The bet.
AI-native financial infrastructure is the companies rebuilding back-office and middle-office financial workflows with LLMs and agents at the core — not as a feature on top of a traditional SaaS product, but as the load-bearing logic of the system. Underwriting, treasury, AP/AR, compliance, credit memos: the work that used to require a person now done by an agent that ships with examiner- ready audit trails. The buyer is the CFO, the chief credit officer, the BSA/AML officer; the budget being displaced is real headcount and real legacy software.
Why now is not “LLMs got better.” It is that, between 2023 and 2025, multimodal models crossed three specific thresholds at once: messy financial-document understanding became reliable enough for human- in-the-loop production use, conversational interfaces stopped embarrassing the buyer’s brand, and structured extraction from government forms hit examiner-acceptable accuracy. The category went from demo-able to deployable inside the same eighteen months that incumbent vendors were still figuring out their AI roadmaps. That window is the bet.
Five subsegments.
AI underwriting & credit decisioning
Companies rebuilding origination, credit analysis and decisioning workflows around LLMs — for consumer, SMB, embedded and SBA lending.
Will banks and lenders trust AI-rendered credit decisions in front of regulators, or stay constrained to AI-as-assistant?
Agentic FinOps
Treasury, spend management and procurement automation where AI agents own approval, fraud, and policy workflows end-to-end.
Does the agent layer become the primary surface area for the CFO stack, or is it a UX skin on a card-and-software platform?
AI-native back-office
AP/AR, reconciliation, close, audit and the broader controller stack — rebuilt around language-model document understanding, not OCR.
Can an AI-native ERP or close stack actually displace NetSuite at venture-backed companies before the incumbents bolt on credible AI?
Embedded finance
Embedded finance infrastructure with AI orchestration on top of BaaS rails — credit, payments, treasury inside a vertical product.
Where does AI orchestration justify a dedicated infrastructure layer above the existing BaaS plumbing?
Compliance & KYC/AML
AML, KYC, KYB, sanctions and regulatory-change monitoring — agents that clear alerts and produce examiner-ready audit trails.
Will regulated FIs trust an agent to clear an alert, or only to triage one — and is the difference a moat or a ceiling?
The rubric.
AI is a feature label on a traditional SaaS product
AI handles a meaningful but optional workflow
AI is the product; remove it and nothing is left
We are looking for products that disappear without their LLMs. Auto-coding, conversational intake, multi-source evidence synthesis, narrative generation. Cosmetic AI is the largest category in the market — a 1 or 2 score is the default, not the exception.
Surface-level tool (notifications, summaries)
Owns one full workflow end-to-end
Owns multiple connected workflows; replaces a role
Depth is the difference between a copilot and a system of record. The companies we want own the credit narrative, not the dashboard above it. Multiple connected workflows are how a product crosses from automation to displacement.
No proprietary data accumulating
Some proprietary data, weak loop
Strong loop: more usage → better model → more usage
The most over-claimed moat in fintech. We probe whether data is actually entering model updates or just sitting in a warehouse. A real loop changes the unit economics of every additional customer; a fake one is just storage.
Founders have no domain background
Founders have adjacent domain experience
Founders have done the exact job being automated
An ex-credit-officer building underwriting tools is a different bet than an ex-Google PM building underwriting tools. We score adjacency vs. exact-job experience explicitly. Engineers building for engineers usually outperforms — but not in regulated finance.
Pre-revenue or pilots only
Real ARR, mixed retention
Real ARR with strong NDR and logo quality
Real ARR with named logos beats a TAM model with conviction every time. We anchor on disclosed numbers only — no inference from round size, no extrapolation from headcount, no benchmarking to the median Series A.
Tangential to the thesis
Clearly within the thesis
Central to the thesis; SFV should know this company
We score how well a company matches the AI-native financial infrastructure brief — not how interesting the company is in general. Adjacent breakthroughs in horizontal AI tooling don’t earn a 5 here.
What kills the thesis.
Regulatory backlash on AI in lending and compliance
An OCC, CFPB or state regulator examines an AI-rendered adverse action and finds the explanation chain inadequate, or a fair-lending audit shows disparate impact in an AI-assisted approval flow. The category response — model risk management documentation, bias testing, examiner-ready audit trails — becomes a hard floor that compresses the gap between AI-native entrants and incumbents who can throw lawyers and consultants at the problem.
Incumbent counter-attack via M&A
nCino, Moody’s, FIS, Fiserv, SAP, Coupa or BILL acquire two or three of the strongest AI-native entrants in their adjacency, fold them into existing distribution, and the standalone-platform thesis collapses into an OEM thesis. Brex → Capital One in January 2026 already showed this is the live exit path. Good for early holders, harder for companies betting on independent scale.
The “AI-native” label getting diluted as everyone bolts on AI
Within 12–18 months every legacy vendor will claim AI-native status, the buyer signal collapses, and procurement teams stop treating AI-nativity as a tiebreaker. The thesis still works — but the marketing wedge disappears. Companies whose moat is real workflow ownership keep compounding; companies whose moat was being first to say the words don’t.
How a memo gets made.
Each memo is one agent run. The agent picks its own tool calls; the prompts enforce citation and numbers discipline.