How to automate invoicing with AI, without handing over the numbers.

The short version

Invoicing is three different jobs wearing one name. Rules handle the predictable part: creating the invoice, scheduling the reminder, updating the status when payment lands. AI handles the reading and drafting: pulling details off a supplier invoice, working out which cost code it belongs to, writing a chase that sounds like a person. A person still makes the decisions.

Most businesses that ask about AI invoicing need the first part done properly. That answer is less exciting and usually cheaper, so it is worth saying first.

1. What AI is genuinely good at here

Reading documents. A modern model can take a supplier invoice it has never seen before, in a layout nobody designed for machines, and pull out the supplier, the date, the line items and the total. This is the biggest single win, because manual entry is where the hours and the typos live. We used the same approach on a client project that reads engineering drawings and checks them against a standards library: the model reads, the person decides.

Classifying and routing. Which cost code does this belong to? Which inbox should it reach? Does it need approval above a threshold? The rules are fuzzy and the exceptions are frequent, which is exactly the shape of problem a model handles better than an if-statement.

Drafting the chase. A reminder that references the actual invoice, the actual amount and the actual days overdue reads better than a template, and gets paid faster. The model writes it, you approve it.

Spotting the unusual. A duplicate invoice number, a supplier charging a rate they have never charged before, a total that does not match its own line items. AI is patient about checking every one of these, which people are not.

2. Where AI quietly causes problems

The final number. Never let a model calculate or restate an amount. Arithmetic should come from your accounting system, not from a language model. If the model reads £4,208 as £4,280, you want the system to catch it, not agree with it.

Anything that needs certainty. VAT treatment, payment terms, compliance checks. AI can flag something for a person to look at. It should not be the thing that decides.

Unreviewed output. An AI that sends emails unsupervised is an AI that can send the wrong one to your best customer. The rule we work to: the model drafts, a person sends, until that specific process has earned the trust to run itself.

3. The three levels, cheapest first

Level 1, rules. Invoices are created from approved work, reminders go out on a schedule, statuses update when the money arrives. No AI anywhere. In most businesses we look at, this is where the majority of the saved time is.

Level 2, AI reading. Supplier invoices and receipts get read automatically and land in the right place with the right codes. This is where AI earns its place, and it is usually the second thing you build, not the first.

Level 3, a supervised agent. Something that monitors the queue, decides what needs attention and prepares the next action for a person to approve. Genuinely useful at volume. A bad first project for a business that has not fixed level one yet.

4. What a sensible first step looks like

Pick one invoice type and one person's week. Time the manual steps honestly for a few days. Then automate the step that takes the most time and needs the least judgement. In most of the businesses we see, that turns out to be data entry from supplier invoices rather than the customer invoicing itself, which people often assume is the problem.

Run it for a month. Compare the hours and the errors against the old way. Then decide whether the next move is a different workflow or the same one extended. That sequence keeps the risk small and the evidence real.

5. Build, buy or wait

Off-the-shelf accounting tools already handle the invoice itself and the reminders well. If that covers your situation, use it and stop reading.

Where those tools run out is the space between systems: the invoice sitting in an inbox, the approval buried in a message thread, the payment that has landed but not been matched. That gap is where AI reading plus a small amount of workflow logic does real work, and it is the part no off-the-shelf product covers.

Custom is worth it when the process is specific to how you operate, or when volume makes the manual version a job in itself. If you raise ten invoices a week, you do not need a system. If you process two hundred supplier invoices a month, you probably do.

If you are still doing this in a spreadsheet, the Excel invoicing guide covers the practical setup and the point where it stops being the right tool.

6. Where this fits in a real system

Reading documents is one capability inside a bigger process. The workflow around it, which is usually business automation, decides what happens when a document arrives, who owns it, and what chases what. The reading itself sits inside an AI integration. When the volume justifies something that watches and acts on its own, that is a supervised agent.

Choosing between those three is the decision that matters, and it is the one most businesses get wrong by starting too big.