From raw to riches: A lender’s path to data readiness in AI era

Trying to operate an AI model with bad data is like asking a Michelin-star chef to cook with rotting table scraps. In other words, many lenders need to clean their kitchens. Roughly 65% of lenders cite lacking data readiness as a top barrier to AI adoption, according to a January report by Experian , which surveyed more than 200 lenders. Less than 10% of financial institutions said their data infrastructure was “very prepared” to support their AI initiatives going into 2026, according to a February report by professional services firm Wolters Kluwer , which surveyed 148 FIs. Any AI model, no matter how advanced, is only as valuable as the data feeding it, Dan Snyder , executive vice president of product at open banking platform Kinective , told FinAi News . “You can't get even close to the real value of AI unless you have a structured data set,” he said. To overcome data hurdles, experts say lenders must prioritize: Data points that are most important for an AI solution; The mechanisms that extract and transmit the most important data; A repository that stores relevant data; An orchestration layer that transmits data from a repository to an AI model; and Data governance to maintain model optimization. [caption id="attachment_57390284" align="aligncenter" width="800"] (ChatGPT/AI-generated)[/caption] Unstructured data Despite AI’s ability to process immense data sets, raw and unstructured data can lead to inconsistent and ineffective model performance because of information overload, Andy Wing , director of indirect lending and dealer management at $29 billion credit union BECU , said during an Oct. 8 panel discussion at the FinAi Lending Summit in Las Vegas. “I would love to take all of our auto loan data and connect it to our AI so I can just extract the data I need and do the analysis,” he said. “But let’s slow down here.” Unstructured data can introduce patterns that were not used to train an AI model, Rohit Arora , chief executive of SMB lending marketplace Biz2Credit , told FinAi News . This is especially challenging for lenders serving multiple segments because “it's very asynchronous” and “very different industry by industry, business by business,” resulting in a high degree of variance in how borrowers submit information, Arora said. “The propensity and the percentage of unstructured data is still very high,” he said. Not all data created equal AI has become a force multiplier in financial services in part because it can extract and analyze thousands of data points quicker and more efficiently than humans. However, “not all of [the data points] matter and not all of them are created equal,” Nadia Aziz , general manager of mortgage at lending fintech Ocrolus , told FinAi News . “We want to make sure that the ones that actually matter are the ones that we are being super diligent about ensuring the accuracy,” she said. For example, a borrower’s name that appears as Robert Smith in some documents and Rob Smith in others might waste time and degrade model performance, Aziz said. “But if you don't build it the right way, it'll highlight all of those things, which just becomes a lot of noise … and I think that's where structured data becomes super important,” she said. The most valuable metrics vary depending on the use case. Cash-flow data, for example, is important for AI-powered underwriting because it provides “lots of useful information that is very hard [for humans] to actually process,” Brett Manning , chief product manager at fintech Plaid , said at the FinAi Lending Summit. AI can effectively parse nuanced cash-flow metrics such as discretional versus essential spending, he said. “If you see someone making large purchases before a windfall or after a windfall, that can tell you something about how they manage their finances,” Manning said. The abstraction layer Once lenders identify the data points they need for a workflow, the next step is build mechanisms that contextualize and transmit information to a centralized repository. That’s why U.S. Bank is prioritizing “intelligent document processing” to get the most out of its AI lending tools, Suzanne Rathbun , senior vice president and lending operations group manager at the $702 billion bank, said during the FinAi Lending Summit. “How do you extract the data and start to think about where that data needs to go?” she said. Building a knowledge graph is one way to effectively transmit data, Joshua Summers , CEO of EnFi , an agentic AI provider for commercial lending, told FinAi News . A knowledge graph describes people, places and things through an ontology, or a structured framework that defines certain data sets and how they relate to each other, Summers said. [caption id="attachment_57390285" align="aligncenter" width="800"] An example of a knowledge graph representing mock customer data. (Courtesy/EnFi)[/caption] Knowledge