Invoice Processing

Every business must pay its suppliers, and every supplier sends invoice documents to businesses containing the essential data to process and complete the payment. It’s up to someone in accounts payable to read the invoice, find the data, and enter that data into the accounting system so the invoice can be recorded properly and paid. Despite several initiatives over the years, as of today there is still no industry standard template for invoices. They come in all shapes, colors, and sizes, and the data is rarely found in the same place from vendor to vendor.

This lack of standards means that every invoice needs special attention from a human. The cost to process a single invoice can be as high as $13, according to the Institute of Finance & Management. It’s no wonder that business managers have long sought ways to automate invoice processing and drive that cost down as far as possible. Automation can also save cash in another way: faster invoice processing means a business can more often take advantage of early payment discounts offered by suppliers.

Extracting data from invoices by using computers instead of humans became a leading use case for document capture and process automation in the 1990s. Several capture start-ups, led by Readsoft, flourished by creating software to read scanned invoices without human intervention and deliver some of those promised savings and efficiencies. As businesses moved from mailing paper invoices to emailing PDFs, capture vendors added PDF data extraction alongside scanned capture. (Paper invoices show no sign of disappearing and will be part of any solution for years to come.)

The largest capture companies today (Kofax, OpenText, Hyland, ABBYY) sell invoice processing, as do their top challengers (Ephesoft, IRIS, KnowledgeLake, and others). In our estimation, well over 100 software companies are marketing an invoice data extraction or invoice processing solution. And there are over 66,000 projects for invoice processing and data extraction on the GitHub development platform, a leading indicator of the mainstream demand for this type of solution.

Most vendors focus on integrations with SAP or Oracle, the leading financial software for large enterprises, to pass the extracted data directly into the ERP system. These solutions can be very expensive, with deals over $1 million not uncommon and professional services engagements lasting several months to years. Other capture vendors focus on the small and mid-sized enterprise (SME) market and enable financial software such as Microsoft Dynamics and Sage.

The largest barrier to entry for competition has been the enormous task of gathering the pre-existing invoice samples required to train machine learning algorithms to achieve a reasonable level of automation (typically 60%-70%) and produce a good ROI. For many years, the legacy leaders had the advantage with their training databases of hundreds of thousands or even millions of invoices. Their solutions were less about “AI” and more about matching an invoice with a template in the database. But that advantage is now slipping as new competitors using deep learning enter the market.

Four big events happened in the second half of 2021 that we believe will transform automated invoice processing from a niche market to a mainstream application used by every business. We also predict that these events foreshadow an eventual changing of the guard, with current market leaders passed by new entrants.

Rossum raised $100 million

Rossum, an AI start-up from the Czech Republic, raised an eye-watering $100 million in Series A funding for its invoice data extraction cloud platform. While Rossum’s valuation wasn’t published, in our estimation – which is in line with the industry’s thinking – Rossum’s valuation alone could surpass the total annual invoice processing revenues of the legacy market leaders (see Figure 1). Clearly, some investors believe in the unlimited market potential of invoice automation. (Rossum also announced plans to offer data extraction for all business documents.)

Rossum is a prime example of the new generation of AI-driven start-ups that are leveraging the deep learning advancements from Google, Amazon, and Microsoft to disrupt the cognitive capture market (see Table 1). Today, the availability of vast and cheap deep learning models has eroded the legacy leaders’ advantage, as start-ups can now deploy advanced deep learning models capable of learning faster with far fewer samples than the old ML models. In addition, the start-ups are using the cloud, so they can potentially offer quick and easy integrations into almost any financial software.

Microsoft entered the ring

At Ignite 2021 in November, Microsoft introduced an invoice data extraction model for SharePoint based on its Syntex deep learning platform. The Microsoft team demonstrated how a “citizen developer” can create an out-of-the-box, end-to-end invoice payment approval process for its ERP solution, Dynamics.
SharePoint is completely integrated with PowerAutomate, Microsoft’s no-code drag-and-drop RPA application, so creating the workflows and the bots is straightforward. To monitor invoice flows, PowerBI, Microsoft’s business analytics tool, adds a full-featured dashboard.

In addition to enabling its own financial platform, Microsoft is also going after SAP and Oracle through its acquisition of Clear Software, which provides deep connectivity to their systems. We’ve talked with legacy vendors who think the Microsoft solution still has a long way to go to reach parity with the existing products, citing barriers such as SharePoint library limitations and lack of transparency into the AI models. But they also expect Microsoft will get there in time.

SAP got in the game

At TechEd2021 in mid-November, SAP announced SAP Process Automation, its new low code/no code hyper-automation platform that integrates its new document extraction AI skill (based on its Leonardo machine learning platform) with its RPA bots (Contextor rebranded), SAP Workflow Management, and SAP Business Intelligence. To demonstrate the new platform, SAP issued a video showing (you guessed it) how to create an out-of-the-box, end-to-end invoice approval process using easy-peasy drag and drop.

With the integration of its powerful BPM capabilities, SAP can now offer a serious alternative to any BPM/BPA/hyper-automation competitor. The key benefit here is the tight integration of invoice processing within the SAP automation and AI ecosystem.

AWS lowered costs with an API

For those who prefer to roll their own, AWS recently introduced a new invoice processing API for Textract, Amazon’s text extraction service. AnalyzeExpense API can extract line-item details and key-value pairs from invoices and receipts. Textract uses machine learning to understand the context of invoices and receipts, and automatically extracts specific information like vendor name, price, and payment terms. It claims to work on any style of invoice or receipt, and there are no templates or configuration required.

The raw cost per invoice processed is among the lowest out there: in the US, you’d pay $0.01 per page for the first one million pages and then $0.008 per page. Also, AWS makes it very easy to try before you buy.