AI is no longer a side project in asset-based finance. It is reshaping how lenders collect data, assess risk, monitor portfolios and run operations. But as the industry races ahead, not every use case is ready for prime time, and not every institution is equipped to manage the security, compliance and data challenges that come with it. In this roundtable, four Power Players and industry technologists offer a candid look at what’s working, what’s overhyped and what the path forward looks like for lenders and investors navigating the next wave of digital transformation.
How are AI and emerging technologies transforming credit underwriting, risk analysis, and portfolio monitoring in asset-based finance?
Joseph Caplan: For the companies adopting or are interested in technological changes, we see some automation in financial analysis, alerts, automation of paperwork and workflows. The rush into AI has pluses and minuses because AI does not seem to know much about commercial lending / ABL details. While it seems to know some basic ABL rules, it does not know about the hundreds of drivers. We’re also seeing cases where people are making incorrect decisions from it. We’re focused on time-saving areas within the realm of our expertise in underwriting, audit and data analytics. Our newest ABL Field Exam technology is a game changer for insight and profitability.
Nelson Chu: We’re finally seeing AI bring real-time intelligence to a part of credit underwriting that’s been largely manual and backward-looking. Instead of teams digging through PDFs and spreadsheets every quarter or with every deal launch, models can now pull structured data from financial statements, loan tapes and collateral documents automatically. That frees underwriters to focus on judgment instead of data wrangling. On the risk side, predictive tools are spotting early warning signs in borrower performance before traditional reviews catch them. And for portfolio monitoring, we’re moving toward live dashboards that flag issues the moment they arise. It’s not about replacing humans — it’s about giving them better visibility, faster insights and time to work through problem solving, or bringing more innovative thinking and creativity into understanding and working through credit risk.
Matt Katz: There’s a rush to explore the benefits of all varieties of artificial intelligence and a lot of excitement about it. The patterns emerging are that AI is very useful for specific things and very weak at others. Analyzing large amounts of data for trends, unusual differences, summarizing text — these are great. I’ve yet to see anyone confident to use generative AI for serious work involving numbers — certainly nothing you’d send to a regulator. But to take information like financial statements or unstructured data and make it easier to get structured, useful information out of them, these ¬are valuable uses. Extracting sentiment and key factors from a large corpus of data is a path that is looking useful.
Some are looking in less proven areas, as well. Analyzing social media activity for creditworthiness, for example, involves being exposed to bias in training data, inaccurate profile linking, and creating self-fulfilling prophecies in risk assessment. Large language models include all the bias from their training data, so firms that use an LLM to make value judgments may be exposing themselves to risk of structural bias. Washing it through an LLM may not be a shield to liability.
Karan Oberoi: AI is helping lenders see the full picture faster and with greater precision. In credit underwriting, integrated platforms that bring together historically siloed lines of business enable new AI agents and models that consider far more data to create brand new insights for the business. This approach creates better risk calibration, faster decisions, and reduced bias through transparent and explainable models.
For asset-based lenders, this means moving beyond static models toward dynamic, data-driven assessments that evolve with each borrower’s performance. For portfolio monitoring, AI agents now continuously analyze exposures, collateral values, and regulatory impacts. They alert teams before risks become losses. When combined with cloud-native data architectures, this creates an intelligent ecosystem where credit quality and portfolio health are visible in real time rather than in retrospect.
What pain points are financial institutions still grappling with when it comes to digitization, automation, or data integration?
Caplan: The willingness of lenders to change is an ongoing battle for software companies, but labor scarcity, rising costs and potential AI time savings are motivating lenders and accountants to automate things.
A significant pain area continues to be writing good prompts in AI and even with training and reading hundreds of pages about it. Having tons of data would help, but that is not necessarily part of the reporting that ABLs get. At this point, we’ve invested thousands of hours in AI prompt development, and we’re still discovering more ways to make it work better. The average well-educated employee is going to be missing the skills and experience it takes to get it right. We’ve seen skilled software engineers correct things in minutes from their experiences, while other programmers worked for days to still fail. AI prompts are a bit like that, and the correct answers take more than an AI engine to get it right.
Chu: Legacy systems are still the biggest drag. A lot of lenders have great tools, but they’re bolted onto decades-old infrastructure that doesn’t play well together. Data is also scattered — operations, risk, and compliance teams often each have their own version of the truth. Even when new analytics are in place, getting everyone to trust and use them consistently is tough. And there’s still a real lack of standardized data formats in private credit, which makes automation harder than it should be. With AI, we have been able to develop technology that allows us to structure standardized transactions at a fraction of the cost and speed.
Katz: Ultimately, all the AI stories become data stories, and all the data stories become, at some point, data governance stories. Most institutions are, at one level or another, confronting the challenge of multiple siloed data warehouses, of data at different grains and with different meanings. Data stores that are fit for the purpose of the data owner may not be ready for other groups in the organization or are hard to discover.
