Three-way invoice matching is such a crucial internal control of procure-to-pay workflows in every industry. Sadly, most businesses and their accountants still wade through it manually, not because they like to but because decades of mediocre automation have only replaced one set of problems with another.
It’s only in the last two years or so that smart, fully automated 3-way invoice matching has become practical, thanks to advances in deep learning and natural language understanding. There’s a revolution happening in accounting automation with tremendous business benefits.
We have explored general invoice matching in the past. In this article, you’ll dive deep into the popular 3-way invoice matching, learn about the severe problems around doing it manually, find out the enormous benefits for your business by replacing it with a fully automated solution, and finally explore the mechanics of how such a system works.
The illustration shows a typical procure-to-pay workflow used in most businesses and organizations along with the departments involved and the documents they issue:
Three-way matching is an internal control of the bookkeeping process to ensure there are no mismatches between the goods, quantities, and terms mentioned in the purchase order, goods-received note, and invoice.
In practice, there are always some mismatches due to unavoidable factors like minor damages, weather conditions, or currency fluctuations. So the three-way matching process is done flexibly as long as the deviations are within accepted tolerances.
Its biggest benefit is in avoiding financial losses in the procure-to-pay workflow. Losses may stem from invoicing mistakes, vendor fraud, or internal fraud. Or they may come as penalties by suppliers for late payments. Invoice matching is a vital control tool to help your organization avoid all that and safeguard its bottom line.
Plus, accounting and auditing standards like the International Financial Reporting Standards and the Generally Accepted Accounting Practices require organizations to do invoice matching.
Depending on the set of documents you reconcile, invoice matching can be 2-way, 3-way, or even 4-way. Let’s compare 3-way and 2-way matching:
If you’re in an industry like precision manufacturing with high quality thresholds, you’ll want to match inspection reports too with the other documents. That’s called 4-way matching.
If you’re using a manual matching process, you’re not the only one. Most businesses still follow manual matching because they have grown wary of automated matching after getting burned by the unreliability and inconvenience of semi-automated approaches like template matching.
But manual matching of the three documents brings some severe problems:
Automated 3-way matching overcomes many problems of manual matching by automating information extraction from purchase orders, invoices, and receipts and matching them using software.
However, the semi-automated approaches that the industry’s seen so far are unreliable and introduce new inconveniences. They expect you to manually create a large number of brittle “invoice templates” to extract data. Sometimes you have to transfer the data to spreadsheets first before automation kicks in. You may also have to verify the extracted data manually.
In contrast, a fully automated system understands invoices and other documents like people do. It can handle all the variations in layouts, terminologies, languages, currencies, and more. To achieve that level of smartness, it uses state-of-the-art computer vision, deep learning, and natural language processing.
Let’s explore all the benefits of fully automated 3-way invoice matching.
Accurate 3-way matching for a purchase takes just a few seconds, not hours. These systems are highly scalable, matching millions of invoices with other documents every day. Their inherent speed and scalability result in additional business benefits as we’ll see.
Fully automated matching understands documents semantically just as we do. When it sees text laid out roughly in a row, it recognizes it as a line item in a list of products. When it sees what looks like a column of information, it identifies them as item descriptions or amounts. It doesn’t need any manual templates, pre-processing, or verification.
Compared to manual and semi-automated approaches, a fully automated system achieves error-free, accurate data extraction and semantically sound 3-way matching.
Fully automated invoice matching dramatically lowers your operating and capital expenses. The capital spent on the software and infrastructure is much lower compared to hiring and training skilled people. The operating expenses for cloud resources are far lower than employment costs. The per-invoice cost reduces by about 75%.
Every discrepancy that’s automatically detected is filed in a database. A fraud monitoring component continuously aggregates them by vendor and time to look for anomalous patterns and other evidence of systematic long-term fraud.
Fast and accurate extraction and matching enable your business intelligence teams to get accurate real-time cash flow and expense data. Real-time data extraction from goods-received notes enables production lines to track part inventories accurately and run smoothly.
Fast matching leads to expedited invoice approvals that help avoid late payments and perhaps even get you early payment discounts.
Error-free 3-way matching enables accurate bookkeeping at all times and keeps your auditors happy.
In this section, we explore details of how a fully automated 3-way invoice matching works.
Fully automated 3-way invoice matching has little choice but to handle all the varieties of invoices, purchase orders, and goods-received notes found in the real world. Otherwise, it suffers the same problems as manual and semi-automated approaches.
These documents vary widely in aspects like:
Regardless of the format or medium, one norm that most businesses follow is arranging the list of goods and their quantities in a table. Identifying tables, or rather, information arranged in tables is an essential step in 3-way matching.
At this stage, we just want the locations of areas where text is spatially arranged in the form of tables. We also want to identify vertical arrangements of text as probable columns and horizontal arrangements of text as probable rows of the table.
We use convolutional deep neural networks (DNN), trained for either object detection or segmentation, for these table/row/column localization tasks. Techniques like clustering and non-maximal suppression are used to identify columns and rows with even higher probability.
With probable rows and columns identified, we use a second DNN for text extraction. Text extraction combines simple character recognition and natural language models to accurately identify entire fragments of text without suffering the typical problems we see with only optical recognition of text characters.
Every text fragment is then run through a named entity recognition (NER) neural network to classify it as a product title, description, serial number, column header, quantity, price, total amount, and so on.
A PO contains a lot of other information necessary for matching. It contains a PO number, vendor name, authorizing employee, receiving department, specification identifier, date of issue, dates of expected delivery, and more.
All these details are also extracted and labeled by the NER network. The named entity labels and values are stored in a structured, queryable data format like JSON in a database or routed to the central ERP system.
Similarly, invoices contain details like invoice number, date of issue, vendor name and identifier, terms of payment (like “Net 30”), and other important information that’s necessary for matching. These too are extracted and labeled by the NER network before storing them.
Receipt notes contain details of goods and quantities that were received and of those that were rejected. This information is extracted by the same NER network and transmitted to the database or ERP.
Matching products and services across invoices, goods receipts, and purchase orders is not simple:
That’s why fully automated 3-way invoice matching has to use smart data matching. Titles, descriptions, and information like SKU numbers are matched using large language models like GPT-3 to achieve human-level understanding.
Quantities of respective line items are matched while honoring tolerance settings, unit conversions, and locale formats.
For example, a PO specifies quantities in units of thousands but the invoice includes full values. Or a purchase order issued in the U.S. uses pounds while the vendor uses kilograms for the benefit of their local logistics partner or customs authorities.
Such smart matching is handled by customization plugins that are configured with tolerance settings and unit conversion rules.
Similarly, the system matches prices while honoring currency conversions and locale formats. For example, some countries use the decimal point while others use the comma. You can set the tolerance levels for matching prices using price outlier detection.
The above steps result in three sets of data:
All these data are stored in a database or ERP. Plus, reports are generated for manual review by senior accountants and management.
The match reports are reviewed regularly by senior accountants in the AP team. The data are also used by automated workflows for putting invoices on hold, releasing them for payments, or forwarding them to senior management for budget approvals.
In this article, you learned about the benefits and mechanics of a fully automated 3-way invoice matching that uses deep learning. With Width.ai’s fully AP automation solution, you can process millions of invoices and 3-way (or even 4-way!) match them in seconds. Contact us for a demo.