A client asks a quick question. Why does it take half a day?

The gist

An accounting firm does Fractional CFO work for small businesses, and its clients ask quick questions all the time. Where did my profit go? What do we owe, and what's overdue? Each one used to take the staff three to four hours. They pulled QuickBooks into Excel, reworked an old template, and checked it by hand.

No two questions are alike, and no two clients keep the same chart of accounts. So there was never a template you could just drop the data into.

I connected the firm's QuickBooks to one database and mapped every client's chart of accounts once. Fixed rules check the numbers. Now a client's email becomes a checked, interactive report in minutes, and the staff only validate it before it goes out.

You can open three of those reports yourself. The company and its books are fictional.

What this solves

Answers in minutes, not half a day. The report is built from the live books, so nobody pulls exports or reworks a template.

Every client's books, mapped once. Each chart of accounts is mapped to the same model, so a new question doesn't mean a new spreadsheet.

Numbers that check themselves. Fixed rules test every report before anyone sees it. Profit ties to cash, cash ties to the bank, and the bank ties to the balance sheet.

Nothing goes out unchecked. The workflow drafts. A person approves every report before the client sees it.

The problem

Ad-hoc questions are where a firm's time leaks. Each one is a little different: a different question, a different period, a different chart of accounts. The template from last time never quite fits. So someone pulls fresh exports, rebuilds the report, and checks the totals by hand. That's three to four hours for a question the owner asked in one line.

Workflow

Here's the path from a client's email to an approved report.

Three calls shaped the design:

• A live dashboard can't go in an email. It opens only inside the data tool. So the client gets a snapshot page behind a link that expires.
• Making sure the AI chooses the right report. AI picks the report, but approval is required. So the system builds context around each request to improve itself.
• AI can misread or invent a number. A report with a wrong figure is worse than no report. So the system pulls the numbers directly from the database, with fixtures to confirm them.

What's next

Each new kind of question becomes another report on the same model: mapped once, checked every time.

Fine printFictional company and people, synthetic books. The reports are snapshots taken 6 Oct 2026.