MindBridge Ai Technology, Mission and Product Overview in Twenty Minutes
Join Robin Grosset, MindBridge Ai CTO as he talks about the MindBridge Ai Auditor technology, why we created it, how we take a hybrid approach cross-correlating business rules, statistical methods and machine learning and why AI is an integral part of modern accounting.
Good morning, everyone. My name is Robin Grosset. I'm the CTO for MindBridge. And today, what I'd like to do is share with you the MindBridge Ai Auditor technology. I'd like to talk to you about, you know, why we created it, what it does and how it works. And most of this presentation is going to be a walkthrough of the product itself, the technology capabilities, and what you can do with it.
So let's get started. So why did we decide to build an Ai Auditor technology? So, data shows that we're actually quite bad at detecting financial irregularities. There's a report that's produced every two years by an organization called the ACFE, the Association of Certified Fraud Examiners. It's called Report to the Nations, and in it, they quantify the financial loss that occurs related to human error and fraud and the things that we know, the things that we are catching, which become public knowledge.
These are the things that are above the surface. They put a number on that, which is 200 billion annually that we're losing. What's perhaps more interesting is what we're not finding. The estimate that they make of the things that go undetected and are happening, again, all around us, and annually, the number that they put on the loss related to that is $3 trillion.
So to put that number in perspective, $3 trillion is about $500 for every person on the planet. So it's a big issue. MindBridge's purpose is to make the detection of the 200 billion faster, easier, quicker. And the intent of that is, if you can find something faster, the damage is less, and you can improve accounting practices.
But further to go under the surface, to start to, kind of, dig into finding some of this $3 trillion and being able to return that to the economy. So the data shows that, if you're a fraudster today, the most likely way that you get caught is by someone calling a tip line or you making a mistake. It's not by a concerted effort by a machine or a system that's monitoring.
It's actually more like happenstance that people get caught. So that's quite a scary statistic. The other one is that analytics is not all that effective in this spot. Only 3%, roughly, do we see fraud being caught through, kind of, analytic means, and that's a fairly poor number.
I've spent many years building machine-learning analytic products, and 3% effective is very low. You're expecting 60%, 80%, or 90% effective. The other thing is the incidence of fraud are increasing, and the time it takes us to detect from when it starts is also on the rise.
It's now 32 months, which is multiple audit cycles. Now, these numbers confuse me, and particularly, the effectiveness for analytics in audit being so low. And as I dug into it, I started to understand why it is that we're not finding much using existing approaches.
And there's a number of reasons for that. One of them is that the predominant way that people audit and what they're looking for when they're auditing is to take a sample of the accounts. Not to look at everything, but to take a sample. And the reason for that is that auditors are human beings. They can't be expected to look at every single transaction.
Particularly, with the data volumes that are going on today, really, it is a big data problem. And a human coping mechanism for that big data problem, when a human being can't look at everything, is to look at a sample, verify that those are okay, and then argue that the rest of the population is going to be the same. Now, the problem with that approach is that you're looking for something that's irregular.
You're looking for something that's an anomaly. And the idea that you would find something like that through a sample, it's not very likely. It's actually about the size of the sample is about how likely you are to find that, and coincidentally, that's between 1% and 3% in most scenarios in most audits today. So this is an issue. And really, what we wanted to do in MindBridge was to see if we can construct an AI Auditing system that could look at every transaction on behalf of the auditor and determine where the higher risks are and to surface those to the users.
That's what we've done. Now, the interesting thing is that...so this is a statement from the audit standards, where...and it talks about reasonable assurance, "The auditor is supposed to provide reasonable assurance that the financial statements are free from misstatements, material misstatements." Now, the interesting thing is the sampling approach is entirely reasonable when you consider that it's a human being who has to perform the audit and they can't look at everything.
But in a world where we have AI technology, where we have machine learning, and we can train a machine to look at every transaction and apply some degree of human intelligence, what's reasonable changes? So it's no longer reasonable, I would argue, that the people sample. We should be using technology to look at every transaction and determine where the highest risks are.
So the approach that we use at MindBridge, we label it as a hybrid approach. We use multiple different techniques including domain expertise with business rules, statistical methods, and machine learning to detect different patterns of irregularities. It's a very robust system for finding unusual circumstances in accounts.
And the combination of these different techniques makes it very resilient to being gamed. It's actually quite hard to fool a system that has multiple techniques in it. So the dotted line in this diagram is supposed to differentiate between simple rules-based systems, which is the convention for computer-aided audit tools.
They typically are built from rules. And what MindBridge is doing...so this is really how MindBridge is different. Our ability to provide machine learning capabilities, whether they're looking for unusual transaction patterns and unusual account communications. It really differentiates us. And it's actually where a lot of our value is. And what we've found in testing this in the real world is that the machine learning approaches are really much better, usually over an order of magnitude better at detecting irregular issues.
