Welcome to AI for FIs, from Dixon Strategic Labs. Each week, this newsletter curates critical developments in agentic AI and explains why they matter for credit unions.
👋 A warm welcome to our new subscribers this week!
On April 9, Perplexity’s agent (Perplexity Computer) gained read-only access to banking accounts through Plaid. Within hours, Fintech Brainfood founder Simon Taylor flagged a user's claim that Perplexity caught $14,000 in errors on a tax return they paid a CPA $2,000 to file. Perplexity calls it a Personal CFO.
On the infrastructure side, Lloyds installed an AI agent in its boardroom, and Nymbus shipped one of the first Model Context Protocol servers purpose-built for core banking.
AI personal finance startup Hiro is joining OpenAI. The UK AI Security Institute (AISI) also published its independent evaluation of Claude Mythos (covered last week). The US, Canada, and the UK all responded the same week.
This week's deep dive argues a credit union's rate advantage stops mattering the moment members ask Perplexity for loans.
Let’s get into it 👇
Perplexity · Apr 9, 2026
Perplexity Computer added read-only Plaid integration. A person can link checking, savings, credit cards, and loans, then ask the agent to build a net worth dashboard, a category-level budget tracker, or a debt payoff plan from a single prompt.
The tools use real transaction data, update daily, and accept plain-English changes. Basic access is free in the US and Canada; Pro and Max tiers unlock deeper analysis.
Why it matters: Frontier labs are now PFM competitors. The agents build around one member's full financial picture and update with every conversation. The feature set is whatever the user asks for on a given day (and it has no ceiling). Vals AI's Finance Agent v1.1 benchmark, last updated April 10, puts Claude Sonnet 4.6 at 63% on SEC analyst tasks as the current top performer. A December 2024 research paper tested leading LLMs on US personal finance questions (mortgages, taxes, loans, investments) and averaged about 70%. No public benchmark tests Perplexity + Plaid yet.
🔔 Acqui-Hire Alert: Personal finance startup Hiro is joining OpenAI
On Monday afternoon, one-year-old AI personal finance startup Hiro announced it is joining OpenAI. The Hiro product shuts down April 20. Hiro will delete user data by May 13, and OpenAI is not receiving it.
Hiro built an MCP server for personal financial data, the consumer-side mirror of what Nymbus just shipped for core banking.
Hiro's founders Bloch and Doshi also built Digit (Bloch founded it in 2013, Doshi later ran product). Oportun bought it for $211 million in 2021.
Why it matters: This is OpenAI's second personal finance acqui-hire in six months, after Roi in October 2025. Frontier labs are setting members' expectations for AI-driven financial advice.
IBS Intelligence · Apr 13, 2026
Nymbus published a Model Context Protocol (MCP) server supporting 19 core banking tools across customer lookup, account management, money movement, and debit card controls. The server enforces role-based access, logs every call, and masks personal data in the logs. Credit unions pick which tools are live and which staff roles can trigger them.
Diagram illustrating integration complexity before and after MCP. Source: Descope
What MCP is: Think of MCP as a shipping container for AI. In 1956, Malcom McLean loaded 58 standardized containers onto a ship called Ideal X, sailing from Port Newark to Port Houston. Before that, every port handled cargo differently. One standard size, one set of interfaces, changed global trade.
MCP is the same standardization move. Any agent that speaks MCP can plug into any business system that publishes an MCP server. An MCP server sits on top of a business system (core banking, document repository, database) and publishes approved actions the AI can take, like "look up a balance" or "freeze a card." A thin, audited layer stands between the AI and the system.
In 1956, Malcom McLean loaded 58 of these onto a ship called Ideal X and changed shipping across the globe. Source: MoreThanShipping.com
Anthropic released MCP as an open standard in November 2024. OpenAI adopted it in March 2025. Visa's Trusted Agent Protocol and Coinbase's x402 (which Fiserv joined as a founding member of the x402 Foundation in early April) complement MCP for payment-specific scenarios.
A chief risk officer reading the Nymbus docs sees role-based access and audit logging, the same tools already used for human employees, now governing AI agents.
Why it matters: Nymbus is among the first purpose-built MCP implementations for core banking. With both Anthropic and OpenAI behind MCP, the next core provider RFP has a new line item on it.
The Sunday Times · Apr 11, 2026
Lloyds Banking Group has put an AI agent inside its boardroom to sift confidential materials and advise directors on cybersecurity, sustainability, financial analysis, and M&A. Board Intelligence CEO Pippa Begg says the next phase has directors running the agent on laptops mid-meeting, "able to almost interrupt and say: 'Hang on, I think you're falling into this trap.'" Begg calls giving AI a legal vote a "dangerous leap."
Lloyds is also rolling agentic AI across 21 million customer accounts. Its generative AI tools produced £50 million in value in 2025, with £100 million targeted for 2026.
Why it matters: Most boardroom AI conversations to date have been about how to govern the system. Lloyds’ system helps govern the board.
