OCR, Agentic AI or Both?

I had the great pleasure recently to speak with Brian Shannon, enterprise account executive for North America, who works closely with Luca van Skyhawk, the Chief Revenue Officer for the US and Canada. He shared some interesting insights, from his perspective, on the market sentiment towards Agentic AI and how these compares to OCR.

Ana: You speak to a lot of prospects and customers. What is your feeling around using AI in general, and more specifically the area of Finance?

‍Brian: I’ve been working in Finance for a long time. I productized an Accounts Payable system and then an Order Management system over 14 years ago. Since 2010 I’ve seen a lot of evolution in the approach to finance workflows and the optimization of business processes. In this time, the capture component has always been lagging. Optical Character Recognition (OCR) solutions were the first attempt at solving the capture issue. In my opinion, the OCR vendors did a horrible job of managing customer expectations about what it was capable of. AI has built on the concept of OCR and provides the capability for capture plus recognition and interpretation.    

Ana:  Many organizations are using OCR solutions and the market for these solutions is still growing. What is the difference? Should I use AI instead of OCR? Or do I use Agentic AI and OCR together?    

Brian: For me, the primary difference is context. Agentic AI can apply context and interpret information. I’ll give you an example.    

We were doing a demonstration for a prospect, and they provided us with purchase order and invoice data, from accounts payable. However, the invoice they received from their vendor didn’t match the textual description on the purchase order.

The purchase order was from a steel manufacturer, so the content was complex – a series of part numbers, codes, measurements, and descriptions. I could not even decipher it. Agentic AI recognized that we should post the invoice to a general ledger account for alloys. The steel manufacturer was really surprised given that the word ‘alloy’ was not on the purchase order or the invoice.

In this situation, Agentic AI was able to review both documents, triangulate the information and come up with a logical response. The prospect confirmed it was correct.  It was impressive. OCR captures information, it does not understand the context and cannot triangulate.  

Ana: That was an impressive example? Do you have any others? ‍

Brian: Yes. Another major difference between OCR and Agentic AI is the ability to understand context based on rules that are provided with natural human language. This means that we don’t have to set up our rules (or templates) in advance to cater for every eventuality. The latter is tough to do. There will always be exceptions. We simply need a knowledgeable and capable individual to be responsible for acting as a super user with oversight capability and the ability to correct when Agentic AI needs it.  

With OCR we focus on having a template. With Agentic AI, our focus is on making sure we can validate and trust the data being processed. Various data inputs could be used to triangulate and validate the data. For example, policy documents, codes, rules and 3^rd party, external data that has been certified as accurate. All these inputs, including the catalog example I used earlier, could be fed into the system as valid data inputs for Agentic AI to reference. This is how the system was able to come up with the correct answer in the earlier alloy example.

Ana: How does the capability to triangulate and validate data impact throughput of invoicing?

‍Brian: The ability for AI to apply intelligence and perform the actions of triangulation and validation is huge. Imagine, for example, that you are configuring a multi-part item with many interdependencies such as a car. If the customer wants a sunroof, then the configuration rule might say this can’t be included in the basic package. The customer might need to purchase a different package. All the packages are defined up-front, and each possible permutation would have a master code. What happens if an error is made in the packaging definitions? Perhaps something is not compliant with a government mandate? With a manual approach, we would have to go back to step 1 and re-define the packages. With Agentic AI, we can resolve the error at the point of finding it, by validating the package against available data. We can validate the changes, apply revised rules and then continue processing. The result is significantly faster throughput. ‍

Ana: Agentic AI can automate the order management process. This is another difference between OCR and Agentic AI correct?

Brian: Yes, absolutely. The agent is autonomous and can start the order process once it has the relevant information. With OCR, you still have a manual step where someone needs to start another job or task/action.  

Ana: If I get 60% accuracy with an OCR solution, what is the benefit in getting 85-100% accuracy with an Agentic AI solution like Hypatos? ‍

Brian: There's obviously a mathematical increase of 20+% but this doesn’t provide the full picture, and you can’t calculate the value like this. The difference is the capability of setting additional rules very easily and being able to bring in additional validated reference material to provide an advanced level of context and intelligence (as I mentioned earlier). It comes down to “How much does it cost you in time and effort to do the upfront training required for OCR?” compared to the flexibility and intelligence of Agentic AI being able to adapt unaided and modify at the point of impact when processing orders – and move on.  

With both OCR and Agentic AI, you need to have your processes well defined in order to automate, otherwise you will automate garbage.

In a part two we are going to address how Agentic AI revolutionizes order management and finance processes. Wait for it!

Discover how much you can save here: https://www.hypatos.ai

‍