And why are they so much better? Well, one of the things is that the business rules that are typically used in audit are well known and anybody who's trying to exploit a system that is rule-based is going to learn the rules and then find ways around them. So rules are only going to be so effective, and they produce a lot of false positives.
So this idea of applying machine learning and a system that can learn to that problem means that you're going to get less false positives. It's going to be much more effective at highlighting what you're interested in. So we're now going to go to a demo. So let me just drop out of Presenter Mode, and I'll log into the AI Auditor System. So this is a cloud-based system, so onboarding is very quick.
If you decide that you want to try the AI Auditor, that can be arranged in the space of a day or so. There's nothing to install, we simply set up a tenant for new clients in our cloud, and they can get started right away. So this is a system, this is our demo system where we have multiple demo clients here.
And what I'm going to do is I'm going to show you the creation of a new client. So here I am. I'm an external auditor. I have a new client, and I want to pull in some data about that client in order to start my analysis of their finances. So, the system has realized that I don't have any data about this client and it's inviting me to pull in a spreadsheet or a CSV that's been exported from their accounting system or I can connect directly to a cloud-based accounting system.
I actually have a ledger that's been given to me. This is a demo ledger that I'm going to load into the system. And you'll see here it goes. One of the things that we did when we built our AI Auditor was we actually asked people who work in this field, you know, "What are the challenges that you have?"
And one of the things was actually just loading data into the tool can be a challenge, making sure that it actually is the right shape to work. So one of the things you'll notice is, as I drag and drop that file, the system actually detects what the file format is. And if it's a common accounting system, it should detect that. But you'll see here it's detected that it contains journal entries with debits and credits.
So our system actually figures out what the format is, and then also how to load it. So the audit professional is not spending time shaping the data to load, and this just removes an impediment there. So what it's doing right now you see it says, "Analyzing and Learning." So this is running all of the control point checks in our system. So there's a lot of them. I'm going to show you them in a second once this analysis is complete.
But the types of things that it's doing are comparing every transaction to every other transaction in order to understand what are common practices for this entity versus not. And this is one of the really nice things about machine learning, is that it can calibrate itself to a given entity and discover what are normal or not normal for that given company, so what's sticking out and what's unusual.
Now, another interesting thing about our platform is that we create a lot of data about your data. So I think in the system, I'm going to show you about 27 different control points. So for every entry in every journal, we're analyzing that and we're creating a lot of data about the data.
So if you give us, let's say, 100,000 transactions, we'll create over a million data points about that data. And that's actually how we construct our quantitative risk-based score for every transaction. So this is why the process takes a little bit of time, usually a few minutes. And you can see that, just as I said that, it's completed, and I can click here to see my results.
So this dashboard and all of its analysis has been constructed while we were waiting, so from the time that I dropped the file to getting here. You see what we've done is we've looked at every transaction. The system has looked at every transaction, and it's created a risk-based score for it, and then it's also stratified the risks into high, medium, and low. So what you can see here is that, if I hover over low risk, you'll see that there are several thousand transactions in this low-risk category.
The dollar value associated with them is 34 million, and that represents 98%of the ledger. Interestingly, you'll see here there's a number of medium risk, you know, only 50 or so, and there's a few high risk, and you see the dollar values associated here. So, what we see is the role of MindBridge, in this case, is that to focus the audit professional on where the most unusual things that they should look at.
And this gives them confidence that, you know, MindBridge has looked at everything and it's found the things that are most unusual. And I can click on any of these categories and drill into the detail of what those high and medium risks are. But let me just go around the dashboard a little bit and show you around. So here we're seeing a breakdown of risk by month, and you can see if there any trends there.
I can also look at it by week and by day. Some of the other interesting things I can do, I can say, "Okay, let's not look at the low risk.It's not so interesting." And when I do that, one of the things that, kind of, starts to stick out is that there's definitely a spike that happened at a particular period of time in those medium risks, so I may be interested in that.
And you can see I can zoom into a particular timeframe and look at that. And if I click on this bar, it will actually take me to the transactions that happened on that day, which we can do that in a second. So you see at the bottom of the dashboard, this is something that we call a risk burst. So this is a view of the entire chart of accounts for this entity.
So you see the rings are basically levels of the chart of accounts. You see the innermost ring is assets, revenue, liabilities, expense, and equity. And as I go further out, essentially I'm going down the chart of accounts into more and more detail. So you'll see over there's something in that kind of asset space. There's some equipment assets that the system has identified.
It looks like the transactions in that category seem to be a, kind of, medium risk. And there's also over here, in the expense area, it looks like I have write-off expenses. And what's this? Management salaries and directors' fees that do seem to be also triggering the system to think there's a higher risk in that area. On the side, I can see all of the different test failures that have occurred.
