Dr. Alok Aggarwal
Invoice processing is a critical aspect of any business. This process includes receiving, processing, and reconciling invoices and then paying the reconciled amounts to various vendors. Traditionally, this process has been manual, time-consuming, and prone to errors. However, with the rapid growth of Artificial Intelligence, businesses can now automate their invoice processing workflows, leading to significant improvements in efficiency, accuracy, cost savings, and reducing errors.
This article first discusses legacy invoice processing systems, reconciliation, and straight-through processing. Next, it discusses how to implement invoice processing by using sophisticated AI software like Collatio® (www.collation.scryai.com) for processing invoices. During this discussion, it provides the pros and cons of both systems and how to potentially compute the return on investment (ROI) if AI-based processing is used.
Legacy invoice processing solutions leverage human professionals and some automation to streamline accounts payable and invoice-processing workflows. This process involves extracting invoice data, reconciling various items within the invoice itself (e.g., ensuring all numbers add up and sales tax has been computed correctly), and then reconciling data from the invoice to purchase orders, contracts, or internal records. Finally, it involves routing approvals, processing reconciled payments, paying vendors, and recording all this data.
Four stages of typical legacy invoice processing solutions are given below:
Reading or capturing invoice data: Businesses often receive invoices in diverse formats and across multiple channels. Legacy invoice processing systems consolidate invoice data received across these disparate channels and convert the invoice data using optical character recognition (OCR) and related technologies. Since OCR is error-prone, an analyst usually checks its output entirely.
Verifying invoice data within an invoice: Using custom validation rules, legacy invoice systems verify vendor information, including address and bank account information.
Verifying invoice data with other documents using business rules: Typically, analysts flag invoices from unknown vendors, highlight inconsistent vendor data, and match invoices against purchase orders, receipts, statements of work, and other internal records.
Paying the reconciled amount and uploading all relevant information in ERP: Finally, Legacy invoice systems pay vendors the reconciled amount and export a final report containing all invoice data in a format compatible with Enterprise Resource Planning (ERP) systems or directly integrated with ERPs and accounting software via APIs.
Legacy invoice processing offers several benefits, including:
Ease of use: OCR software with a user-friendly interface usually reduces the time taken to fix errors that arise due to inaccurate OCR.
Security: Since these systems are essentially deployed on-premises (i.e., within the Information Technology (IT) Firewall of the user), the likelihood of a data breach is reduced considerably.
Customization: Since these systems are usually bespoke systems, they are flexible and extremely easy to incorporate specific needs and approval workflows.
Scalability: The entire workflow system can usually handle several million invoices per year thereby accommodating future growth.
Integration: Most legacy systems integrate quite well with existing accounting and ERP systems (such as QuickBooks and SAP), thereby maintaining data consistency.
Support: Since these legacy systems are bespoke, usually a systems integration partner or the IT department within the user’s organization can manage these systems effectively, provide reliable customer support, and follow the continuous improvement – continuous delivery (CI-CD) process easily and effectively.
Legacy invoice processing suffers from the following debilitating limitations:
Inordinate delays in processing: Legacy invoice processing systems, involving manual data entry, paper invoices, and endless filing, have long been associated with inefficiency and delays. This problem is exacerbated by invoices received through various channels such as mail, fax, and email.
Require an enormous amount of human labor: Unfortunately, most OCR software solutions have an accuracy of 90% or less. Furthermore, these OCR systems do not point out where they may be inaccurate. Hence an analyst is forced to review the entire output of these OCR systems, thereby gaining little advantage during the process.
Manually rebuilding tables and extracting data from them: Most OCR systems provide output that is read from left to right. In other words, they lose the two-dimensional aspect of the invoice, and it is usually left to the analyst to reconstruct tables containing the names of products, services, taxes, etc., which are provided in the invoice.
Prone to errors: Since these systems essentially use human analysts, often these analysts are unable to completely reconcile all invoices with purchase orders and the underlying contracts. For example, many contracts allow the payor to deduct 2% of the value of the invoice if this invoice is paid within ten days of receipt. However, often the analysts forget to take advantage of such a clause.