📡 On the radar
A new study (PYMNTS/Velera) finds that members who recently left their credit union are 122% more likely than the average consumer to want AI chat support.
I attended the April 8 "Is AI for CDFIs?" webinar. A third of surveyed CDFIs gave themselves a 1 out of 10 on AI readiness. Pacific Community Ventures, a CDFI, published a responsible AI policy last year. If you’re drafting your own, it’s a well-sourced reference to build on. Full webinar:
American Express launched the Agentic Commerce Experiences (ACE) Developer Kit for AI agent transactions on Amex cards, plus Agent Purchase Protection covering cardmember losses from registered agent errors. The closed-loop network lets Amex underwrite agent risk that four-party networks cannot.
👾 Mythos update
Last week we broke down Project Glasswing and Claude Mythos Preview, Anthropic's unreleased model that autonomously finds and exploits software vulnerabilities. Two updates since:
April 10-12: Regulators in the US, Canada, and the UK all held emergency meetings about Mythos in the same 48 hours.
April 13: AISI published its independent evaluation. Mythos hit 73% on expert-level Capture-The-Flag tasks and became the first model to run AISI's 32-step network attack simulation, 'The Last Ones,' start to finish, succeeding in 3 of 10 attempts. A human expert needs 20 hours.
Source: AISI
AISI recommends cyber security fundamentals for every institution: patching, access controls, logging, and configuration. Mythos is the reason to prioritize them this quarter.
⚙️ Tools & vendors
Primitive: Fintech vet Derek White came out of stealth April 14 with a three-part stack (system integration, 90-day agent builds, governance) and a "Return on Agent Capital" framework for measuring Al value. MX Technologies, which serves 1,700 FIs, is a launch partner co-building an AI-native Growth Agent for deposit acquisition and direct deposit switching.
Abound's "Financial Autopilot": Abound (the Times of India Group's super-app for non-resident Indians, not to be confused with the wonderful Abound Credit Union) launched an AI agent that auto-sends remittances when exchange rates hit a user-set target, and pays Indian utility bills and school fees on schedule. The super-app serves 800,000 Indians living abroad.
Abound, a “superapp” for Non-Resident Indians, launches AI Financial Autopilot in partnership with NEAR AI
Claude Managed Agents: Anthropic now runs the full stack for 24/7 production AI agents (hosting, code sandboxing, credential handling, error recovery) at roughly $58 per month per agent, plus token costs. Cloud infrastructure stocks dropped double digits the same week.
💡 When a member asks an AI for a loan
Two companies made the same bet in the past week (Perplexity on Plaid, Hiro joining OpenAI), six months after the Roi acqui-hire. The major AI platforms want to be the credit union member's financial advisor, and they are moving fast.
The Financial Health Network's 2025 Pulse puts consumer trust in AI financial advice at 12%. Trust in prior channel shifts started low too. Online banking and mobile deposit capture both broke through once the product worked well enough, often enough, that people stopped asking whether they trusted it and started relying on it.
An anecdote is not a trend, but a few more headlines like the $14,000 CPA miss above could move the 12% fairly quickly.
Good enough can win adoption. The 63-70% accuracy benchmarks describe general-purpose models on general-purpose questions. No public benchmark tests an agent with real Plaid data on the specific questions members bring. Perplexity is betting the number goes up.
Credit unions might consider working towards the following:
Publish rate data agents can read. When a member asks Perplexity "what's the best auto loan rate near me," the agent pulls from sources that publish structured data. That means Schema.org FinancialProduct markup on rate pages, or a rate feed that agent-side tools can crawl. Most credit union rate pages are still built for humans. The fastest move is marking up existing pages with structured data this month and monitoring whether they start surfacing in AI answers.
Audit the core's agent interface. Three questions for a credit union's core provider, in writing: what agent-facing capabilities ship today, what the MCP roadmap looks like through end of year, and how permissioning works when an AI (not a human) initiates an action. If the core cannot answer the third question, that is the answer.
Pick a posture on build-inside vs. build-around. Credit unions have two paths. Show up inside Perplexity and OpenAI's answers through structured data (build-around). Or build a member-facing AI layer of their own (build-inside). Most will need both, but the first path is cheaper and faster to start. The credit unions that move on structured data this quarter buy themselves time to figure out the rest.
Big news in agentic finance arrived at both ends of the stack this week. The consumer and infrastructure ends are converging on the same member. Credit unions that wait for a clearer picture are likely to find the picture was drawn without them. So let’s get to doodling.
In a moment where agentic AI is rewriting operational rules, I help credit union leaders sort out what this means for their strategy. Drop me a note at [email protected].
How this newsletter is made: Brent curates the research and writes the analysis. Claude helps with drafting and editing. Published on Beehiiv. ⚡ Alakazam ⚡.
Know someone who should be reading this? Send them the subscribe link.
Hold on one sec while my agent joins the call…
