Welcome to AI for FIs, from Dixon Strategic Labs. This weekly-ish newsletter tracks agentic AI and explains what it means for community banks and credit unions.
Hello! A month ago I lamented on LinkedIn that hearing “you’re right” all day from AI had me close to unholy rage, and plenty of people said the same.
My fix was a rule telling my agents to verify a claim before they agree with me. When one writes “you’re right” anyway, a program running outside the agent blocks the reply and makes it try again.
One of my agents wrote “You’re right,” and the program booted the reply back.
The agents in this week’s stories handle people’s money, and each one has a different setup for what it can do on its own.
In a bill-paying test at Cornell Tech and Yale, the approval codes for raising the agents’ spending limit went to an inbox the agents could read.
At Southern Cross Credit Union in Australia, regular callers now get their balances from a voice agent. The banking software supplies every number it reads.
At MSU Federal Credit Union, a new agent will help members who are falling behind on a loan. It changes the loan’s terms only after the member says yes.
Source: CoinDesk, Nov. 1, 2026.
Photo: Kevin Ku / Unsplash, via CoinDesk.
Researchers at Cornell Tech and Yale gave AI agents three one-cent bills to pay and a wallet that could spend one cent a day. The agents could also read the wallet owner’s email.
The experiment used cryptocurrency wallets from Circle, each holding 10 USDC, a dollar-linked token. Paying all three bills, three cents in all, required a higher cap.
To raise the cap, the wallet emailed a one-time code to the owner’s inbox, the same inbox the agents could read.
They retrieved the code and submitted it themselves.
The researchers ran the setup 20 times per model. Nine of ten models raised the cap at least once, and six did it in at least 19 of 20 runs. The same model could stay within the cap on one run and raise it on another.
The same thing can happen when a financial institution emails a code to approve a higher limit and a member’s AI agent can read that inbox.
The researchers built Countersign, a prototype that ties each approval to the specific action and to a person’s physical interaction with a phone.
Source: iTnews, Oct. 27, 2026.
Some members of Australia’s Southern Cross Credit Union phone several times a day to see whether a pension or another regular payment has arrived.
Those calls now go to Nova, an AI voice agent. Nova verifies the caller, then software fetches the balance from the credit union’s records and reads it out, so the AI model never generates the number.
Nova passes frustrated or distressed callers to an employee along with the conversation so far.
The credit union has 70 employees and more than 20,000 customers, and it treats its phone team as its largest branch.
As it grew, average waits rose from under two minutes to nearly eight. Its chief customer officer wanted people reporting a scam to reach an employee without waiting.
Southern Cross built Nova with Salesforce engineers in six weeks. Two of them spent time at head office hearing how staff talk with customers and discussing how much humor and empathy Nova should show.
Many customers have never memorized their account numbers, so Nova understands the nicknames and product codes they use.
The chief customer officer says regulars who call for balances “love speaking with Nova,” including some the credit union expected to hesitate.
A Salesforce case study says Nova handled 7% of all calls in one of its first months, work equal to about 4% of staff hours. Of the callers Nova verifies, 95% get their balance, and the credit union is still working toward its goal of verifying 80% of callers.
Source: CMOtech, Nov. 6, 2026.
Screenshot: Constant AI product demonstration with a fictional member.
A member falling behind on a loan will be able to ask an AI agent at MSU Federal Credit Union for help, and the agent can change the loan’s terms once the member agrees.
The agent, Nia, comes from Constant AI, which builds AI agents for loan operations. Constant says board reporting and repossession notices usually follow once a loan reaches 60 days past due, and Nia works on loans before that point.
Requests for help can sit in back-office queues while a borrower falls further behind, CMOtech reports.
Credit unions set money aside in reserves to cover loans that may go unpaid. “Every dollar we set aside in reserves is a dollar that can’t be invested back into serving our members,” said April Clobes, MSU FCU’s CEO.
Nia asks what caused the hardship, how long it is expected to last, and what the member can afford. It reviews the member’s account, applies the credit union’s assistance policy to choose a program, and presents the terms.
In Constant’s demo, a fictional borrower’s work hours were cut. Nia offers a deferment that moves two payments to the end of the loan, with interest still accruing.