AI-based invoice processing represents a giant leap forward from manual processes and is driven by the power of artificial intelligence (AI) and machine learning (ML). This software automates the entire invoice processing workflow, from extracting data to reconciling invoices, and then seamlessly integrating with Enterprise Resource Planning (ERP) systems. It offers a plethora of benefits that range from manual labor, time, and cost savings to improved accuracy, fraud mitigation, and enhanced insights.
Six stages of typical legacy invoice processing solutions are given below:
1. Reading or capturing invoice data and classifying documents: After receiving invoice data from disparate channels, AI-based systems examine the format of the invoice before converting it into an electronic format. For example, if the invoice is provided as a scanned picture, then the AI system may use an OCR system that is based on Deep Learning whereas if the invoice is in PDF, then the AI system would use different machine learning algorithms that usually have a better accuracy. Finally, many AI-based systems would automatically classify each page of a document, namely, whether it is part of an invoice, purchase order, statement of work, master services agreement, etc.
2. Recreating tables automatically and reconciling them: Unlike legacy invoicing systems, AI-based systems use machine learning to automatically recreate the tables that are present in these invoices. Indeed, many AI-based systems such as Collatio recreate such tables with more than 98% accuracy. Furthermore, these systems automatically learn the formulas in these tables. For example, if one column shows the number and another provides the price of an item, then these systems learn that a third column will contain the multiplication of the first two columns. After recreating these tables, they would use such learning to reconcile much of the information within a table and across tables.
3. Extracting key-value pairs and attributes and reconciling them: Next, advanced AI systems would extract the required key-value pairs and relevant attributes from the invoices. These key-value pairs usually include data such as supplier names and addresses, purchase amounts, quantities, sales tax (if any), and the date of the invoice. If these key-value pairs and attributes are available in multiple places within the invoice, then a sophisticated AI system would reconcile them, thereby providing enhanced accuracy.
Reconciling the extracted information from invoices with other documents: In the fourth stage, advanced AI systems reconcile the extracted key-value pairs and attributes from the current invoice with those from past invoices from the same vendor. Using the formulas that these AI systems would have learned during the training phase, they would also reconcile these key-value pairs with those provided in the purchase order, statement of work, and master services agreement. Finally, they would incorporate all the terms and conditions of the underlying contract (e.g., a 2% discount should be applied if paying the invoice in ten days or 1.5% interest needs to apply per month if paying after 60 days) so that all the clauses of the underlying contract have been incorporated.
Straight through processing versus manual intervention: In the fifth stage, if all the relevant data in the incoming invoice has been reconciled, if no errors are found, if the amount to be paid is below a designated threshold, and if there are no additional clauses from the underlying contract that need to be accounted for then these AI systems will directly send such an invoice via straight through the process and the system would provide auto payment. Hence, this process will not require human intervention. On the other hand, if any of these requirements are not met, then an analyst or an accountant may need to review where reconciliation failed or where additional clauses from the underlying contract need to be incorporated.
Uploading all relevant information in ERP and other accounting systems: Finally, in the last stage, just like legacy invoice systems export all the information the invoice and payment data in a format compatible with Enterprise Resource Planning (ERP) systems or directly integrate with ERP and accounting software via APIs. In addition, they would automatically create dashboards and customized reports so that the accounts payable team can quickly visualize how much has been paid to whom, thereby potentially helping them in forecasting the spending trend for the coming months.
The advantages of AI-based invoice processing are manifold and mentioned below:
Reduces time for invoice processing and payment: Advanced AI systems significantly reduce processing time by transforming a three-week process (that uses legacy invoice systems) to one that takes two hours (if it goes via straight-through processing) to at most two days (if an analyst is involved).
Requires significantly less human labor: The time-savings due to these AI systems translate directly into cost savings because of a reduction in the use of pen and paper as well as manual labor, and it also reduces hassles for the invoicing organization. This in turn enhances employee productivity, allowing accounts payable teams to process invoices faster and more accurately. Furthermore, such advanced AI systems eliminate the requirement of growing human labor linearly with respect to the number of invoices, which in turn is beneficial regarding hiring, attrition, and training new joiners.
