Welcome to AI for FIs, from Dixon Strategic Labs. Each week, this newsletter tracks agentic AI and explains what it means for community banks and credit unions.
Who chooses which AI models reach a credit union?
America's Credit Unions (ACU) told Congress this month that vendors usually make that choice. Three core providers serve more than 70% of banks too, so community banks sit in a similar position. The AI arrives embedded in vendor products, shaped by partnerships the institution never negotiated.
ACU proposes a shared option for security. Smaller institutions would get access to advanced defensive models that can triage older software for vulnerabilities.
Stripe is acquiring a different kind of access point: a gateway that routes requests across more than 400 models. The deal could give Stripe a role in how developers evaluate, select, and switch models.
Meanwhile, payments companies are building their own foundation models. Razorpay, Stripe, Mastercard, and Revolut have each trained models on billions of transactions for uses including payment routing, fraud detection, and credit scoring.
Model choice shapes data access and switching costs. Vendor contracts and gateways both influence which models are available and how easily an institution can switch.
Credit unions inherit AI decisions from their technology providers
Sources: America's Credit Unions letter, Aug. 14
ACU calls credit unions "deployers and, overwhelmingly, indirect consumers" of AI embedded in vendor products, with "marginal influence" over the software supply chains behind them. A credit union generally has "no ability to access, let alone modify," the code in its core, the association writes. The financial institution still answers for the result, though. NCUA expects a credit union using third-party AI to understand how the product works, the risks it introduces, and the vendor's safeguards.
Three core providers are building with three model companies
Fiserv, FIS, and Jack Henry supply core systems to more than 70% of banks (most recent number I could find is from 2022), and Fiserv and Jack Henry alone serve about 38% of credit unions, (Callahan & Associates, 2025). Each is building AI products with a different model company:
Jack Henry is building an AI security platform with Google Cloud.
These partnerships shape which models reach an FI through its core.
South Africa's Standard Bank has set up model flexibility through Amazon Bedrock. The bank evaluates models from several providers under one security framework rather than committing to one.
ACU offers one arrangement: regulators such as NCUA, working with Treasury and the FFIEC, could license advanced defensive models through a controlled-access portal. Access would go first to participants who can raise security for more institutions, measured by institution count rather than asset size. Regulators could subsidize smaller institutions' access and distribute the findings.
ACU says many of the smallest credit unions, including those below $50 million in assets, may lack the staff, data, or scale to fully test an advanced cybersecurity model. Early access could improve security, insurance pricing, vendor relationships, and member trust.
Stripe is buying an AI model gateway
Sources: OpenRouter, Aug. 19 · Axios, Aug. 19
Source: OpenRouter
OpenRouter is a general-purpose gateway: developers building AI products can use it to access more than 400 models through a single integration. They set rules for cost, availability, and data handling. The connection stays the same when they switch models.
AT&T says its gateway weighs cost, speed, and expected quality for every request and has cut AI costs by as much as 90%. Open-source models now handle 40% of employees’ AI queries, The Information reports.
Open-source models now handle 40% of AT&T employees’ AI queries. @amir on X.
OpenRouter announced an acquisition agreement on Aug. 19, subject to closing conditions. Neither company disclosed terms. The deal would give Stripe a role in how developers evaluate, select, and switch models.
Stripe charges fintechs per payment transaction. If this deal closes, it could also potentially charge them for AI model usage.
Explainer: What is a foundation model?
Razorpay’s Vulcan, and how a language foundation model differs from a payments foundation model.
Sources: Razorpay's Vulcan announcement, Aug. 18 · Mastercard model explainer, Mar. 17 · Stripe technical post, May 15, 2025 · PRAGMA paper, Apr. 9 · Stanford foundation-model paper
Source: Stanford CRFM.
Razorpay launched Vulcan on Aug. 18, a payments foundation model trained on four billion payments. Razorpay says it can route transactions, detect fraud, flag risky cash-on-delivery orders before checkout, and personalize checkout.
A foundation model is a reusable base trained on a large, varied body of data and adapted for different jobs. Large language models are one kind. Razorpay's launch of Vulcan shows the same design applied to payments.
Training data shapes what patterns the model learns. The product built around it determines how that output reaches a person or system.
Financial institutions can compare models by the job each performs, how accurate and fast it is, what it costs, how it handles data, and where it can run.
A model inside a product
Foundations models are the engines that run AI products.
GPT is a family of language models. ChatGPT is a product built around them, with an interface, instructions, tools, and other systems. The model learns patterns across text and code. The product uses those patterns for tasks such as drafting, search, and tool use.
Banking data arrives as events
Payments and banking foundation models learn from event histories: sequences of transactions where each record can include amount, timestamp, merchant category, payment method, and result.
Revolut's PRAGMA paper describes a banking history as transactions, app activity, trading, and customer communications. PRAGMA encodes each field's meaning, value, and time. Mastercard describes the connection in prediction terms: a language model predicts the next word, while its payment model predicts future transactions.
A shared base supports different jobs
Stripe says its model turns each payment into a numerical representation that supports different predictions. Revolut's researchers evaluated PRAGMA across credit scoring, fraud detection, recurring-payment detection, and product recommendations.
Foundation describes the reusable base. The product around it turns output into a recommendation, a score, a route, or an action.
📡 On the Radar
State Bank of India is piloting an agent that prepares corporate-loan files. The agent pulls data from bank systems, collects and reads loan documents, and writes risk analysis into the loan workflow. Relationship managers validate the analysis before a credit decision.
Oops. In a synthetic test, AI analytics agents produced collections lists that included customers who had paid. Hex's DataBench gave models stale delinquency flags and a separate cash ledger. The agents had to reconcile the conflict before selecting accounts. Only one tested model passed.
Mercury launched corporate cards for AI agents. A person creates each Agent Card and sets its budget, eligible merchants, and spending categories. The card enforces those limits automatically, and the agent cannot raise them/
Source: Mercury.
AI agents set prices together in Anthropic's laboratory experiment. Each agent was told to maximize its profit. By round three, they had settled on price floors and kept matching prices even after researchers cut off direct communication. Anthropic found that agents using the same model and similar contexts tend to make similar choices, which can weaken competition.
Agents struggled to follow tool-use policies while working across APIs and documents. IBM Research's VAKRA paper tested models across more than 8,000 APIs in 62 domains. The agents chose tools and combined records with information from documents. On questions that policy made unanswerable, accuracy fell as low as 2.4%.
via reddit
I help leaders at credit unions and CDFIs figure out what agentic AI means for their strategy and learn by doing. Drop me a note at [email protected].
How this newsletter is made: Brent curates the research and writes the analysis, with AI tools helping with research, drafting, and editing. ⚡ Alakazam ⚡.
Show someone you really love them by having them subscribe to AI for FIs here.