We also see confusion about, “Who’s responsible for making sure this data is accurate?” “A number in a datastore I’m using isn’t correct — how do we find out where the number came from or what the lineage is to get it corrected in the source?”
This is driving the move to more modern data platforms so that big new data initiatives can give answers at higher confidence levels.
Oberoi: Most institutions continue to struggle with the weight of legacy systems. Data is trapped in silos, processes remain manual, insights lack full visibility across the business, and core systems struggle to communicate with modern digital tools. This creates friction for teams, delays in capital, and inconsistency for customers.
The challenge is not the lack of technology, but rather the orchestration of it. Integrating data across origination, servicing, and risk requires open platforms, strong APIs, and a willingness to modernize operations from the inside out. The institutions that succeed are those that treat digitization as a business model transformation rather than a software upgrade.
What are the most promising use cases you’re seeing for AI or machine learning in the lending lifecycle — and what’s overhyped?
Caplan: The Use of AI for total decision-making is overhyped. While AI can assist in identifying potential risks and irregularities or hot spots in data, it offers little more than another tool for human due diligence. We get a ton of great insights using IF and CASE statements that AI never gets, but our accounting and audit background set us apart from wishful and failure ridden endeavors in AI. Having enough data (daily, weekly, monthly reporting) is a significant problem, even if borrowers provide the required reports. The art of decision making with limited data is to ask questions about what you don’t know. Critical thinking humans are needed to assess the data and AI results and to make further refinements.
We went after the “holy grail” items, like cutting field exam report writeup time by 90%. This is a massive improvement for lenders that can’t charge for write-up time, and it allows lenders or accountants to conduct more exams with the same staff. The payback starts on day one, and the profit improvement numbers are substantial from our technology. We also see other significant uses for this technology to expand into other areas.
Chu: Where AI is actually delivering value right now is in automating the grunt work — extracting data from financials, matching invoices, tracking collateral, simplifying very complex models and data sets — and then feeding that into smarter, faster underwriting decisions. Machine learning is also great at continuous monitoring; it can flag subtle changes in a borrower’s performance long before something shows up in a report. Those are tangible wins.
What’s overhyped is the idea that AI can just “do” underwriting on its own, like flushing out a ready to go memo and credit analysis on a borrower or structure. In asset-based and middle-market lending, context matters too much — you need humans to interpret nuance. For instance, companies can use different accounting standards, or can be audited or not. When comparisons are not properly matched, an unsupervised AI can derive inaccurate insights. The same goes for using general-purpose chatbots as compliance tools; they’re helpful, but you still need explainable, auditable, (truly) compliant processes behind the scenes.
Katz: Most promising use cases revolve around summarizing across large sets of data and looking for trends. Organizations are getting a lot of value out of getting a first pass look over large amounts of loan data — some of which is unstructured, intended as documentation from a loan officer for signoff — and being able to see differences. They’re also able to identify items that aren’t currently in a tabular format and get them into a state where they can be analyzed. Using machine learning to look a level deeper into original sources and check higher-level summaries to see if they are supported or if there are differences that aren’t priced in are good use cases for ML.
We are seeing lots of investment around agentic AI workflows, but it’s very early days to see true risk-adjusted returns on these investments.
Oberoi: The most immediate value is in decision support and document intelligence. AI can now read, classify, and extract insights from thousands of documents in seconds, automating tasks that once consumed entire operations teams.
There’s also meaningful progress in customer experience, where predictive servicing and adaptive workflows are enabling lenders to anticipate borrower needs and proactively manage accounts.
At Solifi, we’re seeing AI agents beginning to support multiple aspects of lending operations and deliver high value automation. The real opportunity is to augment human expertise with transparent, auditable AI that drives confidence in every decision.
What’s overhyped? The idea that credit decisioning can, or should, be fully automated. Lending is inherently a judgment-based business, and the best outcomes still come from pairing advanced technology with human insight.
How should firms be thinking about security, compliance, and ethical AI use as they scale digital tools in 2025 and beyond?
Caplan: We wrote a blog on AI Security for ABL Providers two months ago to cover that question. In summary, the secure use of AI requires that all information be encrypted in transit and at rest, but also with the AI not sharing the data with the public and the identifiable names and addresses need to be scrubbed out (obfuscated).
I’m going to beat this drum one more time because I see it as one of the biggest risks: the quality of the AI provider for integrity and trust. AI providers seeking to monetize the value of AI data and prompts, or those with prior questionable acts for user privacy, get a hard “NO” from us as AI providers. Could the AI provider potentially be a bad actor? We have concerns; choose wisely.
