How to build a small-business budget with AI

TLDR: QuickBooks Online is the US ecosystem choice, Xero provides a balanced accounting workflow, and Digits offers the most AI-first analysis. Build scenarios from closed books and keep every assumption visible.

The file called version four final

It lives in a shared drive, it is named something like budget 2026 v4 final, and it was last modified on the fourteenth of February. It took three long sessions in December, it was argued over, it was approved, and nobody has opened it since. In June somebody asks whether you can afford another hire and the honest answer is that you do not know, because the document that should answer it stopped touching reality four months ago.

A budget that never gets compared to what happened is not a plan. It is the minutes of a meeting in December. The work that keeps it alive is not building it, which is the part everybody does. It is the half hour a month where somebody looks at the variance and decides something.

That is where automation belongs, and it is not in generating the budget. Generating numbers is the easy half. Getting somebody to look at them in March is the hard half.

Start from twelve closed months

Begin with twelve reconciled months and separate recurring, variable and one-time items. Fix inconsistent categories first, because a history where the same expense moved accounts three times produces a budget with a trend that does not exist.

A sophisticated model does not compensate for messy history. If last year has unreconciled months in it, the right order is to close those first, and the month-end close guide covers that.

Write the assumptions, not just the numbers

This is what separates a useful budget from an attractive sheet. Every large figure needs a sentence behind it saying where it comes from and who stands behind it. Not revenue of six hundred thousand, but which price, which volume and which acquisition assumption sit underneath, plus the name of the person who owns that assumption.

It buys you two things. In June you can check the assumption instead of arguing about the result, which is a far shorter conversation. And a written assumption can be proven wrong, whereas a bare number can only be defended.

AI is good at computing what happens when one assumption moves, and at flagging when two assumptions contradict each other, such as a sales plan that requires capacity the headcount plan never funded. That is real, tedious work it handles well. Writing the assumptions is not its job, because it has never met your customers.

Three scenarios, and only three

Move few assumptions between them and keep those visible. A scenario that shifts fifteen variables at once cannot be analyzed, because when it misses you have no idea which one missed.

The base case is what gets approved and what you compare against monthly. The downside is not approved. It exists to answer one question: if this happens, what do we stop doing, and in which month do we decide it? If that answer is not written down, the downside case is decoration.

Tie each figure to a decision rule

A budget with no decision rules changes nothing. Hiring, capital spending and expense limits each need a written condition: what has to be true, measured by which number, before it gets approved or paused.

Written that way, in March you are not debating whether to hire. You are checking whether the condition you agreed in December has been met, back when nobody was in a hurry and everybody was thinking more coldly.

The four places a budget breaks

The assumption with no owner. Somebody put the marketing line in during the meeting and nobody claimed it afterward. With no name on it, that figure never gets reviewed, never gets defended and gets spent in full by October. Every large assumption needs a person who answers for it.

The variance that is only a date. A significant payment lands in January instead of December and suddenly you have one bad month and one excellent one. Nothing happened. A date moved. Telling a timing shift apart from a real problem is the first question in any monthly review, and confusing the two leads to fixing something that was never broken.

The upside case that quietly becomes the plan. Nobody approves it, but decisions start getting made against it, because it is the one with room for everything people want to do. You catch it with an uncomfortable question in the March review: are we spending against base or against upside?

The renewal that resets a whole line. Health insurance, a key software contract, a lease. The quote arrives in October and it is not the figure in your budget, and unlike most variances you cannot decide your way out of it in the same month. Note the renewal dates in the budget itself so the conversation starts a quarter early rather than the week the invoice lands.

What to measure when the year ends

Almost nobody grades their own budget, and it is the step that makes next year’s better. Count three things, in January, while the written assumptions are still in front of you.

How many of your assumptions held and how many did not, one by one. How many variances ended in a decision rather than a comment. And in how many months of the year somebody genuinely opened the comparison. If that third number is under six, next year’s problem will not be the model.

QuickBooks Online starts at $75, Xero at $25 and Digits at $65, with the analysis depth varying by tier. Import a history, build a budget and check whether the product traces a variance all the way back to the entries behind it. Do not choose on a generated chart, because every product makes good charts now.

Set the approval boundary

Automation computes impacts, compares scenarios and flags contradictions. It does not set your prices, decide who gets hired or approve capital spending. Those decisions have consequences that live outside the ledger, and a person signs them.

One more limit worth being clear about. A generated scenario does not go to a bank, an investor or a customer as a commitment until a person has reviewed it and taken it on. A model can produce a coherent year of numbers in a minute, and that coherence is not the same thing as a commitment somebody can meet.

Monthly tracking rests on the month-end close and reads better alongside your financial reports. For quarterly cash use the 13-week forecast, which is a different tool with a different horizon. And if the variance is in margin rather than volume, the answer sits in product profitability.

Frequently asked questions

Can AI predict revenue?

It can extrapolate historical patterns and run scenarios off your assumptions. It does not know the contracts you are negotiating, the commercial decision you will make in April, or which customer is about to leave. What it does well is the arithmetic of consequences, not forecasting the world.

Do I still need a spreadsheet?

For a genuinely custom model it may still be the best tool. For tracking, no, because tracking works where the accounting data already lives and a spreadsheet forces you to recopy figures every month. That recopying is exactly what makes it stop happening in March.

Which scenario should I approve?

Base as the operating plan. The downside is not approved, it is a stress test with its decisions written in advance. The upside only counts if it names the specific actions that make it possible and who is executing them.

How often should it be reviewed?

Once a month, half an hour, after the month is closed. Weekly adds nothing because the numbers have not moved enough, and quarterly arrives too late for almost any useful correction.

Is it worth it for a small, predictable business?

With three recurring customers and a stable cost base, a formal budget adds little and the effort belongs in cash flow instead. What always pays off is writing down the two or three assumptions your year depends on, even if they fit on half a page.