AI can seem like a mysterious “black box somewhere out there!” Let’s take a look inside to explore how AI models compare to rules-based systems and see how AI is changing fintech.
It’s hard to give up what you know for something that seems mysterious and vague. To compare AI models to rules-based decisioning is a bit like comparing a rotary phone to a smartphone. We know they’re really different and we reap so many benefits from the smartphone, but does everyone totally understand how it works? No!
To clear up confusion about AI models, machine learning and legacy rules-based systems, let’s start with a definition of each term:
- Artificial Intelligence (AI): the ability for a machine to analyze external data based on a set of parameters, learn from the data and improve outputs over time.
- Machine Learning (ML): refers to systems that can learn from experience. As more information becomes available, machine learning teaches the AI model to use the new data to make better decisions.
- Rules-based systems (rules-based decisioning, business management systems): enable companies to easily define, deploy, monitor, and maintain new regulations, procedures, policies, known fraud schemes and workflows.
- AI Models: comprise a limited number of algorithms and rules that have the capability to “learn” from the data they are built on.
- Legacy systems: often references systems that paved the way for the standards that would follow; may also imply the system is out of date or in need of replacement.
Here’s a primer on how AI differs from using a purely rules-based solution.
#1 AI models use fewer rules
Rules need constant updating whenever a new behavior is detected. Algorithms seek and compare data based on a wide range of options.
A rules-based system needs tens of thousands of rules which are quickly outdated as soon as new fraud schemes, payment codes, AML rules, and more evolve. Rules are effective only when actively monitored and managed by fraud teams. Many fraud investigators spend 20-40% of their time providing feedback on rules that require updating and working with that team when they could be investigating fraud. Rules are simple and absolute.
AI uses a limited set of rules and algorithms to process data, but also accesses many other data analytics tools as determined by our unique Smart Agents technology. For example, in the payment card world, an algorithm might detect unusual behavior based on the latest fraud trends learned through screening other transactions and that card’s usual locations and transaction amounts. In healthcare FWA, an algorithm might look at a specific set of codes in concert with patient age range and frequency, then compare those submitted by other healthcare providers for similar diagnoses. Algorithms are multifactored.
Benefit: Fewer rules results in less maintenance
#2 AI models are not constrained by hard-coded rules
Rules are hardcoded and depend on certain conditions. Because rules take a binary approach – if this, then that – they depend on the information provided. This leaves room for human error and bias. We don’t know, what we don’t know.
AI models are not constrained by rules. The algorithms look at a broad range of data, comparing it to similar data in the system. Based on the solution’s goals, such as monitoring credit risk or detecting potential fraud, the algorithm flags any anomalous behavior.
Tim McBride, Director, Product Development, Cyber and Intelligence Solutions, Mastercard, uses a fishing analogy to explain the difference. In a rules-based world, we would see that all the fish eat at 5pm in a sunny spot on the lake. Our solution would tell us that’s the time to go bass fishing. Eventually the lake will be depleted of the type of fish that likes to eat at 5pm where it’s sunny, and we stop fishing. In the context of tradition detection, the fishermen are hardcoded, therefore also limited to one activity.
An AI solution would take in other factors: When is it sunny on other parts of the lake? Do fish only feed once a day? Do they only feed when it’s sunny? Are there different types of fish in this lake? By comparing data, your AI fishing app may recommend other parts of the lake where you discover fish feeding at different times of day. The AI model has analyzed multiple data points to discover new behavior.
Benefits: Models can analyze multiple data points and detect anomalies
#3 Continuous, automated refinement
Rules are static, which is why they need constant manual updating. Your information lives in silos.
Because AI is self-learning, models can use an automatic feedback loop for constant improvement. When your model is built, data scientists will input your historical data and the activities you wish to detect. AI uses that information to build a model for your particular business goals, creating a base of collective intelligence where all your data is considered. Brighterion codes our models to give the rationale for decisions; your AI is transparent and your team can easily explain what happened. Once a model is in production and its user interactions are captured, the data is collected and used to inform and fine tune the model so that it provides you more of what is good and less of what doesn’t matter.
Bear in mind that AI models do use some rules, such as for “if-then” decisions. If you are concerned that machine learning won’t happen quickly enough, you can always layer in a temporary overriding business rule for extra assurance. A rule is easily fixed or removed once machine learning is shown to be effective.
“AI gains a deeper understanding by using one-to-one data relationships, via actual user interactions, as to what is important and what is not. This feedback loop enables continuous, consistent automated refinement,” McBride says.
Benefit: Models learn and adjust automatically based on experience
#4 Scalable AI outperforms other solutions
Data in rules-based decisioning solutions have to run through thousands, even hundreds of thousands, of rules. Unsophisticated AI can use such complex algorithms that it takes hours or even days to run. This timeframe makes it very difficult to stop fraud in progress as it bogs down when the volume of data increases.
To improve processing time, some rules may be turned off but that comes with risk. If you delete or turn off a rule in a rules-based system, it won’t detect a fraud scheme that resurfaces in six months. That results in a drain on resources. Staff have to spend time investigating the fraud or increased risk, while having to comb through and decide which rules to reinstate. Meanwhile, the company needlessly loses revenue.
“A significant amount of damage can be done in 5 days,” says McBride. “A medical supply provider can bill a healthcare payer millions of dollars in a few days. They get in and out undetected, then move on to a new scheme or new audience.” AI will not only detect the behavior, but also learn any new twists to schemes.
At Brighterion, our distributed architecture allows lightning speed response times (less than 10 milliseconds), end-to-end encryption and traceability. Our AI models make decisions at up to 100,000 events per second, while delivering 99.9999% uptime as it isn’t reliant on a single server.
The Aite Group, leading analysts of financial services technologies, named Brighterion the most scalable AI platform used for anti-money laundering (AML). Aite found our platform enables more than two times the volume of transactions than of our nearest competitor.
Benefit: Our customers benefit from 99.9999% uptime with limitless scalability.
#5 AI reports in real-time
Now that we see that rules can take time to process huge amounts of data, you may also see that volume problems are also time problems. Without limitless scalability, real-time detection can’t exist.
The advantage of Brighterion’s distributed architecture is that its incredible speed enables instant analysis. We process billions of transactions per day, 80 percent of those within milliseconds.
Benefit: Transaction decisioning within milliseconds, stopping fraud in its tracks.
How do AI models compare to rules-based decisioning? Because AI’s machine learning makes multifactored decisions, AI returns higher accuracy rates, and reduces false positives and manual reviews. Just like your smartphone’s voicemail and walking app, it’s there doing its work in the background all day long – and that’s reassuring.
Learn more about Brighterion’s technology and how our AI tools work together to create high-performing, custom models for our customers.