Artificial intelligence is clearly here to stay, but what’s less clear is how it’ll impact financial risk management. In our most recent webinar our CEO, Vinodh Poyyapakkam, sat down with David Snitkof, VP of Growth at Ocrolus, and Ryan Hildebrand, EVP & CIO at Bankwell, to discuss how AI might impact risk management strategies.

Check out a summary of the conversation and link to the recording below.

‍

AI Opportunities across the SMB customer journey

There’s an increasing amount of SMB data in the world, but most of it is unstructured and difficult to consume in a programmatic way. Risk teams have historically relied on manual analysis to glean insights from this data, but recent progress in AI has made it easier to aggregate unstructured data and generate insights systematically.

‍

‍

Pre-qualification & Onboarding 

Automated document classification and pre-filling data can help teams cut through the noise and speed up the pre-qualification and onboarding processes exponentially. 

Meanwhile, unstructured data extraction helps generate insights from multiple data sources quickly, while maintaining the accuracy of a human risk analyst. Automated merchant industry classification is one use case for this, and customers using Merchant Real Industry have noticed significant time savings by switching to this LLM-powered process.

‍

Underwriting

Fraud has gotten tougher to solve with deep fakes, but AI can be used to fight that much more effectively. For example, automating the processing of incorporation documents, financials, merchant processing statements, and other documents is becoming a much more common way to uncover documentation fraud.

Additionally, because AI enables insight extraction from multiple data sources, teams can now pattern match across various data points and associate them with known fraud vectors. At Coris, we’ve implemented this strategy in our adverse media screening tool, which allows risk teams to automatically screen businesses and their owners for negative information indicative of key risks.

‍

Ongoing Monitoring

Risk teams often overlook opportunities to streamline SMB monitoring, but there are plenty of automation options available. 

One common priority for risk teams is understanding when a business might be in trouble or has filed for bankruptcy. Traditional bankruptcy notifications are delivered on a lag: by the time a risk analyst receives an alert, it’s too late to act. AI can provide much faster contextual information on businesses based on their activity across all monitored channels, and proactively flag when a business might be in trouble.

Beyond loss prevention, AI can also be used to monitor for revenue generating opportunities. For example, it can help business teams identify important cross-sell opportunities, and better understand which customers would benefit from additional products.

‍

When does gen AI have an edge over traditional ML models?

Generative AI is most useful for decisions and processes that are probabilistic. For example, classifying a merchant’s industry involves a high degree of uncertainty and can have multiple acceptable answers. It requires analysis of multiple data points to get to a potential answer - Gen AI like LLMs are a great fit for this decision-making process.

According to David, other SMB risk decisions require deterministic outputs, especially when it comes to credit decisions. While generative models are not a good fit for deterministic decisions, they can be used throughout the decision-making process.

Take data structuring, for instance. Traditionally, financial institutions have employed large engineering teams to build integrations with multiple data sets, clean incoming SMB data and reconcile different schemas, and merge everything together into a clean record. 

A generative model can turn this data engineering problem into a semantic search problem. It can structure data into useful inputs for deterministic decision-making, and then traditional ML models can be used for the decision output itself.

‍

Building custom models in-house vs. fine tuning an existing model

As EVP of Bankwell, Ryan has been laser focused on understanding how their team can leverage AI for their SBA product. Current SBA processes are long and cumbersome: it can take 90-100 days to close a loan and this involves lots of manual processes. 

Bankwell’s team wanted to prioritize speed over perfection, and decided to leverage an off-the-shelf solution to prove out their AI use case before building bespoke. We often see companies start out with fine tuning an existing model for their proof-of-concept before deciding whether to go down the custom route. 

The drawback of this approach is model risk management: Traditional model risk management at financial institutions, and as specified by regulators such as the OCC, focuses on having very strong lineage and controls over company data. Foundation models trained by 3rd parties that are fine tuned is a very different concept, and it might take some time for the industry to keep up with this development.

Regardless of whether a company goes down the fine tuning or custom model route, David stresses that every company should first evaluate their data quality: does the company have a strong data architecture and data pipeline?

‍

Watch the webinar to learn more

Check out the full webinar below, and reach out to our team if you’d like to learn more about the future of AI in risk management.