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 your credit union.

Last week, Jack Dorsey published an essay arguing AI should replace middle management. The same day, Marc Andreessen said every large company is overstaffed by up to 75 percent. Silicon Valley is building one version of the future org chart. Your credit union gets to build a different one.

This issue: new agent payment infrastructure from Coinbase and Visa, a credit union governance framework worth borrowing, a peer-reviewed study on AI flattery, and Google’s new top 3 open model built to run on-site with no data leaving the building.

At the end, a deeper dive into what Dorsey gets right, what he misses, and questions credit unions should be asking themselves right now.

coindesk.com

Coinbase's x402 protocol for AI agent micropayments moved under the Linux Foundation as an open standard. Google, Stripe, Visa, Mastercard, American Express, AWS, and Fiserv are among the initial supporters. The protocol handles high-frequency transactions worth fractions of a cent.

Why it matters: Fiserv and Visa are on the founding member list. If x402 becomes the SSL of agent payments, it will touch every FI's transaction infrastructure.

ABA Banking Journal · Apr 1, 2026

Two separate surveys found growing numbers of consumers use AI and non-traditional sources for financial guidance. Wells Fargo found 19% of adults and 38% of Gen Z use AI for financial advice.

Why it matters: 38% of Gen Z already seeks AI-driven financial advice. If that AI does not mention a credit union's products, it sends members elsewhere.

CreditUnions.com · Apr 6, 2026

CreditUnions.com profiled how Members Cooperative Credit Union governs AI use internally. Staff request access case by case, only Microsoft Copilot is approved, member data cannot be submitted, and all AI use requires human review and disclosure.

Why it matters: This profile shows what one credit union’s AI governance process looks like in practice. Very much worth reading if you're building your own.

A National Science Foundation-funded study tested all 11 major AI models and found every one of them sycophantic. They affirm users' actions 50 percent more than humans do. Even when users describe manipulating someone, the AI validates them. People who interacted with sycophantic models became more convinced they were right and less willing to repair relationships.

Users consistently prefer the flattering models. They rate them as higher quality and more trustworthy. A 2025 study from HBR and IMD Business School found the same pattern in a business context: executives who used ChatGPT for forecasting became more optimistic, more confident, and produced worse predictions than those who consulted human peers.

If credit union staff use AI for member interactions or internal decisions, the tool may be reinforcing bad judgment right now. One offset is to stop prompting an AI for a single answer. Assign it a role that forces disagreement. Ask it to argue against your position. Ask it to find the three weakest points in your plan.

Bridgewater Associates uses AI systems that generate investment theories and then deploy separate AI techniques to aggressively interrogate those theories, iterating between creation and critique. The principle applies at any scale. The agents surface strategies no single perspective would produce. Knowing when the AI agrees because you are correct versus because it was trained to is a skill worth developing.

Last week, Jack Dorsey and Sequoia Capital's Roelof Botha published a joint essay arguing that corporate hierarchy exists to route information, and AI now does that. The same day, Marc Andreessen told the 20VC podcast that every large company is overstaffed by 25 to 75 percent and AI gives them a "silver bullet excuse" for cuts that were coming regardless. Block, the fintech behind Square and Cash App, already acted on this in February, cutting 4,000 employees (forty percent of its workforce) and replacing its management layers with a system that tracks decisions and progress in real time. Three roles remain: builders, outcome owners, and player-coaches.

AI can absorb the coordination work that fills many managers' days: status updates, project tracking, report compilation. PwC found that organizations need fewer managers focused on routine oversight as AI agents take on routine tasks. Dorsey and Botha make a second argument worth sitting with: hierarchy creates delay. Every “let me check with my boss,” every meeting to schedule a meeting adds days to decisions that should take hours. A member who needs an answer today gets one next week. AI can compress that. Most organizations are slower than they need to be, and middle management is where that slowness accumulates. The org chart is changing. It should.

Dorsey's framing reduces management to information routing. If that were the full picture, he would be right to cut it. But good managers do something different. They coach staff through situations that do not fit a script, read behavioral patterns data does not capture, and translate boardroom strategy into branch-level execution while carrying ground-level reality back up. When one person uses words that mean something different to someone else, a good manager catches the gap. An AI system might widen it.

The research backs this up. Filene Research Institute studied 94 top-performing credit union middle managers and found them defined by fast contextual learning (87% above average), intuition over raw data (79%), and a willingness to push back on leadership when something on the ground does not match the plan (73%).

The retention data from the same body of research reinforces this. Two out of three young CU professionals rated their supervisor relationship above cash compensation. Credit unions with the team cultures these managers build outperform peers by almost 40 percentage points in Return on Assets.

The org chart still needs work. In Issue 8, we covered PwC's hourglass model for org design: keep the middle lean, use AI to strengthen early-career talent, free leaders for judgment and coaching. Dorsey's Block went in a different direction. They cut the middle entirely and bet on a system to do what managers used to do.

Jess Rimington and Joanna Levitt Cea's research in Beloved Economies: Transforming How We Work adds depth to this. Their 8-year study found that organizations generating standout success do it through shared decision-making and sourcing from multiple ways of knowing.

An AI can simulate multiple viewpoints, but every persona channels through the same substrate: language and probability. Lived experience, intuition, and relational knowledge built from years in a specific community are not something a model can tap its training data to replicate.

Block answers to shareholders. Credit unions answer to members, and might start with different questions: Which functions are information routing? Those are candidates for AI. Which depend on coaching, judgment, and institutional knowledge? Those need investment. Where are your middle managers spending their time? How will you bring them into this conversation instead of making decisions about them without them?

The org chart is going to change. Your managers should be part of that conversation.