How Process Mining Revealed Hidden Invoice-to-Pay Costs and Control Gaps

A supplier sends an invoice. The invoice is checked. It is posted. The payment is made. But in large organizations, the actual process can look very different.

Invoices may be changed several times. Purchase orders may be created after an invoice has already arrived. Payment terms may change between the purchase order and the final payment. Some invoices may take weeks to get posted.

These issues are easy to miss when teams review only a small sample of transactions. Process mining gives a different view. It looks at real transaction data and shows how the process actually runs. It can reveal delays, repeated manual work, unusual process variants and control gaps.

A recent Invoice-to-Pay process mining exercise shows just how much can be hidden inside a routine finance process.

What was the problem?

Industry: – Manufacturing

No. of Employees: – 8,000+

Problem Statement:

• Six company codes were each following a different variant of the same Invoice-to-Pay process, making standard controls hard to enforce across roughly 80% of business volume.
• Purchase orders were created after the supplier invoice in 4.5% of cases, and invoices took an average of 19 days to post, with only 10.2% posted within seven days.
• Payment terms were changed on 35% of invoices (an estimated USD 463.8 million paid early), while manual invoice edits and master-data re-entry consumed the equivalent of 63 full-time employees a month.

Solution Implemented: – SAP Signavio, SAP LeanIX, Mind-A-Mend Accelerators, ProcessOrbit

The case involved a large manufacturing organization with more than 8,000 employees. The review covered three end-to-end processes across seven lines of business. For Invoice-to-Pay, nine months of data across six company codes were analyzed using a Plug & Gain approach.

The aim was not simply to find slow invoices. The goal was to understand where the process was breaking down and where the business was losing time, money and productive capacity.

The results showed more than 1,500 blockers. Some were process issues. Others were linked to manual changes, poor use of master data and differences between company codes.

The findings showed why looking at the complete process matters.

One Invoice-To-Pay Process Was Actually Six Different Processes

One of the clearest findings was process variation. The six company codes followed six different variants of the same Invoice-to-Pay flow. These six variants represented around 80% of the business volume.

This creates a basic question:

If the same business process is meant to work in the same way, why are people following different paths?

Different process variants can create extra work. They can also make it harder to enforce controls. A rule that works for one process path may not work for another.

Process mining helps make these differences visible. Instead of asking employees how the process works, teams can look at the transaction data and see which paths invoices actually take.

That is an important difference.

4.5% Of Invoices Had A Basic Control Problem

One finding stood out. In 4.5% of cases, the purchase order was created after the supplier invoice had already been issued.

In simple terms, the business received the invoice before creating the purchase order. That happened in about one out of every 20 cases. This is more than a process sequence issue.

The purchase order is meant to provide an important reference for the transaction. When the order comes after the invoice, the expected sequence is broken. A simple process mining rule can detect this.

The system can compare the purchase order date with the invoice date and flag cases where the PO comes later. This turns a control requirement into something that can be checked continuously.

The 19-Day Invoice Posting Delay

Another major issue was invoice posting. The average delay between receiving an invoice and posting it was 19 days. Once invoices were posted, they generally cleared within a few days.

Only 10.2% of invoices were posted within seven days in the detailed analysis. That delay matters. A late posting can push the entire payment process back.

The supplier may not receive payment on time. The finance team may have to handle follow-ups. Employees may spend time checking invoices that should already have moved forward.

The issue is not necessarily the payment step itself. The problem can start much earlier. This is one of the benefits of process mining. It connects individual steps and shows where a delay begins and what happens after it.

Payment Terms Were Being Changed

Payment terms were another major area of concern. The analysis found that payment terms were changed between the purchase order and payment in 35% of invoices. In some cases, terms moved from 30-day credit to immediate payment.

This can have a direct effect on cash. If an invoice is paid earlier than required, the company may lose the opportunity to keep that cash for longer.

The case analysis estimated that USD 463.8 million of invoices were paid early. At a 5% cost of capital and an average of 19 days paid early, the estimated annual financing impact was about USD 0.74 million.

The lesson is simple. Paying early is not always a problem. But it should be a conscious decision. If an early-payment discount is smaller than the cost of using the company’s cash, paying early may not make financial sense. Process mining can help identify these patterns at transaction level.

Manual Invoice Changes Were Using Valuable Capacity

Another finding was linked to manual work. Around 63 FTE worth of productive man-days were being spent each month on manual invoice changes. That does not necessarily mean 63 people were doing nothing else.

It means the amount of manual effort was equivalent to roughly 63 full-time employees working on these activities. That is a significant amount of capacity.

The question then becomes:

• Which of these activities really need human intervention?
• Some may need review.
• Others may be suitable for automation.

This is where process mining can support automation decisions. Instead of choosing automation candidates based only on assumptions, teams can look at transaction volume, frequency, rework and the effort involved.

The Master Data Problem

Master data was another weak point. Payment terms were re-entered on 65% of invoices. Posting dates were also manually re-entered in 49% of cases.

Why does this matter? If the required information already exists in the system, entering it again creates another opportunity for error.

