How do you automate invoice processing, and what happens to the exceptions?

In most companies of a few hundred people, supplier invoices land in a shared accounts payable inbox as PDFs. Someone opens each one, types it into the accounting system, checks it against the purchase order by eye and chases the approver. On a quiet week that's a few hours spread across two or three people. In the week before month end it's the whole job, and the mistakes made at speed come back as credit notes, duplicate payments and awkward calls with suppliers.

It's often the first process a finance team hands over. Invoices arrive every day, they follow the same shape, and the decisions that need a person are easy to describe. Below is how the automated version works, where it still needs your team, and how to set it up without disrupting a close.

What happens to each invoice

The automated process has four steps. Two of them end with a person when something needs deciding.

1. Capture. The invoice arrives by email, through a supplier portal or as a scan. The system picks it up, records where it came from and when, and reads the supplier, invoice number, date, line items, amounts, tax and payment terms.
2. Match. It finds the purchase order and, if you record them, the goods receipt. It compares the fields you've said matter and marks each one as agreeing or not.
3. Route. If everything agrees within your tolerances, the invoice goes to the right approver, or is approved on a rule if you've allowed that for small routine amounts. If something doesn't agree, it goes to the person who owns that kind of difference.
4. Post. Once approved, the invoice is posted to your accounting system with its coding, the order is updated, and the file is stored where your auditors expect to find it.

Reading the invoice is what people worry about, and it causes the least trouble. Current document reading copes with different layouts, scans and the occasional phone photo. Whether the system works for you depends on the second step, and on rules only your team can write.

Decide what counts as a match before anything is built

Sit down with the person who does the matching today and write out what they check. The list is usually shorter than they expect, with a few judgment calls hiding in it.

• Fields that must agree exactly: supplier, purchase order number, currency.
• Fields that can differ within a tolerance: a rounding difference on the total, or a quantity within what was received.
• Differences that shouldn't block anything: a freight line that isn't on the order, or a price that moved within an agreed range.
• Invoices with no purchase order: utilities, subscriptions and some services never get one. Decide whether they go to a person every time or follow a simpler rule for a named list of suppliers.

Write each rule as a plain sentence with the actual numbers in it. Those sentences become the rules the system runs on, the cases it's tested against, and the document your auditor will ask for. If the team can't agree a tolerance, start with the strict one. Loosening a rule later is easy. Tightening one after a few bad payments is a harder conversation.

What to do with the invoices that don't match

An automated process surfaces exceptions faster than the manual one did. In the first week that feels like more work, because you're seeing all of them at once instead of finding them at month end. It settles once two things are in place.

The first is that each exception has an owner. Price differences go to whoever placed the order. Quantity differences go to whoever receives the goods. Unknown suppliers go to accounts payable. Nothing sits in a shared queue that everyone assumes someone else is watching.

The second is that the exception arrives ready to decide. The person opens one message and sees the invoice, the order, the receipt if there is one, and exactly which line differs and by how much. They approve it, reject it or send it back to the supplier from there, and the system records what they chose. If they still have to go and find the paperwork, the automation has saved them very little.

Watch which exceptions keep coming up. A supplier who always adds freight is a rule waiting to be written. A buyer whose orders never match is a conversation to have. After a few months the exception list shows you where your purchasing is loose.

Where a person stays in the process

There are two places, and everyone who touches the process should know them. The first is exceptions, as above. The second is approval above a threshold you set, even when the match is perfect. Some finance teams let a clean match under a set amount post on its own. Others want a person to look at everything for the first quarter. Make that decision on purpose. Copying the old approval chain often means copying approvals nobody reads.

What a person stops doing is typing. Nobody copies a total from a PDF into a field or searches the inbox for the second page of a scan. That's where most of the hours go, and it was never a good use of a trained accounts payable clerk's day.

How to start without disrupting month end

1. Pick one entity and one group of suppliers: the ones who always quote a purchase order and whose invoices usually match. Leave the awkward ones for later.
2. Take a baseline. Count last month's invoices, the hours spent on them and the number corrected after posting.
3. Run the system alongside the manual process for a few weeks, on the same invoices, and compare the results. Start mid-month, away from the close. This is where the match rules get corrected.
4. Switch the matched invoices over first. Exceptions keep going to the same people as before, now with the documents attached.
5. Add the next supplier group once the exception rate for the first is steady and you understand why each exception happened.

If you're still weighing invoices against other candidates, the four-question test for a first use case will help you score them. Invoices pass all four for most finance teams, which is why they come up so often.

What to measure

Agree the measures before the build and take a reading before anything changes, so that when you report the result you're comparing two numbers.

• Cost per invoice: the hours spent on invoices each month, costed, divided by the number of invoices.
• Hours back each week for the people who did the typing and matching.
• Errors and re-work: corrections after posting, duplicate payments caught, and credit notes raised because of a processing mistake.
• Time from receipt to approval, which is what suppliers notice and what early-payment discounts depend on.

How to measure AI ROI covers how to turn these into a figure your CFO will accept.

What to do this week

1. Count last month's supplier invoices and estimate the hours your team spent on them.
2. Sit with whoever matches invoices today and write down every check they make, with the numbers.
3. Name an owner for each kind of difference: price, quantity, unknown supplier and no purchase order.
4. Pick the entity and supplier group you'd start with.

This is the kind of system our build work covers: the reading, the match rules, the exception path and the connection to your accounting system, set up around the way your finance team already works.

Get the next article when it's published.