If you personally approve every invoice over a certain size, you are not being careful. You are the last line of defence in a system that does not have one.

One owner was checking line items at nine in the evening, because catching a mistake now beat finding it three weeks later. The people were not the problem. The process was, and the caution was justified.

What the approval is standing in for

Approving everything is a control. It is the most expensive one you have, because it spends your time, and it stops working the moment volume goes up.

It exists because three things are missing. Nobody has separated who raises a payment from who approves it. Nothing checks the invoice against what was ordered and what actually arrived. And no limit has been set that says which payments genuinely need a decision from you.

Until those three exist, you are the control. Stepping out without replacing yourself would be reckless, which is why the habit is so hard to break.

What replacing yourself actually looks like

In one distribution business, the whole purchase-to-pay process was rebuilt so that it ran without anybody in the middle of it.

A purchase order goes to the supplier. The goods arrive and a goods received note is raised at the door. The invoice comes in and is checked. If something is wrong with it, it goes straight back to the supplier without anybody touching it. If it is valid, it is matched against the purchase order and the goods received note. Where all three agree, the transaction is processed. Where they disagree but the difference falls inside a limit set in advance, it is accepted anyway. Where the difference is larger than that, the supplier and the buyer both get an email and finance gets a ticket, so somebody owns the resolution rather than it sitting in an inbox.

Once matched, the payment is scheduled on terms. The day before it falls due, it is loaded at the bank and a listing goes to whoever releases payments. The payment goes out. When the bank confirms it, the remittance goes to the supplier.

Nobody touches any of that except on exceptions. The owner is not in the path at all.

Controls are about speed, not audits.

The word controls has been captured by the audit conversation, which is why it sounds like friction. In practice it is the reverse. A control is a written answer to the question "how do I know this is right", agreed once, so that nobody has to ask it again every time.

Payment runs that took three days clear in one, because nothing is queued behind one person's inbox. Suppliers stop chasing. Month-end stops waiting on a single review.

Passing an audit more easily is a by-product. It is useful when the auditors arrive or a buyer starts asking questions, but it is not the reason to do the work.

What has changed is the cost of building it

The first time that process was built, it took serious development. That is why this kind of thing has largely been the preserve of businesses with big technology functions, and why everybody else has carried on approving invoices in the evening.

The same process was rebuilt recently for a smaller business. It was much faster to build and it cost a fraction of the first one, because AI did the work that used to need a development team.

That is where AI earns its place in a finance function. Not as judgement, and not as a CFO. As the thing that makes the repeatable, rules-based part of finance run without people, so that the people are left doing the part that needs a view.

A test worth running on last month

Go back through the last month of approvals. For each one, ask whether your answer changed anything, or whether you were confirming something the process should already have known.

Count the ones where your answer changed something. That is your real list, and it is the work worth your attention. The rest are checks, and checks belong in the process rather than in your evening.

Why it decays, and what stops it

Splitting the work, matching the paperwork, and setting the limits is a piece of design. Keeping it true as the business changes is not, and that is the part that usually goes missing. New suppliers arrive. Limits set two years ago stop making sense. Somebody starts working round a rule because it is quicker. Controls that nobody owns quietly go back to being you.

Owned properly, a finance function is built so that you are asked for judgement, and not asked to verify.