AI for CFOs

A three-person finance team is often where it matters most, because there is nowhere for the work to go. In a thirty-person function, a bad month gets absorbed by somebody working a weekend. In a three-person function, one person on leave during close is a genuine business risk, and the single point of failure is usually the person who knows how the AP inbox actually works. Automation in a small function is less about cost reduction and more about removing the key-person dependency and making the close survivable when somebody is away. The honest counterpoint: if your transaction volume is genuinely low — a few dozen supplier invoices a month — the processing burden may not be where your pain is, and you should not be buying this.

Almost never in the parts people assume. Consolidation, judgement calls on accruals, and the review itself do not compress much, because they are analytical work. What compresses is everything upstream: the AP invoices that were sitting uncoded, the supplier statements nobody had reconciled, the receipts an employee never submitted, and the two days spent chasing approvals from managers who did not read the email. In most finance functions the close is long because the ledger was not current on day one, not because the close activities are slow. Getting the ledger current continuously is what shortens the calendar, and that is what automation does well.

The failure mode you are describing is real and it is usually caused by making approvers log into something. The AI routes an invoice to the right approver based on your delegation rules, chases them where they already are, and records the approval against the transaction. The approver does not learn a new interface for the sake of it. The important design point is that the AI routes and chases, but the delegation of authority is your policy, not a setting the tool invents — if your board has approved a delegations schedule, that schedule is what gets configured, and exceptions to it get escalated rather than quietly accommodated.

They deserve genuine thought rather than reassurance, because AP is where fraud lives. The design principle is that the AI proposes and a human disposes: it never creates a payment, never changes supplier bank details, and never approves anything. Supplier bank detail changes are the single highest-risk event in AP, and they are treated as an exception requiring human verification through a channel that is not the email that requested the change. Everything the AI does carries an audit trail showing what it saw, what it proposed and who approved it — which is generally more evidence than a manual process leaves behind. But you should be walking your auditor through the control design before rollout, not after, and we would expect you to.

It complicates payroll tax far more than it complicates automation, and it is worth separating the two. The processing work — AP, coding, reconciliation prep — is essentially the same regardless of entity or state. What differs is the state-based obligations sitting on top: payroll tax is a state tax, so Victoria, New South Wales and Queensland each set their own threshold, rate and grouping rules through their own revenue office, and a group operating across borders has to deal with all of them plus the grouping provisions that can pull related entities together. The AI keeps the underlying data clean and consistent across entities, which makes that analysis easier. It does not make the payroll tax determination — that is advice, and it should come from your tax adviser.

We would rather you did not build a headcount case on it, and that is not false modesty. The reliable outcome is that your existing team stops processing and starts analysing — the AP officer who spent most of the month keying invoices instead owns the supplier relationship and the exception queue, and the accountant who spent a week on close prep instead spends it on the variance analysis your board actually reads. If you are growing, the realistic benefit is that the next hire is deferred rather than that the current one is cut. If you go in planning to remove people on day one, you will discover that somebody still needs to own the exceptions, and that person needed to understand the process to do it.