So there are, you know, 500 weekend posts. There's a number of reversals that were flagging here, things that contain suspicious keywords. And again, I can click on 76 manual entries. I can click on these. I'll do that now. And it takes me to the list of manual entries, and I can take a look at the transactions in some degree of detail.
So let's look at one of these, let's pick this one here. You'll see these are each journal entries and dates. And I can actually expand them, and I can look in the details of each one. So again, this is a demo ledger. So this data is...it has all been created for the purpose of the demonstration. So, there are several things that we've, kind of, seeded into these transactions that are are going to trigger the system.
So we've given this transaction, which you can see up here, 52% overall risk score, which puts it in the high-risk category. And everything that's read on the screen is a control point that has triggered for this particular transaction, this particular journal entry. And you can see also all the things that have not triggered. All the tests that were run that, you know, this was not reversed as an example, and it's also not a duplicate transaction.
But the sort of things that are triggering it is, you know, it's a high monetary value. Given the transactions we're looking at for this ledger, it's a high monetary value. It was manually entered. It does seem to be a reporting adjustment, so it seems that there was another transaction related to this at a future time or this one was reversed the previous one. It says, "Job Cancelled." So this transaction reversed the previous one, which contributed it to revenue.
So it triggered a reporting adjustment. Again, reversal. It happened at the start of a period and the start of a year. It contains suspicious keywords. You see, it says, "Cancel." And the other control points that are triggering here are some interesting machine learning ones.
So this is a rare transaction flow. So it's unusual for this entity. It also triggers our expert score control point. So you'll see here, the description here is, this transaction was identified as high-risk by a domain expert, it gets a 67% score, and that there are 128 transactions that are identified as high-risk by the expert. So how did we build an expert system to do this?
Essentially, what we did is we worked with audit professionals in building the system. And we actually learned how they work with journal entries, how they think about transactional risk, and what they are looking for. We even went to the extent where we gave them a survey, asked them to score transactions for us.
And this allowed us to construct an expert system which can look at every transaction and apply the same kind of thought process. It also triggers flow analysis. It's also an outlier. And these are machine learning techniques which are triggering for the transaction. Some other things I can do here is I can look at the detail of each and every line of the transaction.
And if there were more fields in the file, they would appear here, so I can get all the way back to the original journal entry. You'll see also that as I click on the different lines in the journal entry, you'll see different parts of the transaction are lighting up. So, for example, not all of the transaction is triggering the expert score.
It's just these two lines here that seem to be unusual. So again, it will look at the entire transaction, but it will flag which parts that concern for a given control point failure. So in this view, you're actually looking at multiple different machine learning techniques as well. We have unsupervised learning and supervised techniques at play here.
And the last thing I wanted to show you here in the Control Point Views. Let's say, I've decided that this transaction is something I want to follow up on, I can flag it as suspicious just by clicking this button here. I can say, "Hey, this is a weird transaction." I can ask one of my colleagues, Jim, to take a look at it. And, let's say, perhaps I'm very interested in the tax allocation here because I don't think it's quite right, because the expert has flagged it.
So I'm going to get Jim to verify that it's right or not. And in this follow-up, he can go then talk to the client and figure out what happened. But one thing that's interesting about our system is just this interaction. Just flagging this transaction for follow-up is a signal to our AI system. And from the data that we capture from this interaction, our system can actually get better.
It can find out what are the things that people are actually interested in looking at for a given entity and what are the things that are not interesting. So the "I'm not interested in this, stop showing it to me" is this kind of "flag as normal" button. So this is...again, interesting thing about machine learning and AI systems is that they get better with use. So unlike a rules-based system, they actually evolve and they can reduce the number of false positives you're looking at, which is one of the big problems with rules-based systems.
So this is a quick kind of whirlwind tour of the AI Auditor. So we're very happy to follow up and go into more detail with you. So that concludes our demo for today. Thank you for your time. And please feel free to reach out. Send me an email if you have any questions, I'll try to answer them.
You can also sign up online for a demo and to get access to the system. Thank you very much for your time. Have a good day.
Demystifying Artificial Intelligence and the Impact on Auditing
AI is helping of a number of professions, including auditors, deal with a changing landscape. Learn how the Ai Auditor platform leverages AI to provide faster, more highly assured audits than ever before.
Interview with Becky Shield of Kingston Smith
Kingston Smith, a top 20 UK chartered accountancy and audit firm is leveraging MindBridge’s Ai Auditor. Becky Shield talks about the value that Ai Auditor brings to the firm.
How AI is Disrupting Audit and Introduction to Ai Auditor
Learn about the history of AI and why it is emerging now and how AI is currently disrupting audit, followed by a demonstration of the Ai Auditor platform.