin margin that half your customers arrive through. Cutting it because the analysis says it loses money is the classic mistake in this exercise. Before deciding, check what share of your good lines’ customers came in that way.

Dead stock and rework. Neither shows up as a cost of any line, because in the books they sit somewhere else, but they get consumed there. A line with heavy returns or a lot of tied-up material has a real cost its accounting margin does not show.

Ask for causes, not paragraphs

A useful answer does not say costs went up. It names the category, the period, the amount and the transactions responsible, and it lets you open them. If what comes back is well-written prose with no link into the ledger, that is not financial analysis, it is copywriting.

Test on two lines at once, one performing and one struggling, and ask the same question about both. A system that can only explain the good one is describing rather than analyzing.

And do not paste a full ledger export into a general assistant to get prose. You lose permissions, accounting context and the ability to return to the source, which is the only thing making the analysis worth anything. Reasoning over live traceable data is what the financial reports guide develops.

Finish with a decision and a date

An analysis that ends in a report has not ended. Pick one of three: change a price, change the catalog or renegotiate a cost. Before executing it, write down which number has to move and when you will look.

Then measure two things. The margin on that line before and after, under the same allocation rule, which is the part everyone forgets. And how many lines you can explain today without opening an email or asking anybody, which is the real measure of whether tagging works.

Run the exercise after the month-end close and never before it, because a margin computed on an unreconciled month changes once the month closes. If the variance came from volume rather than margin, the answer lives in the annual budget. And if a line is profitable but pays in ninety days, that is not a margin problem, it is a cash flow problem.

Set the approval boundary

Automation applies rules, compares outcomes and finds the transactions explaining a change. It does not choose the allocation rule. That choice determines which line looks profitable and which does not, so a person makes it, writes it down and keeps it even in a quarter when it hurts somebody.

Changing the rule mid-year is allowed and sometimes correct, but then the history gets recalculated under the new one and you say so. Comparing March margin under one rule with October margin under another is not analysis, it is an error with two decimal places.

Frequently asked questions

Which margin should I use?

The one matching the decision. Gross margin when deciding on direct costs, contribution or line profit when the decision touches shared structure. Picking the margin that looks best is the most common failure in this exercise.

Can AI allocate shared costs?

It can apply a rule, compare it against another and show how much the answer moves depending on which you pick, which is exactly the useful work. Choosing the defensible rule remains a business and accounting judgment worth discussing with your accountant.

How often should I run it?

Quarterly is enough for most small businesses. Monthly only if you make pricing or catalog decisions that often, and almost nobody does. What should stay monthly is the tagging, because that is the part you cannot recover later.

What if shared costs are most of my cost base?

Then margin by line depends more on your rule than on your business, and it is worth knowing that before making big decisions with it. In that case the useful exercise is not margin by line, it is counting hours per line for a month and looking at that number with nothing allocated at all.

Which product should I choose?

QuickBooks for ecosystem depth, Xero for balanced reporting and Digits for AI-led analysis. The real difference between them matters far less than the quality of your tags, so if you are going to spend effort somewhere, spend it tagging one month properly before comparing products.