AI Financial Reporting That Tells You What Changed and Why

A Statement Is a Record. You Needed an Explanation.

Here is the pattern. The month ends. Three weeks later you get a profit and loss statement with a note from the bookkeeper. You look at the bottom line. If it is higher than last month you feel briefly good, if it is lower you feel briefly bad, and either way you close the attachment because there is nothing in it that tells you what to do differently on Monday.

That is not a failing on your part, and it is not a failing on the part of your bookkeeper, who is doing exactly the job you hired them for. A statement is a record of what happened. It is built to be accurate and complete, which is the opposite of being useful in the moment. The document that would actually help you is a different product, and most small businesses never get one because the gap between a P and L and a fractional CFO is wide and nobody sells into the middle of it.

The dashboard is the other thing people try. It gets built during onboarding, looks impressive, and is opened exactly twice. That is a design failure rather than a discipline failure. If the value depends on the owner remembering to go somewhere and look at something, the value is not going to be realized, because the owner is running a business and remembering to open a dashboard is the seventeenth most important thing in their day.

And then there is the harder problem. A tool that answers your questions requires you to have the question. An owner who knew that cooling repair margin had quietly compressed eight points on flat volume would already know what to ask. The reporting system that helps is the one that tells you before you knew to wonder.

What Actually Lands in Your Inbox

Switch between the daily, weekly, and monthly cadence. Same system, different altitude. Notice that the headline is a sentence, not a number of dollars.

Yesterday collected $18,240 against $21,900 invoiced. Two jobs closed with no invoice, and one deposit balance from September is still open.

Two completed jobs have no invoice

WO-8842 and WO-8851 completed yesterday and no invoice references either job number.

One estimate has been quiet for 34 days

Marlowe Dental Group, $8,400 commercial replacement. Past their usual decision window.

Parts cost variance on two jobs

WO-8830 and WO-8836 both ran parts cost more than 18 percent above estimate.

One technician logged 11.4 hours against 8.0 billed

Three jobs on that technician account for almost all of the variance, and two of them are the same repair type. This is worth a look before it repeats all week.

Sample figures for illustration only, not client data.

A Number Becomes an Answer

Click any variance to open it. This is the difference between being told margin dropped and being told which three jobs and which supplier caused it.

Labor Variance

Logged labor exceeded billed labor across the month.

Same repair type on all three, suggesting a labor guide or a parts issue rather than a slow technician.

Diagnostic hours were absorbed into flat-rate jobs 22 times in September.

Both comebacks traced to the same substituted part supplier.

No concentration above the tolerance threshold. Normal friction.

Figures Are Deterministic. The Narrative Sits on Top of Them.

The architectural decision that makes this trustworthy is a simple one, and almost everything that goes wrong with AI reporting goes wrong because this decision was made the other way. The figures are computed by queries, and the narrative is generated over the figures rather than over raw data.

Concretely: the agent runs a query, gets back an exact number, stores the query alongside it, and only then writes a sentence around it. The sentence can be wrong in tone, can be dull, can miss an implication, but it cannot invent a revenue figure, because the figure was never generated by the language model in the first place. Every claim in the narrative carries a link back to the query that produced it, so you or your accountant can verify any statement in two clicks.

The Ledger

QuickBooks Online Accounting API entities and report endpoints: Profit and Loss, Balance Sheet, Cash Flow, and Aged Receivables. Incremental reads use the changed-since query, so the nightly pull reads what moved instead of re-reading the whole books.

The Payments

Stripe and equivalents for payment volume, processing fees, refunds, and payout timing. This is why the collected figure in the briefing is what actually reached the bank rather than what was invoiced.

The Operations

ServiceTitan V2, Jobber GraphQL, Tekmetric Open API, or a report export surface for completed volume, average ticket, and the estimated versus actual split that produces most of the interesting variance.

The Pipeline

Open estimates and outstanding quotes by age. The leading indicator of next month is not next month. It is the shape of this month's unconverted pipeline, which is why it belongs in the briefing.

A note on integration. Most small businesses do not have their systems talking to each other, and that is the normal starting point rather than a blocker. Where a documented API exists, the agent reads it directly. Where one does not, it reads scheduled report exports. Either way, the briefing states which source every figure came from, so nobody argues about which system is right.

Three Lines We Are Careful Not To Cross

It is not a fractional CFO

A real fractional CFO is worth every dollar and does something this system cannot: exercises judgment about strategy, tax structure, and capital allocation using years of experience across many businesses. What this system does is make sure you arrive at that conversation already knowing what happened, which turns a $2,000 monthly retainer from discovery into advice. If you already have an advisor, this makes them better. If it is not yet the right time to hire one, this is what fills the gap until it is.

It is not a forecast

The report describes what happened and names the trend. It does not project. The difference matters because projections built on a business's own history and dressed in confident language are how owners end up making a capital decision against a number that was never more than an extrapolation. Reporting, trending, and projecting are three separate exercises with three separate levels of confidence, and conflating them is how trust in the whole system gets destroyed.

It is not a tax position

The monthly briefing will flag the two or three questions worth taking to your accountant, such as an allocation method or a timing question around a large purchase. It will not answer them. Tax treatment depends on your entity, your state, and facts the system does not see, and an automated answer to a tax question is worse than no answer because it carries unearned confidence.

Read and Assemble Alone. Recommend With a Person.

Straight Answers

Let Us Build One From Your Data

In thirty minutes we will look at what you have, identify which two or three figures actually drive your week, and assemble a sample briefing from your real numbers. You keep it whether or not you hire us.