Not our software area, but AI fakes in documents and voices from bad actors are attempting to get funds with fake vendor payables, fake invoices, fake wire transfer authorizations and other methods. Lenders and Borrowers need to use lower thresholds for when to require double authorization amounts, secret ID words, multi-factor authentication and direct communications from lenders to Borrowers for approval.
Some are suggesting that studying philosophy, the humanities and ethics might be the foundation of corporate AI governance. But is our case that complex? We have financial data ranging from financials to ageing and inventory reports, and in some cases, collections, sales, etc. We’re not developing “deep fakes” here. The risks of employee theft and self-dealing with knowledge are the same as having paper or a spreadsheet. Corporate policies for confidentiality and ethical behavior apply to AI.
Chu: You can’t separate innovation from governance anymore — they have to grow together. Every firm using AI should have a clear playbook for how models are trained, validated, and monitored. You also need strong data controls — encryption, access management, and clear audit trails. And just as importantly, humans should either take action or manage all complex or high-stakes decisions. Regulators are watching how firms manage third-party data and model risk, so treating vendors like true partners under the same oversight is key. At the end of the day, compliance isn’t a bottleneck; it’s what lets you scale safely and sustainably.
Katz: Firms should very seriously think about security. If you are shipping data or prompts off to a third party, you should be sure that this isn’t going to be the training data for the next iteration of the model! Clever prompt engineering can often get data back out of model vector embeddings. At the very least, make sure you’ve got an understanding of the OWASP Top 10 threats. Many large language models are trained on…the whole internet. Some corners of it are really dark out there.
Ethical use is also important to consider from an employee perspective and a customer perspective. From an employee perspective, are you empowering your staff to solve the problems they are responsible for, or are you designing systems where they are simply a responsibility or blame sink for the hallucinations of an LLM?
For customers, are you sure about the underlying assumptions in the tools you’re deploying? You don’t want to be unintentionally doing the 2025 version of redlining, but with a digital assistant.
Oberoi: Trust is the currency of finance. As institutions deploy AI, they must ensure every model and decision can be traced, tested, and explained. Data governance, model validation, and audit trails must be built from the start.
Cloud and SaaS providers play a critical role in this. But compliance is only the beginning. Ethical AI requires that models are trained on fully representative data and that outcomes are consistently monitored continuously for bias, overfitting, and skew. Transparency builds confidence, and confidence is what the finance industry runs on.
What’s next on the innovation roadmap for financial software and technology providers serving middle market lenders and investors?
Caplan: Middle market is not the only market for us, and we have significant and world-leading clients as well as smaller non-bank lenders. The hot technology market changes are in energy (fusion), quantum computing and AI, the trifecta of the future. It is evolving rapidly, but the capital investment picture into AI may not be sustainable in terms of ROI because current AI servers are only Gen1 or Gen2; we’re still in the infancy era of AI. We continue to focus on software that offers significant time savings and focused insights, just as we have been doing since 1996. With AI, we’re able to add insights that would take an A-game employee to find, summarize, and convey, but with 90% time savings. Tangible, attainable, real and awesome.
Chu: We’re heading toward a much more connected ecosystem. Data standards will make it easier to compare loans and collateral across platforms. Lenders will expect modular, API-first systems so they can plug in best-of-breed tools instead of being locked into one vendor. AI copilots will become part of the day-to-day workflow, helping teams make faster, better-informed decisions without losing control. Structures will be more similar across the board and will enable a more transparent competition in the market, bringing more efficiencies and less information asymmetry across constituents. As a result, we will also see better infrastructure for liquidity, documentation, and model governance built right into these platforms. The ultimate goal — and what we’ve been building at Percent — is to make private credit as transparent and efficient as public credit, without sacrificing the diligence that makes it work.
Katz: We expect to see a recap of the Web 2.0 mashup era replayed through this sector. You’ve got a grand unifying technology like ML or AI that allows you to join disparate ideas and data sources. Right now, it’s an era of excitement where you’ll see lots of new APIs and MCP servers stood up to help with processing this data. I’d expect that, over time, we’ll see something to refine the security and access stories across these use cases, to enable you to more clearly trace the lineage of information.
People will ask the question, “Where did this information come from?” There’s going to be new technology to help us answer that and trace why the AI said a strange thing, and how we can correct it going forward.
Oberoi: The next phase is about intelligence in action, what we call agentic automation. Rather than static workflows, systems will reason and act within defined guardrails to move transactions forward.
For middle-market lenders, this means faster onboarding, adaptive pricing, and proactive risk management without increasing headcount or complexity. For large enterprise organizations, this means fast, predictable scalability and deep insights that span lines of business.
At Solifi, we see the future as connected, composable, and cloud-native. Open finance platforms will link data, AI, and automation seamlessly. The winners will be those who modernize safely, scale intelligently, and innovate responsibly.
