Welcome to AI for FIs. Each week, this newsletter curates the most important AI developments and explains why they matter for your credit union.

42% of consumers say they're comfortable with AI completing transactions. 80% of Gen Z already use AI for financial planning. Agentic payment rails are taking shape. The industry still lacks a shared standard for verifying that an AI agent has permission to spend on a member's behalf. Credit unions that figure out their piece of this early will define how it works for their members.

This week: Velera's VP named the identity question every CU fraud team should be working on. MSUFCU deployed an agent that completes skip-a-pay requests without a human. Meta classified a rogue agent incident as SEV-1 (a highest‑criticality incident severely impacting many users or core services). And banks passed credit unions in member satisfaction for the first time in ACSI history.

The issue wraps with a deep dive, Your Credit Union's First AI Move Should Be Boring, on why the smartest first deployment is the one nobody notices.

PYMNTS.com · Mar 18, 2026

Velera VP Elizabeth Wadsworth warned that AI-powered fraud now targets every stage of the member lifecycle. Voice cloning, synthetic identities, and behavioral simulation are defeating traditional verification. Passkeys, device-bound authentication, and risk-based decisioning are the emerging countermeasures.

Why it matters: Credit unions will soon need to authenticate AI agents acting on behalf of members, a use case most identity frameworks were never designed for. Wadsworth's framing puts a timeline on a problem many CU leaders have not yet scoped.

The Credit Union Connection · Mar 23, 2026

MSUFCU deploys a skip-a-pay agent that completes the request end to end. No human involvement. Deterministic state machine with an agentic conversational layer. Integrates with Symitar, Fiserv DNA, and Keystone. CTO Ben Maxim says they're prioritizing operationalizing agents over experimentation.

Why it matters: This is one of the first production AI agents in credit unions handling a full transaction cycle, from member request to core system update. It sets a benchmark for what autonomous processing looks like and raises immediate questions about audit trails, error handling, and member consent.

TechCrunch · Mar 18, 2026

Meta's AI agent posted an unauthorized response on an internal forum and exposed sensitive company and user data to unauthorized engineers for two hours. Meta classified it as Sev 1. A Meta safety director's agent deleted over 200 emails despite instructions to confirm before acting. Both incidents happened while Meta aggressively pursues agentic AI deployment.

Why it matters: Any credit union deploying AI agents faces the same risk profile. An agent acting outside its approved scope can put your members at risk and trigger NCUA or state examiner scrutiny. This incident is a concrete case study for boards asking "what could go wrong."

When Paul Hudson was CEO of Sanofi, he walked into a meeting where 32 people had gathered. He asked what they were working on. The leader said, "We're deciding how not to give you the data." Hudson pressed. The leader explained: "It's not wise to give it to you without us telling you what to think about it."

Hudson said it scared the hell out of him. His team in Germany built a resource deployment agent that could analyze the same operational data without career anxiety or political incentive. The agent told him where capital was sitting idle. He redeployed close to €1 billion.

That’s Paul’s story. But most of us earn our way by starting somewhere boring.

Lamont Black at Wide Open Ventures got the sequence right: align your leadership team on AI first. Then the next step is practical.

Lake Michigan Credit Union (~$16B in assets in assets) looked at its lending pipeline and found that every loan file triggered the same massive checklist of tasks, most of which didn't need to happen. The system generated work regardless of whether the work was necessary.

They deployed an AI agent (software that can work through a multi-step task, follow rules, and take limited action inside a defined workflow) to pre-screen every file, run the routine steps, and clear the unnecessary tasks before a human ever sees the loan.

Fulfillment labor on home equity files dropped from 55 hours to 10. The team saved 3,400 hours in less than four months. Noel Watts, LMCU's SVP of Lending Operations, said the same disclosure team now handles 40% more files with the same headcount. They're offering same-day disclosures.

The agent coordinates the workflow. Humans make the judgment calls. And the work it handles is structured and reviewable by the team it supports.

Klarna automated two-thirds of its customer service conversations, replacing the equivalent of 700 human agents. They projected $40 million in profit improvement. Then service quality dropped. CEO Sebastian Siemiatkowski admitted the company had prioritized cost over quality, and they started bringing humans back.

When you move fast on automation, you can measure the speed gains. You can't measure the expertise you stopped building, or what’s been lost along the way. Structured internal workflows like loan processing can be different in kind from member-facing conversations where empathy and judgment carry the weight of the interaction.

A useful operating model is what Boston Consulting Group calls "graduated autonomy" -

LMCU followed this pattern. They watched, measured, and expanded based on what they learned.

PwC extends this thinking to org design. When you automate entry-level roles, you can end up with managers overseeing AI systems and junior staff who never learn the business. Who becomes your next SVP of Lending if nobody built intuition by doing the work?

The people stay and the grunt work goes to the machine. The expertise keeps building as long as the work is redesigned so junior staff still build judgment along the way.

McKinsey estimates that only about 6% of respondents in their latest survey report meaningful bottom-line impact (defined as 5% or more of EBIT attributable to AI). Most enterprise AI spending goes toward accelerating existing processes. The workflows stay the same. This is a missed opportunity.

Hudson's 32 people were filtering his information. His agent told him what they wouldn't. Shopify's CEO now requires teams to show why AI can't do the work before they can request headcount. Operational proof first. Bigger moves second.

Govern the work carefully. Measure what changes. Then earn the next move.