Podcast Episode 3: How Machine Learning for AP Automation Works

Henri Rove, product manager for Dooap, an AP automation company in Austin, TX describes:

• The point of machine learning in accounts payable
• What does machine learning tangibly look like?
• How the business cases for machine learning in accounts payable were discovered
• Approach for designing the machine learning features
• How will machine learning for accounts payable evolve in the coming years?

Brent Hametner (00:00):

This is episode three of the AP automazing podcast where we talk about all things accounts payable automation. Hi everyone. This is Brent Hametner, your host for the AP automazing podcast. Today. I'm happy to be joined by Henri Rove product manager for Dooap, an AP automation company based in Austin, Texas. Henri, thanks for joining us.

Henri Rove (00:26):

Hi Brent. Thanks for having me.

Brent Hametner (00:29):

So the topic of today's conversation is on machine learning for AP automation and specifically on how it works. But before diving into the details, I was hoping you could briefly describe really what the point of machine learning is for accounts payable automation and why should AP staff care?

Henri Rove (00:49):

Right. So basically when we do invoice handling it usually has a lot of repeated tasks so we need to kind of adhere to the accounting standards and make sure that our invoices are posted in the same financial dimensions as last month's invoices were. We also want to make sure that the right people get to check the invoices so that the vendors are charging us what they were supposed to. So, when we first set up an invoice handling process and we are starting up the business or just getting into digital invoice handling, we usually do it somewhat manually, kind of one invoice at a time. And, it's really dependent on the AP person to kind of check each invoice, input the coding and select the correct workflow, for the people to review and approve the invoices.

Henri Rove (01:55):

And once the volume starts growing and we get more used to the AP process, we probably want to start setting up some rules that we can utilize in the invoice handling, such as vendor default main account default approvers and things like this. So a rule based automation, really good thing has a lot of benefits and Dooap actually has a really solid structure to kind of support these rule based automation in a lot of ways. But the downside for this is that it's kind of a big job to maintain the rules. So what we actually can do the next step of a modern AP process that all the big enterprises are moving towards is using machine learning. So it's more of a hassle free way of setting up the invoice automation because you can do it without the maintaining of roles. So machine learning can really mean a lot of things and it's kind of a fancy term, but what we mean by it is that we take all of the historical invoice handling data we tell the machine to look for relationships between the input and output data, and then we use that data to predict how the next invoices are going to be handled.

Brent Hametner (03:37):

Ah, interesting. So, yeah, and I would guess that most people right now get really the I guess the main idea of machine learning and where an algorithm is used and as you mentioned, your data gets fed into it and as more and more data is fed into the algorithm, it learns over time and is able to make increasingly more accurate predictions. But what I'd like to do is pull out the magnifying glass for a bit and understand really what's happening beneath the hood. So to start, what does this algorithm actually look like? Is it in the form of code that's written on a server? How does someone even begin figuring out what the algorithm is?

Henri Rove (04:24):

Yeah, so our machine learning algorithm brains live in the same place as Dooap. So it's in the Azure cloud as a server less function. So there's a couple of reasons why we chose this. Firstly it's in the same ecosystem as Dooap data. So no information leaves that Azure cloud at any point so your data is safe there. Secondly, it provides us the processing power that we need for all of the data analyzing that we need. So how the learning starts in the Azure cloud it's basically the machine breaks down the invoice handling information into inputs and outputs. So inputs would be basic invoice information such as company, vendor, invoice amount, order information and things like this that you can read on the actual invoice and which can be then scanned into digital format.

Henri Rove (05:36):

And the outputs, they would be kind of the end results that we're looking to get. So how was the invoice handled? How was it coded on which financial dimensions and who reviewed it and who approved it. So this is what we're looking for the machine to tell us, taking the inputs and giving us the outputs. And to do this, the machine learning takes about 70% of the available history data and examines the different combinations that you can have and tries to find these statistical relations to kind of figure out how the invoices were handled. So let's say we have a vendor that supplies a couple of different departments at our company and these invoices are coded a bit differently. So the machine, you could use a simple model for example, to determine that invoices over 10 grand always go to the production plant. And invoices around 500 bucks or so go to the I.T. Department. But this is a pretty simple guess and we sometimes want it to be a bit more educated so we can allow the machine to use a little bit more time and to find out what's actually the determining factor there.  

Henri Rove (07:15):

So it could probably even find out that when the invoices go to the I.T. Department they are usually sent out on a Monday or that the guy that's always sending us the invoices is called John. So the thing there is that it actually doesn't matter what the actual data is, as long as there is some data that's repeated on the invoice and has some type of connection on how the invoice was handled, the machine can and will find it.

Brent Hametner (07:50):

Great. Great. So that's helpful in understanding the machine learning in general for AP. I'm curious now to know how you started with machine learning. I would imagine the beginning of the process would be to start with understanding the business cases first and then developing the capability itself. I'm curious how did y'all go about discovering what those business cases were?

Henri Rove (08:19):

Yeah, so that's a great question. So what we did in the beginning was take an actual company's invoice handling process and started investigating the data, also what the AP professionals actually do. So we wanted to know what kind of decision making is involved and what kind of details their coding selections are usually based on. So what we quickly realized, and pretty much anyone who's done this kind of work knows that this can be based on even the smallest details and completely unrelated to the contents of the invoice. So for example, an invoice could come in with a colored logo and a day after that we get the same invoice, but with a black and white logo, this tells us that something's different on it and we handle it differently. So it's something you, the AP professionals pick up over a long time years of experience and basically do this because it's more efficient this way and gets you home early. So what we wanted to do with machine learning is basically simulate this behavior but with the data. So it means that we're not looking for the text I.T. Department on the I.T. Department invoices. So if the vendor could provide us that information, that would be great. But in most cases it's harder than that. So we are looking for something, some data that differentiates these invoices from other invoices such as John sending us the invoice on a Monday, which tells us it's for the I.T. Department.

Brent Hametner (10:10):

Gotcha. Gotcha. So then you discover the business cases and then the next step then would be to design the features around that, that machine learning tech. What then was your approach for the design?

Henri Rove (10:24):

So our design was we wanted to make the user experience as simplified as possible. There are a lot of powerful algorithms under the hood. For example every time you make a change in the invoice data or coding suggestion and probability numbers are updated so that you always have the option to let the machine fill in the rest of your thought. So we wanted to kind of convey this to the users in a way that's not threatening, so they feel comfortable with the machine, kind of helping them with their own coding process. So you can always choose to use the machine learning suggestion, but you're not forced to do it. And every time you make a decision by yourself or from a suggestion the system will know what you did and add this to the learning data. So tomorrow this decision will be included in the machine learning suggestions.

Brent Hametner (11:32):

Okay. Okay. So we've now talked about the machine learning from the design perspective, the tech perspective and business perspective. I'm curious to know how you see machine learning for accounts payable evolving here in the next few years.

Henri Rove (11:50):

So there's currently quite a bit of buzz around the topic and I would personally predict that in the next couple of years there's going to be a huge increase in machine learning in the large enterprises mostly due to decision making that is driven by efficiency. So it will be interesting to see how smaller, midsize enterprises actually take this new technology in, but something we will learn.

Brent Hametner (12:27):

Gotcha, gotcha. Great, great. So, you know, thanks for this information Henri and thanks again for joining us. This concludes episode three of the AP automazing podcast on how machine learning works for AP automation. And if you'd like to learn more about Dooap, visit www.dooap.com. Thanks everyone!