Where procure-to-pay really breaks: 251,734 order lines from one real company

One in seven purchase order lines was changed after the order went out. On one in five lines where the price changed, it changed after the invoice had already arrived. And barely a quarter of the lines that were supposed to wait for a goods receipt before the invoice followed the process as designed.

Those numbers come from a real procure-to-pay log: every purchase order line, goods receipt, invoice and payment a multinational paints and coatings company recorded in 2018. The company published it, anonymised, for the 2019 Business Process Intelligence Challenge. It holds 251,734 order lines on 76,349 purchase orders and almost 1.6 million timestamped events. Nobody filled in a survey. This is what the system recorded.

The control that worked

Where the system was set to require a goods receipt before an invoice could be booked, it held. Of 11,128 lines in that category with both events, not one invoice was booked first.

The vendors tried. On a third of those lines (3,491 of 10,709) the supplier's own invoice was dated more than a day before delivery. The block was doing real work, every week.

Where it leaked

Most of this company's orders (221,010 lines) were set up the other way: the invoice may arrive first, and a payment block holds it until delivery is recorded. That design leans on people to release blocks correctly.

• A payment block had to be lifted on 26.8% of invoiced lines in that category (53,569 of 199,564).
• 654 invoices were cleared with no goods receipt before them: 89 before delivery, and 565 with no goods receipt anywhere in the log. That is 0.4% of cleared lines, small as a share and not small as money paid for deliveries nobody recorded.

The log cannot say why each block was set or who lifted it. A quarter of all events carry no user at all, so this study says nothing about individuals.

Changes after the order

39,582 lines (15.7%) were changed after the order was raised. Quantity changes led (7.0% of all lines), then price changes (4.5%) and deleted lines (3.5%). Another 6,471 lines had an invoice receipt cancelled.

The price changes are the uncomfortable part. On 2,195 of the 11,224 lines with a price change, the first change came after the first invoice was booked. Some are legitimate, for example a contract price that changes for later deliveries on the same line. The log does not show how many were made to match the invoice, which is exactly the question an auditor would ask next.

How long it took

From order to cleared invoice: a median of 77 days, and one line in ten took longer than 119 to 127 days depending on the category. Goods arrived after a median of 4 to 10 days. The invoice was then booked after 23 to 41 days. Payment terms are not in the log, so part of the wait is agreed terms and part is delay, and the two cannot be separated.

At the end of the extract, a fifth of the goods-first lines raised in the first half of 2018 (1,637 of 8,184) still had no cleared invoice. Two thirds of those had been delivered and had no invoice at all.

The process on paper and the process in the system

The log holds 11,973 different sequences of steps. The ten most common cover 59.5% of lines. Three quarters of the sequences happened exactly once.

Measured against the designed flow for each category, 72.4% of consignment lines followed it exactly, 46.5% of invoice-first lines did, and only 27.9% of goods-first lines did. A line with several deliveries, such as monthly rent, counts as off the design without anything being wrong, so distance from the design is not the same as a breach. It is the gap between the process the documentation describes and the one people actually run, and it is where every finding above lives.

What this does not tell you

It is one company, in one year, in one industry. The company field was anonymised down to four codes against a published description of 60 subsidiaries. Values were scaled, so there are no money figures here, and the approval workflow and payment terms were never in the extract.

Method and data

Source: van Dongen, Boudewijn (2019): BPI Challenge 2019. 4TU.ResearchData. Dataset. https://doi.org/10.4121/uuid:d06aff4b-79f0-45e6-8ec8-e19730c248f1, licensed CC BY 4.0.

A line is one purchase order line item, which is one case in the log. Ordering tests compare the first occurrence of each event on a line. 104 events dated between 1948 and 2020 are all vendor-supplied document dates ("Vendor creates invoice", "Vendor creates debit memo") and are left out of every ordering and duration test. Changes count only when dated on or after the line was created. Because the extract ends on 18 January 2019, durations and open items are also checked on lines created January to June 2018, which gives the same picture. Every figure reruns from the raw file in under a minute.

APQC's Process Classification Framework® (PCF) is an open standard developed by APQC, a nonprofit that promotes benchmarking and best practices worldwide. To download the full PCF or to view definitions and measures, please visit www.apqc.org/pcf.