Q&A: What is Adaptive AI?

Adaptive AI is the next great advancement in leveraging AI for credit risk. Equifax's Chief Innovation Architect, John Fenstermaker, developed the product based on his customers' struggle with making optimal decisions in a ever-evolving business environment. In this interview, John explains the benefits of Adaptive AI. 

Let’s start with the basic question: What is adaptive artificial intelligence?

Fenstermaker: Adaptive AI solutions are models that self update over time with new data. With Adaptive AI, Equifax puts a self-updating modeling process -- as opposed to a static modeling process -- into production. This ensures that models are fully optimized as the economy and consumer behaviors change over time.

Which industries is this technology most applicable to?

Fenstermaker: Adaptive AI solutions are applicable to a variety of use cases, including fraud, risk, and marketing across many different industries like banking, lending, telecommunications and insurance. While they already leverage Adaptive AI solutions in areas like fraud and marketing, Equifax’s Adaptive AI solutions are unique because they can produce regulatory-compliant credit risk models. This is groundbreaking in the area of credit risk. 

What are the main benefits of Adaptive AI for the Equifax customer?

Fenstermaker: First, financial institutions typically develop credit risk models every 2-3 years. After developing and deploying credit risk models, financial institutions monitor their performance. They look for signs of model deterioration; if found, they redevelop models with newer data. New model redevelopment and deployment can take up to 6 to 12 months, which means financial institutions will use a suboptimal model for another 6 to 12 months. Whereas with Adaptive AI, new models are developed each month. Therefore, lenders and financial institutions can instantly deploy a new model that significantly outperforms what is currently in production. As a result, an optimized model is always in production.  Secondly, Adaptive AI enables financial institutions to deploy better models when no deterioration is detected using the traditional approach. If a model shows no signs of deterioration, a financial institution won't know if there's a better model to deploy. In many cases, models that show no signs of deterioration perform far worse than an optimal model, but lenders don't have that knowledge without Adaptive AI. Finally, the traditional model is developed for use across all seasonal periods -- and this is a negative. In the wireless phone industry, for example, applicant populations vary by season. The applicant population skews toward prime credit scores in the November through December holiday shopping season; it skews toward subprime scores during the tax refund season. With Adaptive AI, the modeling process can be designed for seasonality.

What about the consumer?

Fenstermaker: When a model does a better job of differentiating applicants who will pay their bills from applicants who will not pay their bills, lenders can approve more loans without taking on more risk. As a result, more consumers get access to credit and better offers. In one study, Equifax found that with Adaptive AI, one lender could increase approvals by 130,000 every two years. That's a 2% increase, without taking on more risk. These incremental approvals will yield almost $16MM in profits for the lender every two years.

Can you provide a scenario for a few industries that illustrates how Adaptive AI would work?

Fenstermaker: Each month, whether it's for telecommunications, credit cards, auto loans or personal installment loans, we obtain one new month of credit applications with performance information (typically 1 – 2 years of account performance after application) for those applications. Adaptive AI solutions take the additional month of data, combines it with other credit applications, and updates the model. This process repeats with each new month of applications with performance.  In marketing and fraud, feedback is much more immediate because we don't need to wait 1-2 years to obtain performance for new applications. We obtain campaign performance information within days or weeks after campaign execution. This allows models to update within weeks or months of campaign execution. Manual reviews identify many types of fraud during the application process. And those not identified are usually found within the next few months. As a result, there is more immediate feedback to drive model updates. Marketing and fraud models can update more frequently than monthly, depending on how fast feedback is generated. Fraud in particular will benefit from Adaptive AI. Fraud schemes are constantly evolving because as models catch up to fraudsters, fraudsters modify their schemes to go undetected.

Watch: Equifax data scientists discuss Adaptive AI for credit risk.

Is there human oversight of Adaptive AI?  

Fenstermaker: Yes, there has always been some level of human oversight of Adaptive AI solutions. That's especially true in the highly-regulated credit risk space. However, the oversight is far more streamlined than with traditional, static models. Because of regulatory requirements, Adaptive AI solutions for credit risk incorporate compliance, legal, and model risk management (MRM). With traditional, static models, developers do their best to build compliant models that will pass legal, compliance and MRM reviews. But legal, compliance, and MRM manual reviews often discover issues that require adjustments. The reviews themselves are time-consuming, and any rework adds significant time to the development process. With Adaptive AI, legal and compliance collaborate in the model development processes. Therefore, developers are building “guardrails” into the Adaptive AI process that restrict the universe of possible solutions. Model developers also collaborate with the MRM team to design the Adaptive AI process in a way that ensures MRM’s approval of each model update. Due to this collaborative approach, detailed manual reviews are not needed at the end of the model update process; instead, legal and compliance simply review a report showing appropriate guardrails were applied, and MRM simply does a high level review of the execution. In addition to legal, compliance and MRM, human oversight determines whether to deploy updated models. During some months updated models do not yield significant performance improvement. A risk manager will review performance of the updated model versus the existing model, as well as other documentation and reporting before deciding whether to deploy the updated model. 

What does the future of Adaptive AI look like?

Fenstermaker: Financial institutions are starting to leverage Adaptive AI-like solutions for less-regulated models like fraud and marketing. I think we will see widespread adoption of Adaptive AI-like solutions in less-regulated areas within the next five to ten years. As banks and other FIs become more comfortable with Adaptive AI-like solutions, you'll see them begin leveraging Adaptive AI for more regulated solutions like credit risk.

Read more on this topic in our e-zine, Summer of AI.