The distinction between AI applied to financial services and AI-native financial infrastructure is not a semantic one. Most of the AI deployment in financial services over the past decade has taken the form of the former: machine learning models grafted onto existing processes — fraud detection rules supplemented by ML classifiers, credit underwriting scorecards supplemented by open banking data features, customer service chatbots built on top of existing CRM systems. These are genuine improvements over the processes they augment, but they are not restructuring the architecture of financial services. The infrastructure layer — the data pipelines, the model serving infrastructure, the decision audit frameworks, the regulatory reporting systems — was designed before ML models were in the loop and largely remains so.
AI-native financial infrastructure means something more specific: systems where the machine learning model is not an addition to an existing decision process but is the decision process, and where the surrounding infrastructure — data ingestion, feature engineering, model governance, decision logging, regulatory audit trail — is designed from the ground up to serve that model rather than retro-fitted around it. A credit underwriting system designed this way looks different from one designed around a scorecard with ML added: the data model tracks feature importance over time to detect model drift, the decision logging captures the full feature vector at inference time for audit purposes, and the retraining pipeline is integrated with the production serving infrastructure rather than being a separate offline process.
The regulatory dimension of AI-native financial infrastructure is where the real architectural constraint lies. The FCA's emerging approach to algorithmic accountability in financial services — drawing on the Consumer Duty requirements and the FCA's 2022 discussion paper on artificial intelligence — signals that regulators will expect financial institutions using AI for consequential decisions to be able to explain those decisions, to demonstrate that they do not produce discriminatory outcomes, and to maintain an audit trail that supports supervisory review. Building this into an AI financial product after the fact is extremely expensive; designing it in from the start is a matter of architectural choices at the data and model governance layers.
The infrastructure opportunity we are most focused on is in the model governance and regulatory audit layer. As more financial institutions move to AI-native decisioning — for credit, for fraud, for customer onboarding — the infrastructure to manage the model lifecycle, maintain the decision audit trail, and produce the regulatory documentation that supervisors will require becomes a standard component of the financial services technology stack. This is not a glamorous product category: it exists entirely in the compliance and risk function of its customers, not in the consumer-facing layer where fintech products typically generate their brand identity. But infrastructure that sits in the compliance stack of regulated financial institutions exhibits the same characteristics as the best fintech infrastructure generally — high switching costs, excellent net revenue retention, and a sales process driven by risk avoidance rather than optional feature preference.
We are selective about what we call AI in this context. The models used in financial services inference must meet a different standard than the models used in, say, content recommendation: the consequences of an error — a misclassified fraud alert, a wrongly declined credit application, an incorrect AML flag — have regulatory, financial, and sometimes legal implications. Infrastructure built for AI-native financial services must therefore treat model quality and interpretability as first-class engineering requirements, not as features to be added when customers ask for them. The companies in our portfolio that work in this space have all arrived at a similar conclusion: the technical barrier to building AI financial infrastructure is not the ML model itself — it is the surrounding architecture that makes the model deployable, auditable, and defensible in a regulated environment.