Fewer errors and fewer chances of missing key attributes: Because of automated reconciliation, such an advanced AI-based software dramatically reduces the occurrence of errors and substantially reduces the potential of incorrect payments. Furthermore, it helps in avoiding interest and late fees and increases the possibility of getting a 2% discount (since the invoice can be paid in ten days).
Providing deeper and more valuable insights: Sophisticated AI-based invoice processing systems collect valuable data and provide deeper insights, thereby allowing businesses to make informed decisions about vendor relationships, payment strategies, and cost-saving measures. Such software also enhances control and transparency by simplifying invoice tracking and offering real-time visibility into productivity.
Ultimately, advanced AI-based invoice processing is a catalyst for business growth. It equips organizations to handle a growing volume of invoices without an immediate need for expansion, thereby freeing up capital for strategic investments. By making AI-based invoice processing a cornerstone of modern invoice management and optimizing the invoicing process, businesses can navigate their growth trajectories with greater control and confidence.
Implementing sophisticated AI-based invoice processing involves several key steps:
Assess your current process: Analyze your existing invoice processing workflow to identify bottlenecks and inefficiencies.
Include stakeholder input: Engage key stakeholders, including finance, procurement, accounts payables, and IT teams, to gather input and address concerns.
Establish appropriate workflows: Define the workflow that aligns well with the workflow of your current legacy invoice system and with your organization’s policies and procedures. This would help users in getting appropriate buy-in from various stakeholders.
Compare various AI-based systems to determine their advantages: This is probably the hardest step in the entire process but needs to be done wisely. Unfortunately, currently, AI is hyped to the extent of being magical and most AI companies are unable to live up to this hype. Hence, users should compare various companies that are providing sophisticated AI-based systems and determine which one works the best with respect to their chosen key performance indicators (KPIs), e.g., cost, timeliness, accuracy, quality, and ease of use. Fortunately, many AI companies provide their AI-based solutions using SaaS (software as a solution) systems for users to test their systems for little or no cost. Users can use these systems to compare one against another. Of course, if an AI company does not provide such a SaaS system, users should do a “bake-off” among various solutions in the market, which unfortunately some AI companies do not agree to. In any case, it is vital to conduct thorough testing to address any issues and gather feedback from internal stakeholders.
SaaS versus On-Premises: Many AI companies only provide SaaS (software as a service) solutions where all the user’s data (including invoices, purchase orders, and contracts) go to the AI company’s IT system and where it is processed. Many of these companies would use this data to train their AI systems further. Hence, users should determine upfront whether their compliance department would allow their data to leave their premises or allow this data to be used for training the AI system, which can be used for others. Indeed, many organizations especially in the banking, finance, insurance, healthcare, biotech, and pharmaceutical verticals have strict controls regarding their data leaving their premises, and this is partially dictated by governmental rules and regulations. In such cases, users would need the AI company to deploy their solution on-premises and within the users’ IT Firewall and the company may not be allowed to train their AI system using users’ data.
Upfront investment should be low: Most AI companies have already provided a pre-trained AI-based invoicing system that is ready to be installed. Hence, in this regard, they are no different than traditional software companies. Furthermore, if these companies have a SaaS (software as a service) model, they will allow users to test their systems for free. Hence, the users are not expected to provide any money upfront, except for the licensing (or the use) of this software and for its installation on-premises (if needed). Usually, such on-premises installation costs run between 10,000 and 25,000 US Dollars and the license fee is a combination of a fixed fee per year plus the number of pages or documents that are fed into the system. Since these license fees are charged on an annual basis, users can terminate the contract after a year (without incurring much additional cost).
Train the users and analysts appropriately: Provide comprehensive training to ensure your team can effectively use the new system. Many AI companies will provide such training at very little additional cost.
Evaluating the success of an AI-based invoice processing system involves tracking key performance indicators (KPIs), including:
Overall, to compute the return on investment (ROI) by using a sophisticated AI-based invoice processing system, it behooves the users to compare the benefits related to the above-mentioned KPIs to the additional costs (in software, hardware, training, etc.) that are incurred.
Blog Written by
CEO, Chief Data Scientist at Scry AI
Author of the book The Fourth Industrial Revolution
and 100 Years of AI (1950-2050)
