A friend recently told me that their department is looking to replace their AP Automation product with a more modern one that "has AI" because her company wants to stay current and not fall behind on AI trends. I asked "What do you guys use today?" and she said "Product X". I was surprised because "Product X" to my knowledge has a lot of AI in it.
What ensued was a discussion about technology vs. use case. Should you use the most modern tech? Even if it doesn't solve a problem you have? Or should you use software that solves your problems, regardless of the technology used to do so?
Honestly, I don't care much about how a vendor manages to extract 98% of the data from my invoices. Maybe they use massive Large Language Models? Or more "classical" AI? Or templates? Or do they have a bunch of data typists keying in the data in real-time? What does it matter as long as the job gets done? Other things matter more to me, like whether I need to send data to OpenAI. Is it a SaaS product? What is the payment model? Let's leave the technology decisions to the vendor, they are the experts.
But I gave my friend a few insights to show off, next time she talks to her IT manager.
"Product X" is a typical market-leading AP Automation product that has been around for 10 years or longer. The vendor has thousands of customers using the platform for invoice automation. It has a lot of features for accountants (UX matters a lot!) and uses AI where needed, and other technology where that works better. But here is where it uses AI, while the oblivious users don't even realize it:
Yep, OCR uses AI and always has. At least for machine-printed text, even the oldest engines like FineReader, OmniPage, IRIS, Tesseract, and others have an AI model in there somewhere. OCR lends itself to AI. Typically these engines segment the image into characters and then classify them. That's only 50-100 classes (characters and symbols) and there are easily accessible samples to train models. This is a fairly easy task for AI.
More on that topic here:
If the AP Automation product requires templating, then something needs to assign each vendor's invoice to the correct template. Sometimes these templates are explicitly defined, but some products just hide the templating approach well. Either way, classification is needed. Also, invoices need to be treated differently than utility bills or credit notes, so these types are also usually distinguished through classification. And unless you are required to set up a myriad of keywords to classify documents, this job is done by machine learning, which is AI.
More on Classification here:
A viewpoint on Templating can be found here:
This is probably obvious to the AI laymen as well. Many AI techniques are being used here. From template-based fixed Extraction to fancier algorithms like Conditional Random Fields to more modern base models like LayoutLM, AI is and has been used in such products in many ways, for a very long time. Especially if your software learns from the corrections made by the reviewers, coders, and data keyers, AI is part of the game.
Many AP Automation products assist the account coding process with AI. This trend started maybe 5-8 years ago. Accountants need to select often multi-level, long general ledger codes and assign them to the invoice line items unless they are not pre-coded. This is often the most time-consuming manual step in AP Automation. Imagine invoices with 300 line items and a company with 10000 GL codes to pick from...
Machine Learning models can be trained on historical data or they can learn on the fly and can eventually suggest the proper GL code for every line item, allowing the Coder person to just review and confirm.
The above use cases are very well supported by various established AI technologies. With the rise of Large Language Models and NLP before that, many vendors jumped on the hype bandwagon and tried to find applications for NLP (remember the BERT hype?) and LLMs.
Applying Natural Language Processing technology to structured documents such as invoices never really worked well. While we see interesting applications in AP Automation for Large Language Models, many vendors implemented some quick showcase "use cases" to prove that LLMs are used so they appear modern. The value of these use cases is dubious. For example, who hasn't seen a chabot now in the invoice review screen where you can ask questions about the invoice? "What is the invoice total?" or "When is the invoice due?"... It is impressive to see that an LLM can understand the question and answer it, but why would an accountant do that while looking at the invoice and the already extracted data? Well, that doesn't keep the manufacturers from implementing it. After all, it is a nice demo.
Despite the exciting times ahead of us, I am pretty sure that you will not find an AP Automation product, however old, that has no AI in it!