Just how close are we to machine learning in Internal Audit? The better question is, “How much money and time do you have?”
Big Data. Data Mining. Machine Learning (ML). All of these buzz words communicate the same thing: Artificial Intelligence (AI). We often don’t take notice of how AI is in the fabric of our lives because we’re used to it. From spell correct to online purchase suggestions to facial recognition, we’ve all experienced AI at work, learning our habits and patterns, and we’re relatively unworried about it. Individuals and companies benefit from machines learning our patterns and predicting what we’ll do, need, or want next.
Misconceptions abound around artificial intelligence. When us “common folk” hear the word “artificial intelligence,” we take a bullet train ticket to the sci-fi world, equating AI with human intelligence and decision-making. But a machine thinking like humans isn’t even close to what AI actually is.
Rather than robotic humanoids or machines who have become “self-aware,” AI might be better described as computer systems that can predict human behavior. For internal audit, AI can be a handy tool for specific processes within audit and analyzing overall sets of data with greater accuracy and even predicting risk (but not judging risk, which we’ll dive into in just a moment).
Andrew Clark, Principal, Machine Learning Audit at Capital One, is one of the few AI whizzes in the audit industry. “A lot of the biggest gains for audit will actually come around robotic process automation, essentially a more intuitive way to create programs that automate routine tasks. No machine learning or artificial intelligence required.”
Mr. Clark has already built several systems that use AI technology in audit, but for specific tasks only. When you talk to Clark, you get a more realistic view of what AI is. AI/ML isn’t tomorrow’s human android, it’s a machine that learns how to process data. Contrary to the marketing hype, AI is helpful but sometimes slow and fraught with inaccuracies.
Humans think, machines learn
Machine learning gives time to let people do what people do best – make decisions. A machine makes conclusions based on algorithms fed to it, which can help to predict where risk could exist in a company; however, the machine can’t make proper judgments decisions, like how to assess risk level or mitigate a risk. “These are very judgmental and ‘thinking’ capabilities,” says Mr. Clark, “not something that a machine, no matter whatever math is powering it, will be able to take away.”
What AI can do, is reduce the amount of noise that auditors have to wade through before auditors can assess and draw valuable conclusions. AI algorithms can be created to compute larger test sections, or evaluate an entire population (rather than just a sample), or classify checklists. AI will eventually allow audit to be more effective, just by providing more coverage and consistency.
Why you won’t see AI in your shop tomorrow
Here’s the drawback: the process for machine learning is long and expensive and requires minute attention to detail. For example, in Andrew Clark’s article, he writes about an ML classification model that was designed to classify huskies from wolves. But the test data used to generate the algorithms presented inaccuracies. The test data used to train the machine showed all pictures of wolves with snow in the background; as a result, the machine incorrectly classified huskies in snow to be wolves.
This example tells us that the minutest details of test data must be accurate, and at any point, the things we fail to notice when creating test data can generate inaccurate conclusions from machines.
But let’s get back to internal audit and AI. We see how machine learning has broadened the playing field for what’s possible. How does internal audit even begin to take a bite?
Companies interested in getting on the AI path must know where to start first. Think of AI as the top of the pyramid, not the bottom. According to Clark, to effectively use AI/ML, companies must have a variety of foundational tech elements before they can consider having a successful AI rollout. These elements include a lot of other less “sexy” data (see the data science hierarchy picture), software development, and talent development (including a big checkbook because data talent is pricy).
As with the Maslow’s hierarchy of needs, it’s easy to try to jump to self-actualization before having the more important things, like safety, in place. “The same goes for companies and their AI journeys,” says Clark. “Until you have spent years supporting and developing your data acquisition, cleansing and skills development, the business won’t even have the data available to fully use these technologies.”
Although your company will probably not achieve self-actualization/deep learning/AI for the New Year, having a blueprint for where to begin is helpful.
So everyone starts at the bottom of the pyramid. Focus on the data needed and what’s available in the company. This points to creating effective data analytics. So much of machine learning can come about once all the fundamental stuff is in place. Know how to use ACL and do it well; take classes in Excel, and get the most out of your current software.
Once AI becomes commonplace in the profit-generating lines of business for companies, then internal audit will have a shot at AI/ML too. Although AI has been around for 50 years, AI/ML is still finding its function in a current and changing world. For now, Silicon-based goliaths, like Google, Amazon, and Apple, will continue to pioneer AI, and more companies will jump on board and take algorithms for a ride.
Want to learn more about machine learning?
- VisualVisual understanding of how machine learning works.
- An overview of sample ML [SG1] algorithms (slightly deeper dive).
- An easy understanding of the data science hierarchy is explained here.