FinovateFall 2026: AI Underwriting Vendors Had the Same Answer for Explainability
By Jill Robb
Finopotamus was on the ground at FinovateFall 2026, held September 9-11 in New York City. Across a show floor packed with generative AI demos, the most interesting for lenders were the ones that could show and make its work auditable.
Explainability has been a buzzword in fintech for years, usually deployed by vendors as a checkbox feature bolted onto a black-box model. This year, a number of companies, all working on three different parts of the lending and compliance pipeline built it into the actual architecture of what they were selling, and made a point of saying so, unprompted, before anyone in the audience asked.
Titan AI builds a knowledge graph behind every loan decision. Titan's demo opened with a mock home equity refinance application. The company's small language model flagged a mismatch between stated income, assets and verified financial behavior, and recommended an escalation for manual review.
"There is a limit amongst large language models to how explainable they can be," the presenter told the room. "Some are mathematical, some are actually geopolitical. Our small language model does not have those limits for explainability. We are not after artificial general intelligence, we are after endgame intelligence, and therefore explainability is a first-class citizen in the model that we built."
The demo then opened what the company called the audit log: a step-by-step trace of the twelve reasoning steps the model took before rendering its recommendation, along with a confidence indicator and a description of exactly what context it drew on. Titan's pitch is that this context comes from what it calls an institutional knowledge graph: roughly two million nodes encoding a bank's own lending policy, built from what the presenter described as "the largest institutional knowledge graph of banking." The model, the demo highlighted, reasons from the specific ability-to-repay rule that exists at that particular institution. "Without that knowledge being invoked," the presenter said, "the model — large language model or otherwise — would have resulted in a different outcome."
Kita Technologies keeps the underwriter in the loop.
"Our AI agents are able to reason through complex files with human-like intuition, with full explainability and source-level citations, so you can be fully audited by any regulator you might have," the presenter said. Kita was explicit that final decisioning authority stays with the institution's own underwriting team. The company positions its product as a recommendation engine, rather than an autopilot, with a portfolio-analysis layer intended to surface what's actually driving approval, denial, and default rates so lenders can refine credit policy accordingly. The closing pitch: "With Kita, lenders we work with underwrite anyone in minutes."
FSAi by McCarthy Hatch turns customer complaints into an early-warning radar.
The most personal framing of the explainability theme came from McCarthy Hatch, founded by a former Consumer Financial Protection Bureau official who helped stand up the agency in 2011. "The institutions weren't listening to their customers," he told the room, describing his time at the CFPB. "They thought they were. There were systems in place, there were mechanisms, but collectively they weren't listening... $22 billion was returned to 250 million Americans."
The solution, FSAi, is built to bring that "consumer voice as radar" capability inside the institution, reading complaint narratives, call center transcripts, and surveys in real time. In the live demo, a market-condition screen showed 277,000 complaints in the system, with mortgage applications and refinancing flagged as an accelerating category.
Then came the test: restricting the dataset to only pre-2026 records and asking FSAi whether it could have predicted a real 2026 enforcement action against lender GreenSky, which regulators found had issued loans to consumers who were unaware they'd taken one out. The system surfaced 76 relevant complaints from the restricted window, with 47% involving payday, title, personal, or advance loans, and specific patterns around inaccurate credit reporting and blocked dispute resolution. "Your customers know more about your organization than the leadership knows today," he said. "Your job is to make sure that none of that becomes a surprise."
Explainability Was the Common Thread
None of these three companies led with raw model accuracy or processing speed, one of the metrics that dominated AI vendor pitches just a year or two ago. All three demos led with the same underlying claim: that their system's reasoning could be traced, audited, and defended after the fact. For credit unions weighing AI tools in lending and compliance (the NCUA and state examiners are likely to pay even more attention to how a decision was reached and not just what the decision was), that convergence is worth paying attention to on its own.