After the member agrees, Constant says, Nia posts the new terms to the credit union’s core banking system, which holds member accounts. It logs each offer, outcome, and change. Cases that need judgment can stay with human collectors.
Explainer: Know Your Agent
A doodle I doodled.
Know Your Agent, or KYA, applies the idea behind Know Your Customer, the identity checks before opening an account, to AI agents. It ties each agent to the person or institution it works for and to the actions that party has allowed.
Ken Huang, an AI security researcher, splits KYA into four parts in an October 29 analysis:
Identity links an agent to the person or institution it represents. FIS says banks need a way to tell a customer’s real agent from a fraudulent one.
Permission lists the accounts, actions, and amounts the agent may use. JPMorgan Chase is applying 15 safeguards across the firm, covering how agents prove who they are and how permissions work.
Enforcement tests that permission in the system that makes the change or moves the money. Goldman Sachs builds its rules into the environments its agents work in, with gates around tools and sensitive systems.
Evidence records the authorization, action, and outcome. Starting November 24, Wells Fargo’s online access agreement requires customers to keep evidence that they revoked an agent’s access.
In the wallet test, the one-cent cap was the permission, and the emailed code was the enforcement. The enforcement failed because the agents could read the email.
On November 2, the Hong Kong Monetary Authority, Hong Kong’s banking regulator, launched Project SAGE, a program on AI agents in finance. It includes work on agent identity and on what a person has authorized an agent to do.
On the Radar
Kitsap Credit Union says AI reviews all its loans nightly. It built the system with D8TAOPS. The system applies policy and risk checks and sends findings to staff.
Claude can analyze spending in linked bank accounts. Anthropic’s U.S. beta uses Plaid for read-only access. OpenAI added similar account links to ChatGPT in May, after acquiring personal finance startup Hiro. Makes you wonder what would happen if Claude or Open AI got a banking charter, eh?
Chime is building connections for members’ own AI agents. Members will control which agents reach their accounts. JPMorgan Chase, Citi, and Bank of America restrict outside agents.
AI agents cut decision time on payment holds by two-thirds. Fiserv reports the result for its Clover merchant payments system. Incorrect risk flags fell by more than 40%.
U.S. bank examiners’ new AI guidance names an autonomous payment agent as its example of critical risk. The OCC, which supervises national banks, directs examiners to weigh potential consequences and autonomy, along with bank size and complexity.
A UK bank’s AI fraud agents have now worked over 10,600 live cases. Lloyds, which described the agents in June, reports decision accuracy above 90% in September. Staff use roughly half as many systems during a fraud call.
Agents cut a bank’s research on a small-business loan applicant to under an hour, from eight to ten hours. They pull company records, identify owners, and search sanctions lists. Bretton AI’s case study reports the result for First Internet Bank.
Nous switches utility and insurance providers for UK households. It plans to shop for savings accounts and consolidate card debt.
Anthropic released Claude Fable 5.2, and it’s a doozy. In a test by Oracle, two other AI models came up with debt totals five to eight times too high. Fable 5.2 got them right. Ethan Mollick shares some examples of its impressive capabilities, including coding this video from one prompt in one shot. I’ll be honest, this video demonstration really knocked me out. Enjoy:
Video: Claude Fable 5.2, made for Ethan Mollick. Fable asked Mollick to reveal nothing about it until viewers had watched to the end.
I’m curious who out there in financial services are using high-performing models like Claude Fable 5.2 or OpenAI’s GPT-6 Astra in their work. And what they’re using it for. Is that you or someone you know? If so, drop me a line at [email protected]. I’d love to hear about it.
By Tom Fishburne via ModelBehavior
Work with me
I help leadership teams at credit unions, community banks, and community development financial institutions choose which work to try AI on first.
Since August I've been working with the executive team at a $1 billion credit union. I audited the AI running in its vendor systems, spent three days on site watching how staff do their work, and planned two pilots that start in January.
I'm now designing AI training for all of its staff in 2027.
If your team is choosing where to start, write to me at [email protected] and tell me what you're considering.
How this newsletter is made: Brent curates the research and writes the analysis, with AI models helping with research, drafting, and editing.
⚡ Alakazam ⚡.Show someone you love them by having them subscribe to AI for FIs here.
