What happens when you put AI on top of a broken invoice payment process?

Bruce Mullan

Aug 13

10 min read

Key points

The hardest part of automating invoice processing isn't teaching the AI to read them properly, it's fixing everything around them.
This is a real-world case study showing AI can deliver invoice processing at speed and scale. This meant fewer exceptions, fewer overdue invoices, fewer supplier queries, less staff time, better cash visibility, stronger supplier relationships and less operational risk.
To achieve that, you'll need to break apart big rocks and move them, to realise the benefits of the AI

Once upon a time, every invoice arrived in the mail.

That’s how many business fairytales begin. A beautiful little Finance department where everyone lived happily ever after until the big, bad AI came along.

Except this story isn’t really a fairytale. Invoice automation started back in 2007, when Kofax began working with Australian companies to digitise paper invoices. By 2013, machine learning had arrived. Systems could read unstructured invoice formats without requiring a separate template for every vendor. By 2020, native AI capabilities emerged to improve straight-through invoice processing by dynamically adapting to new supplier layouts and reduce the need for human intervention.

And by 2025, AI-powered Accounts Payable invoice automation had moved beyond extracting data from invoices to handling line-item data, identifying anomalies, alerting for potential fraud, and automatically routing invoices through approval workflows.

Finance teams have always been playing catch-up

Over the past 30 years, I’ve implemented many of the major ERP and Finance systems, along with the business systems that sit around them including travel & expense, time & attendance, payroll, inventory & supply chain logistics, client services and projects.

And one thing I’ve noticed is that Accounts Payable (AP) has traditionally been treated as part of Finance Operations: the business of doing Finance. The focus was transactional throughput but not necessarily value-add. How quickly can we get invoices paid? How do we stop suppliers chasing us?

Invoice in. Payment out. A sausage factory in full flight. As such, invoice processing was an obvious candidate for automation. After all, automation works best when a process is stable, predictable, standardised and rules-based.

When you say those four words out loud, I instantly think about Finance Departments who routinely moved data around to support recurring processes and monthly reporting cycles that rinse and repeat over and over again.

Invoice processing should therefore be a perfect candidate for automation. Shouldn’t it?

Not necessarily. One of the most benign and mundane of business processes, processing an invoice, is often nowhere near as stable, predictable, standardised or rules-based as it appears.

In fact, I’ve worked with many Finance teams where the exception is the rule. The Finance teams I got to know over the years, were always in catch-up mode between month-ends.

Why? Finance would be feverishly trying to get on top of the adhoc requests, projections, business cases, the next big thing, new reports, deep-diving into random analysis and so on. It was about choices. What gets priority and what’s ignored between reporting cycles. Finding time to standardise and automate mid-month never really got a look in.

An AI case study in moving big rocks

I recently had the privilege of implementing an AI-driven AP automation solution. During the vendor selection, the AI automation capabilities appeared eminently impressive. The vendor was already effortlessly processing millions of invoices around the world worth billions, at a speed and scale that CFOs get really excited about.

The AI itself was a tiny rock. Standing next to it was one big rock: the organisation.

The client (who has been de-identified) was a national company spending more than $80 million a year across 50,000+ invoices received from 2,500 suppliers. Invoices were being processed across hundreds of sites around Australia, with accountability decentralised to site managers. A high volume, low value invoice payment operating model with only 10% of suppliers accounting for more than 80% of the spend. Most transactions were small with local suppliers such that 90% of the suppliers accounted for only 20% of the spend.

Averages across operating sites:
- invoice amount was around $1,500
- annual spend per supplier was around $30,000
- on-time invoice payment performance was 57%
- invoices received per site was roughly one per work day

I could easily see the dozens of different ways of doing essentially the same thing in different locations. Some site staff were studious, others were haphazard due to staff turnover, time constraints, lack of training or lack of business acumen. It wasn’t going to be pretty at the other end: Accounts Payable. And it wasn’t.

Only 57% of invoices were being paid on time. The Accounts Payable team spent more time chasing missing invoices, fixing exceptions and processing one-off payments than actually processing invoices.

A perfect candidate for AP automation, but first I had to find some skeletons in closets.

