Since the public availability of large language models from late 2022 onward, the category of "AI fintech" has expanded dramatically in both substance and noise. The substance includes genuine advances in AI-native underwriting, in real-time fraud detection, and in the automation of financial operations that were previously manual. The noise includes a considerable number of companies that have placed a language model interface in front of an unchanged legacy system and described this as an AI product. We have seen this pattern in every category we invest in: a conversational interface for a loan application, a natural language query layer for financial data, a chatbot front-end for customer service at a financial institution. These are improvements to existing workflows, but they are not the AI-native finance thesis that we are investing behind.
The architectural distinction matters because of where the competitive moat lies. A language model placed in front of a legacy loan origination system adds user experience value — customers can describe their needs in natural language, the interface is more forgiving of unusual inputs — but the underwriting decision is still made by the legacy system’s rules engine. The AI component can be swapped out or replicated by a competitor relatively easily. The competitive value of the legacy system — its credit performance data, its compliance framework, its scheme membership — is unchanged. Conversely, a credit decisioning system where the core underwriting model is a machine learning model trained on consented Open Banking data, where the feature engineering is designed for ML inference rather than human-readable rules, and where the model retraining pipeline is integrated with production serving, cannot be replicated by adding an LLM interface to a rules engine. The competitive moat is at the architecture level, not at the interface level.
The companies we find most interesting in AI-native finance are those where the product would not exist without the AI component — where the underlying service is only feasible because machine learning makes it tractable. Flowcast, which we backed in 2024, is building AI credit decisioning infrastructure for lenders who need to underwrite borrowers without adequate credit bureau coverage. The decisioning model they provide could not be built as a rules engine — the number of features required, the non-linear interactions between transaction behaviour signals, and the need to retrain against new performance data continuously make a rules-based approximation functionally unusable. The product is only possible as a machine learning system, which means there is no legacy system to wrap an interface around.
The model governance question is where we most frequently probe. An AI-native financial product must answer three regulatory questions clearly: how does the model make a decision on any given input (explainability), how does the firm monitor for model drift and bias (ongoing validation), and what happens when the model is wrong and a customer is harmed (recourse and audit trail). The FCA’s increasing scrutiny of algorithmic decision-making in financial services — informed by the Consumer Duty obligations and the FCA’s work on algorithmic bias in credit — means that these are not theoretical questions for future compliance teams to answer. They are live requirements that affect product architecture today. The companies that have designed their AI systems with explainability and governance as first-class requirements are better placed not only for regulatory examination but for the enterprise sales process, where financial institution buyers increasingly include model governance questions in their vendor due diligence.
We will pass on AI fintech companies where the AI is a feature claim rather than a structural capability. The pattern we have seen consistently: the pitch describes an AI-powered product, the technical diligence reveals a rules engine with one ML feature for scoring, and the data moat story dissolves when you ask how the model is retrained and what the firm will do when model performance degrades. This is not a criticism of any specific company — building genuine AI-native financial infrastructure is hard, and many early-stage teams are doing the difficult work of getting there. But the investment thesis is specific: we are backing AI-native architecture, not AI-adjacent user experience. The distinction is load-bearing for the competitive moat, and we hold to it.