graphs are a type of semantic or abstraction layer, which refers to data infrastructure components that translate complex information into shared meaning, according to cloud-based AI company Databricks . For example, a knowledge graph can help an AI agent or system understand the link between a borrower, lender, loan terms and loan covenants — even when each of those represent a different building block within a data set, Summers said. When these graphs are established, AI agents can “take the raw data and figure out how it maps into the graph,” he said. Building data lakes, warehouses Another step to optimizing AI involves data storage techniques One such repository is called a data lake, Piermont Bank Chief Executive Wendy Cai-Lee said during the summit. Data lakes comprise unstructured data with toggles that enable structure depending on a user’s query, according to IBM . Data warehouses are similar but have frameworks already in place. Piermont is a digital bank that was founded in 2019 and built on modern technology. When building the bank’s data lake, “the first thing we did was do a full assessment of the source of the data,” Cai-Lee said. “‘Where do we get all our data from? What condition are they in right now and how are we even storing them? Where does this data go both internally and externally?’” she said. For the $505 million bank, the answers to these questions vary based on the lending segment, Cai-Lee said. For example, Piermont’s commercial real estate documents contain “very different formatting information compared to my fund finance” documents, she said. After gaining an understanding of these nuances, the bank assessed how it wanted to use certain data sets for AI and how that might evolve over time, Cai-Lee said. “It's not even about the construction of the building; it's about the ability to maintain it,” she said. The orchestration layer Data lakes and other repositories are often enhanced by an orchestration layer or "multi-agent system” that sits between the final AI output. [caption id="attachment_57390286" align="aligncenter" width="800"] (ChatGPT/AI-generated)[/caption] This layer constantly reads the knowledge graphs and repositories to maximize the ontology feeding a model, EnFi’s Summers said. “Another set of agents on the other side … they have tools that allow them to understand how to query the graph, how to use the data,” he said. “Some of those tools may be analytics capabilities and how to build dashboards and graphs and charts off of that. Others may be how to do a cash-flow analysis. Others may be how to write an annual review report.” The orchestration layer helps deliver the exact information needed for an AI system to execute a task by working "upstream" instead of reading raw documents and unstructured data, Summers said. Actionable insights Ultimately, FIs have many tools to help them structure and analyze data for optimal AI deployment. Once a strong data foundation is in place, ongoing governance is crucial to ensure that “the data used in any AI instance is valid and clean and trustworthy,” Marty Miracle , chief digital and risk officer at Thread Bank , told FinAi News . “In the end, I think AI will play a large role in us governing data,” he said. “It's really good at seeing a disparate data source or seeing an anomaly out there and finding it, alerting you and helping you drive user decisions on that. We just need the right infrastructure.” At the end of the day, the governance layer sits atop of all the foundation work that has been set within a bank's data investment, like the finishing garnish a chef places on a meal before sending it out. Register here for early-bird pricing for the FinAi Banking Summit, taking place March 1-2 in Denver. View the full event agenda here .

From raw to riches: A lender’s path to data readiness in AI era

65% of lenders cite lacking data readiness as top barrier to AI deployment

Why Nvidia, Paypal are betting on Meta while major FIs block Muse

Related Posts

Truist explores sale of specialty lender Afco credit

TD plans to automate nearly 70% of underwriting process

DLL offering ‘holistic’ financing for cybersecurity, AI investments

EMERGING FINTECH DIRECTORY

FinAi Podcast

SPONSORED

The Next Generation of Cyberthreats: Powered by AI

Build an Antifragile Strategy to Outperform the Market

How AI and Product Experts Turn Fuzzy Requirements Into Focused Dev-ready Roadmaps

Connect

Welcome Back!

Retrieve your password

Continue Reading with FinAi News Premium - Less than $2/Day

From raw to riches: A lender’s path to data readiness in AI era

65% of lenders cite lacking data readiness as top barrier to AI deployment

Why Nvidia, Paypal are betting on Meta while major FIs block Muse

Related Posts

FinAi Podcast

Connect

Welcome Back!

Retrieve your password

Unlock This Article

Continue Reading with FinAi News Premium - Less than $2/Day