Speaker 1: Hi, my name is Rohit Ambarkar. I'm a Finance Director in Finance Operations, and we look at the operations that is required to support the finance function. What we will talk today is about how we have leveraged Artificial Intelligence to automate one of our process. We will talk about automating a key SOX control using Artificial Intelligence within a royalties processes. Without further delay, let's just jump into the presentation today. Here is a problem landscape. The royalties team is responsible for processing, payment of royalties for all the intellectual properties, Microsoft licenses from third-party vendors, suppliers, companies. Annually, we process about 30,000 statements. Traditionally, these statements are very high-value statements, some very low-value statements, and these are manually reviewed because we have a SOX control that requires us to review these statements and then approve these statements before payments are released. We spend an inordinate amount of time in reviewing the statements. As you can imagine, these statements require a lot of human oversight to make sure that we don't make inaccurate payments. As with any process, we do have some error here. Every time there's an error or every time there is some reconciliation problem, we delay processing these statements, which results in other problems like penalties and other things. Overall, it's a very inefficient process that we used to have. If you look at the current process, what it looks like. As with any of this, what you see here on the screen is the real workflow of how we used to process and approve and review the statements. Anywhere between five to 10 layers of reviews were performed on these statements, primarily because every mistake created a new control. There were about 10 different decision points, and about 20 control points along the way of approving the statement, reviewing the statement, and every mistake created a new checkpoint. You can imagine over the last 15, 20 years, the review function got complex and complex. We looked at this and we said, how can we use technology to augment the human process here of approving and reviewing the statement? What we did was we looked at different options that we have in re-imagining this review and approve function. What did we re-imagine? We said, look, whatever we are looking at in terms of reviewing the statement, it can fall within three or four different checks that humans can perform. Number 1 is people apply heuristics based on their prior understanding, judgment, and their expectations. They look at a statement and they verify if the numbers and the transactions on the details on the statement meet their expectations. The second, generally people apply, new people is they're given a checklist, some kind of a checklist, or the SOCKS control also defines what a pass scenario looks like and what a fail scenario looks like for any statement. Based on that checklist, people perform review function. The last one is some kind of a human intuition. Something just doesn't look right. Either it's based on the trending or based on any other conditions people may have that just doesn't look right for people. We said, look, of all the different ways in which people review this statement, we can think about certain statements that generally go through something what we call as a happy path. A happy path is low volatility, thresholds are very low, and the control works as designed. Take, for example, a fixed fee payment, that for the next 12 months, you will pay exactly the same amount because your subscription for this intellectual property is based off of a fixed fee contract. What can go wrong? It's very easy to look at different periods of payments and just look at, does it meet the requirement and does it meet the expectations? That's precisely what we captured and we said, let's capture those simple rules, simple heuristics, simple trends that you can take away. Instead of people working like robots, just doing the same thing over and over again, can we actually get robots and machine and artificial intelligence to perform and augment the human effort? We focused on what you see on the screen here. A red scenario is where something doesn't look right, and yellow scenario is something appears for the first time. We did define a set of controls, set of rules, set of trends that we expect to see, and then we deployed artificial intelligence machine learning to review every statement. Those statement that are below certain thresholds, we called it a happy path and we said, can we tag them as green if they all pass the test? From an audit perspective, just seeing that something is green is not adequate. You have to be able to say why something is green, and one of the screenshot on the bottom right-hand corner you see there is the proof that why something is indeed a green. This proof is given to the approver to be able to say, for whatever reason, if this proof is not adequate, they can override this tagging that the artificial intelligence has performed. Based on this framework, we deployed machine learning and artificial intelligence using Microsoft's own AI stack. Then we ran every statement below a certain threshold and we talk about happy path, to see that if the statements can go to the happy path. What you see on the screen is how our statement portal looks like. This is a portal where approvers, reviewers come to look at the statements and release these statements for our partner's consumptions. Upon release of these statements is when the payment happens, so this is a check before the release of the payment. On the second column where you see anomaly, that is exactly where the AI is augmenting the human intuition, human judgment. Every time you see a green, it tells you that the artificial intelligence, the machine learning has tagged that statement to be extremely low risk and passing all the rule-based and trending-based controls that we look for. Now, we have an option here, right? We completely make this touchless and we can release it for payment. And we reviewed these statements, continuously reviewed the statements just to look at if this artificial intelligence is indeed doing what it's supposed to do for six months. And after six months, our confidence was very high and then we completely made upon certain thresholds, definition of certain thresholds, we made these statements completely touchless. And every time you see a red, those statements require an additional layer of human review. So the overall benefit of doing this, as you can see on the next slide, is we have reduced about 70% of the manual hours as a result of this automation. 100% of the statement earlier, these statements were sample tested. Now, every statement goes through a very rigorous testing. And those statements that don't pass through the judgment of AI goes through the human review. And that clearly gives a reason why something is red. So thereby, the risk of inaccurate payments is substantially reduced. So what we have seen here, you can see a 99.9% of all the known errors are all eliminated as a result of this. So we are seeing a lot of benefits overall. If you look at one segment of the business, about 90% of the statements today for one segment of the business today goes through this automated path or the happy path as we call it, completely touchless and it has saved a lot of hours as a result of this automation. The beauty of this automation is just that we have tested this and only in the statement. But if you look at this framework of using artificial intelligence to augment the human judgment for anything that requires either a rule-based testing or some kind of a trending-based testing, let's talk about PO approval. Let's talk about expense approval. Let's talk about journal entry approval. The same framework that we reviewed today can be applied in very many different finance processes and we have just started scratching the surface. I hope you enjoyed this presentation and it inspired you to reimagine your own processes in a way to not only increase your efficiency but substantially reduce your risk. Thanks a lot for your time today.