Where Skeletons Hide in Companies

1. Data Quality

The first one was inactivation of suppliers no longer being used. No-one had time for housekeeping. We inactivated more than 4,000 suppliers no longer in use. This closed one door on potential fraud.

Next, we also discovered hundreds of obsolete employee accounts with major corporate suppliers such as CBA credit cards, Telstra phones, Bunnings purchase cards, travel agency accounts and uniform suppliers. For example, more than 30% of mobile phone plans were cancelled that belonged to people who had already left the organisation, saving thousands.

Then there was the invoice numbering problem. In addition to the invoice number, some suppliers included the sales order number, purchase order number and delivery number on the same document. Different site administration staff were entering different numbers as the invoice number! That might sound like a minor problem. Only Finance people will know it made subsequent reconciliation unnecessarily difficult and created another source of manual exceptions.

2. Immature Supplier Billing Systems

Then we looked closely at our suppliers. Around 90% of suppliers were incorrectly invoicing individual sites (i.e. Invoice To: Site Name) rather than the company’s legal entity name. That created a serious GST compliance breach. Approximately 2,000 active suppliers were contacted to invoice us correctly.

We also asked suppliers to include the site name on their invoices so AI could easily allocate the expense to the correct Finance cost centre. This sounds simple. It wasn’t. Many suppliers were operating legacy or home-grown billing / accounts receivable systems that were extremely difficult to reconfigure for a simple change of adding ONE field to an invoice. Good grief!

Asking major suppliers to produce one monthly consolidated invoice for dozens of sites was also a challenge for the very same reasons.

One last, minor, insignificant detail. Some suppliers were still using paper-based invoice books. No AI was going to solve that problem.

Supplier challenges:
- Suppliers were issuing non-compliant ATO invoices
- Suppliers had IT difficulties to modify invoices to suit the AI
- Some major suppliers were unable to produce consolidated invoices
- Hand-written paper invoices were still in use in regional areas

3. Finance Business Rules and Processes

Sometime in the deep, dark past someone in Finance made a rule that required every purchase to have an approved purchase order (PO), even for purchases for as little as $20. The administration burden meant the admin staff created the purchase order at the time the invoice had arrived.

That is, after the purchase had already been made. The PO wasn’t pre-approving anything. Controls were in place, but implemented in an uncontrolled way.

Credit notes were being processed inconsistently and incorrectly, creating significant manual effort for what should have been a straightforward process.

Staff reimbursements were being processed and paid as supplier payments, while employee bank details were being maintained in both Payroll and Finance systems.

We also found about 50 new suppliers were being requested every month, many of which were one-offs. The reason: Site staff weren’t using preferred suppliers, partly because there wasn’t a published list available and partly because of restrictions on credit card purchasing forced them to use the AP workflow for nearly everything.

The business operated on a calendar month-end, yet closing took AP more than four days.

Summary of business challenges:
- Every purchase had to have a PO irrespective of value
- All staff reimbursements were being processed as vendor payments
- 50 new suppliers were being requested every month
- No unused suppliers were being inactivated
- AP end-of-month closing took four days

4. An Unwieldy Chart of Accounts

The decentralised operating model had produced another problem: every division of the business had developed its own approach to coding expenses over many years. This had been allowed during years of corporate expansion where the focus was on flexibility and support for growth.

For example, food safety audits could be coded to Health & Safety, Training or Food depending on who was entering the transaction. The consequences meant the organisation couldn’t get a reliable picture of Health & Safety expenditure without some serious financial gymnastics. This is one example of dozens.

The ever-evolving general ledger had grown into something far more complicated than it needed to be. It was being asked to capture minute details that should have been managed by specialist systems and processes. The result was more accounts. More choices. More confusion. More inconsistent coding. And, ultimately, poorer data for analysis and reporting.

5. Purchasing Policy Gaps

There were purchasing policy gaps too. Suppliers could be engaged for services worth up to $10,000 without any formal contractual terms and conditions. That might seem harmless until something goes wrong.

General consumer laws may provide a safety net, but it is far better to establish expectations, responsibilities and standards of service before the work begins than to rely on legislation after the relationship has broken down.

