Ask the owner of a business with ten or thirty people who their CFO is and you will usually get a pause, then one of three answers. “Me, sort of.” “Our bookkeeper, but she only does the books.” “Our accountant, but I only hear from him in June.”

What they mean is that nobody is doing the job. The compliance gets done. The thinking does not: what is our cash position really, who owes us and for how long, which supplier invoice does not match what we ordered, can we afford the hire, what happens if that big client pays thirty days late?

A real CFO costs more than most small businesses can carry. But look at how a CFO spends the week and a great deal of it is not strategy. It is assembling.

The repetitive majority

  • The cash position, every Monday, from the bank feed and the accounting system: what is in, what is due out, what is due in, how many weeks of runway.
  • Receivables. Who owes what, how late, and the polite chase, then the firmer one.
  • Payables. Matching supplier invoices against what was ordered, pushing the matches through, bouncing the mismatches back with the adjustment.
  • The calendar. Tax dates, super, payroll, renewals, the insurer’s deadline.
  • The monthly pack. Revenue, margin, costs against last month and last year, with the three things that moved and why.
  • The what-if. A simple model you can ask a question of.

I would call that sixty to eighty percent of the role by hours, and every line of it is reading, matching, calculating, drafting and reminding. I have this working as a proof of concept across my own businesses: the accounting system and the bank feed as sources, an assistant built on Claude doing the assembly. I do not yet run it every week, and I will say why further down, because the reason is instructive.

The lines it must never cross

Finance is where you want an AI to be most useful and least trusted, so the design matters more here than anywhere.

It never moves money. It prepares the payment run; a person approves it, by name, every time.

It never sends to a supplier or a customer on its own. The chase email is a draft. So is the query about the mismatched invoice.

Every number traces back. Each figure in the pack is tied to the source it came from and the version of the logic that produced it, because the first question anyone asks when a number looks wrong is “where did that come from?”

One official number. When analysis gets cheap, two people bring two different margins to the same meeting. Decide which report is the source of truth for each headline figure, and label everything else as exploratory.

It does not replace your accountant. Tax advice, structuring and sign-off stay with a qualified human. What changes is that your accountant receives clean, reconciled books and a list of real questions, instead of a shoebox.

What it feels like

The difference is less dramatic than you might hope and more valuable. Nothing clever happens. You simply know, every Monday, where you stand, and you stopped being the person who works it out. Decisions that used to wait for month-end get made on a Tuesday. The invoice that would have aged to ninety days gets chased at thirty-five. The founder gets Sunday night back.

The honest state of mine

My own version is at the proof-of-concept stage. It can pull the position together and it can match invoices. What it is not yet is a habit, and the reason is the lesson. A finance loop only earns its keep when it arrives unprompted at the same time every week and someone has agreed to read it. Building the assembly took days. Deciding who looks at it on a Monday, and what they do next, is the part that turns a clever demo into a finance function. If you build one, settle that first.

Where to start

Do not start with the forecast. Start with the most boring loop you have, which is almost always either the weekly cash position or supplier invoice matching. Run it alongside the manual process until the numbers agree for a month. Then cut over and pick the next one.

If your finance function is currently “me, sort of”, tell me what you run the books on and what eats the most time. I will tell you which loop to automate first and what it would take.