It also adds work. A better approach is to use trusted master data wherever possible and reduce unnecessary manual entry.

The case proposed a control where payment terms and posting dates would be derived from master data, with manual re-keying blocked and logged. The target was to reduce re-keying to below 5%.

From Finding Problems To Preventing Them

Finding a problem is only the first step. The next step is to turn the finding into a control. The case study proposed continuous test cases for the Invoice-to-Pay process.

For example:

Buy-before-invoice check: – The purchase order date must come before the supplier invoice date.

Post-and-pay check: – Invoices should be posted within the defined SLA and paid by the due date.

Payment-term integrity check: – Payment terms on the invoice, purchase order and vendor master should match unless there is an approved exception.

Straight-through processing check: – Manual business-field changes after posting should be reduced.

Master data check: – Important fields should come from master data rather than being entered again.

This creates a simple cycle:

Detect → Prevent → Sustain

The process is monitored. Exceptions are flagged. Teams act on them. The same checks are run again.

That is much more useful than running a one-time process review and moving on.

Why Process Mining Works Well For Invoice-To-Pay

Invoice-to-Pay has many steps. There are purchase orders, goods receipts, invoices, approvals, postings and payments. There are also many people and systems involved. That makes it difficult to understand the full process through workshops alone.

Process mining uses transaction data to reconstruct the process.

It can show:

• Where invoices are delayed
• Which process variants are being followed
• Which fields are changed most often
• Where manual work is concentrated
• Which suppliers have unusual patterns
• Where controls are being bypassed
• Where payment terms change
• Which activities may be suitable for automation

In the case study, 449,230 invoice cases were analyzed, with 72.36% containing a change event. Payment terms were the most frequently changed field, accounting for 28% of changes.

That level of detail is difficult to get from a small sample of invoices.

What Companies Can Do Next

A good Invoice-to-Pay improvement program does not need to start with a complete process redesign.

• Start with the data.
• Look at the actual process.
• Find the largest delays.
• Find the most common changes.
• Find the process variants.
• Then identify the issues that have the biggest business impact.

From there, teams can decide what needs to be:

• Retained — because it works.
• Re-engineered — because it creates unnecessary work.
• Automated — because it is repetitive and rule-based.
• Controlled — because it creates financial or compliance risk.

This approach keeps improvement work focused on the problems that matter most.

The case study followed a similar path. More than 5,000 KPIs and process metrics were used to move from broad process visibility to high-impact areas, root causes and improvement candidates.

The Bigger Lesson

Invoice-to-Pay problems are often not caused by one big failure. They come from many small issues.

• A purchase order created too late.
• An invoice sitting unposted for weeks.
• Payment terms changed without a clear reason.
• The same process is handled differently by different company codes.
• Information entered again even though it already exists.
• Each issue may look small on its own.
• Together, they can create a large cost.

That is why process mining is useful.

It gives finance and process teams a way to move from assumptions to evidence. Instead of asking, “Why are invoices taking so long?” They can ask:

• Where exactly is the delay?
• Which transactions are affected?
• What causes the delay?
• How much work does it create?
• Can the issue be prevented?

Those questions lead to better decisions.

Which SAP Business AI Agents to Be Considered

The findings above are also a good starting point for agentic AI. SAP’s Business AI catalog on the SAP Discovery Center (discovery-center.cloud.sap/ai-catalog) lists Joule Agents that map closely to the gaps this case study uncovered:

• Invoice Data Extraction Agent: Extracts invoice data from PDFs and images with high accuracy, cutting down the manual re-keying of payment terms and posting dates seen in 65% and 49% of invoices in this case.
• Missing Goods Receipt Agent: Detects missing goods receipts early and prompts receivers before invoices fall due, targeting the same PO-after-invoice sequence gap found in 4.5% of cases.
• Payment Term Conflict Resolution Agent: Resolves conflicting payment terms across the purchase order, invoice and vendor master using learned decisions and guidelines, addressing the 35% of invoices where terms were changed.
• Invoice Exception Handling Agent: Handles invoice exceptions autonomously or guides users through resolution options, helping clear high-volume blockers like the 1,500+ identified in this case.
• Non-PO Invoice Coding Agent: Identifies tax, general ledger, WBS and supplier codes for non-PO invoices based on historical patterns, reducing dependence on manual coding and re-entry.

These agents can be evaluated individually or orchestrated together through SAP’s Accounts Payable Assistant, alongside custom agents built in Joule Studio for a company’s specific Invoice-to-Pay process.

About BPX

BPX focuses on business process management, process mining, SAP Signavio and business transformation management.

The team has more than 13 years of BPM experience and has delivered 300+ engagements, with 25K+ processes optimized. Its SAP BTM capabilities cover areas including SAP Signavio, LeanIX, SAP Cloud ALM, Tricentis and WalkMe.

BPX also uses a data-led approach to process discovery, process analysis, process optimization and AI enablement.

For organizations looking to improve Invoice-to-Pay, Procure-to-Pay, Order-to-Cash or Record-to-Report, process mining can help identify the issues that are worth fixing first.

The focus is simple: understand the real process, find the root causes and turn those findings into measurable improvements.