6. People With Tacit Knowledge

Some AP staff had been hitting the keyboard for more than 15 years and knew everything about every supplier and every business user in their heads. That’s the tacit organisational knowledge that the AI never knows about. AP were continually intervening in the current process to correct expense codes or data entry mistakes made by the sites.

By capturing the invoice correctly, the new AI automations would eliminate this kind of intervention along with approximately 80% of their day-to-day processing work.

People’s jobs were changing. Instead of invoice data entry, the team would need to monitor the automation, review exceptions, oversee AI performance, analyse business spending and look for opportunities to optimise the process. New systems, new skills and new ways of thinking about the work.

Meanwhile, managers and site staff would benefit enormously from faster approvals and less administration. But routing invoices into a central process also created concerns about losing local control.

The technology was changing. The process was changing. And people’s roles were changing. The disruption to AP staff was significant.

7. No Baseline Performance Measurement

Finally, I asked for some AP performance metrics. How long did it take to process an invoice? How many invoices were processed without AP intervention? How many were exceptions? How much time was spent chasing invoices? How many invoices were paid on time?

There weren’t any reliable answers. The focus had simply been on catching up. Get the invoices in. Get them approved. Get them paid. Chase the unpaid ones. Fix the exceptions. Make the occasionally “one-off” payment. Which, of course, wasn’t really occasional at all.

So it was difficult to know how well (or not) the current process was working. Using some industry benchmarks, I was able to establish AP staff were only half as productive as the benchmark indicated they should be. The data was telling me this was due to the constant interruptions, interventions and follow-ups that plagued their day jobs supporting hundreds of sites at varying degrees of maturity. They just couldn’t get a clean run at it.

But Every Fairytale Has a Happy Ending

So I found some of the skeletons upfront, namely:

Data
Suppliers
Finance Business Rules
Chart of Accounts
Policy Gaps
Tacit Knowledge
Metrics

This meant we couldn’t automate an old process. We adapted to the built-in process the AI offered. The technology component was important, but the real transformation was in the business itself. Invoice automation meant fewer exceptions, fewer overdue invoices, fewer supplier queries, less staff time chasing up, better cash flow management, stronger supplier relationships and less operational risk.

What the AI delivered

92% on-time payment, up from 57%, within four months
Standardised, automated and tracked end-to-end workflows*
Automated application of authority and delegation thresholds
Accurate month-end spend and expense accruals in management reports and profit and loss reports

*Note: Touchless processing eventually arrived but the AI took a little bit longer than we expected to get there.

What the broader business transformation delivered

Achieved 100% GST compliance with ATO invoicing requirements
Consolidated invoicing from preferred suppliers reduced invoice volumes by 90% in those suppliers
Transitioned 100% of staff reimbursements into Payroll processing flows
Cancelled direct debits where no corresponding service or invoice existed
Established a detailed preferred supplier list to reduce new Supplier requests
Identified significant savings and unnecessary spend during the analysis stage
Standardised expense account coding across the business units for consistent reporting
Reduced the Finance / AP close from Day 4 to Day 1
Trained staff to focus on business spend review, oversight and optimisation

AI Doesn’t Fix Broken Processes

It’s easy to think that buying an AI-powered automation platform will make a process faster, cheaper and better. AI can automate any process to be faster but it won't be better. If the Finance process is poorly designed, inconsistent, enables too much flexibility or built on poor data, AI can simply help you do the wrong thing faster or make a bigger mess more quickly.

The technology might be the little rock. But the people, process, supplier systems and beliefs about how we should work are the big rocks.

And before you automate using that little rock, you need to understand the big rocks first. You’ll need to do the deep analysis to build a case for change ad bring those affected along with you. Keep in mind what they say about people in glass houses too.

Over the years, in every business software I’ve deployed, it was impossible to implement and preserve what people do today. There will always be a level of disruption. What people do now exists because someone wanted it that way. New AI-driven business systems will challenge the status quo.

What the AI offers you today, is a reason to finally fix things.

If this case study reminds you of some big rocks you're trying to shift, then AI might be that impetus to design your own happy ending to the Finance fairytale.

Stay safe,

Bruce

AI. Use responsibly.