For over a century, credit unions have operated on a principle that the rest of the business world is just now arriving at: the people closest to the work should have the context to make decisions.
A branch manager at a 200-person credit union doesn’t sit atop a hierarchy — she knows half the members by name. The compliance officer isn’t buried under seven layers of reporting — he’s two doors down from the CEO. Information doesn’t route through management layers because the organization is small enough, and the culture collaborative enough, that it already flows naturally.
That operating model — flat, cooperative, community-driven — was considered a limitation for decades. Too small to compete. Too simple to scale. Too human to automate.
It turns out it was a head start.
The Shift Nobody Expected
Jack Dorsey and Roelof Botha published an essay in March called From Hierarchy to Intelligence. In it, they trace 2,000 years of organizational design — Roman legions, Prussian general staff, American railroads, Frederick Taylor’s scientific management — and arrive at a thesis: hierarchy is an information routing protocol, and AI can now replace it.
The argument is that a manager’s primary function is not leadership — it’s information routing. They aggregate context from below, translate it for above, and push decisions back down. The Roman centurion coordinated 80 soldiers not because he was the best fighter, but because humans can only hold about eight direct relationships before communication breaks down. Two thousand years later, every org chart on earth still runs on that same constraint.
AI changes the equation. A continuously updated model of the entire business — what Block calls a “world model” — can hold it all. Instead of information flowing up through layers, getting compressed and distorted at each level, the model maintains it. People operate at the edges, close to the work, with the full context of the organization available on demand.
Block then reorganized around this thesis and cut over 4,000 positions. The stock surged 22% in the days that followed.
This isn’t the Spotify squad model. It isn’t Holacracy. Those experiments failed because they removed hierarchy without replacing the information routing function. What’s happening now — at Block, at Coinbase with their multi-agent decision systems, at Ramp with agentic financial operations — is replacing the function itself.
But here’s what nobody in Silicon Valley is saying: the organizations closest to this future aren’t in San Francisco. They’re the 4,500 credit unions that have been running flat, cooperative organizations since before anyone called it a competitive advantage.
Why Cooperatives Were Already There
The intelligence-native model, as Dorsey describes it, pushes decision-making to the edges. Individual contributors get direct context instead of filtered management summaries. Player-coaches build alongside their teams instead of managing from above. The world model provides the organizational knowledge that used to live exclusively in managers’ heads.
Credit unions have been doing a version of this for generations — not because they chose to, but because their structure demanded it.
A $500 million credit union with 120 employees doesn’t have the luxury of a six-layer management hierarchy. The loan officer, the BSA analyst, and the branch manager operate with direct access to institutional knowledge because there’s nobody between them and the information. The culture is collaborative by necessity and by mission — member-owned, democratically governed, built to serve rather than extract.
The problem was never the organizational model. The problem was that the technology stack forced information into silos that the culture didn’t create. Six disconnected systems for a single compliance investigation. Member data trapped in a core processor that speaks a dialect no other system understands. Institutional knowledge locked in the heads of people approaching retirement, with no way to capture it at scale.
The intelligence-native model doesn’t ask credit unions to change their culture. It gives them the technology to finally remove the barriers their culture never wanted.
The Honest Counterargument
I can hear the pushback already, and it’s fair: credit unions face real structural headwinds that make this harder, not easier.
Legacy core processors are the obvious one. Symitar, Corelation, DNA — they all structure data differently, they resist integration, and replacing them is a multi-year, multi-million-dollar project that most CUs can’t stomach. An intelligence layer that needs access to normalized member data doesn’t do much when the data lives in a 30-year-old system with a proprietary API.
IT budgets are thin. NCUA data shows the median credit union under $1 billion in assets operates with an IT budget that would make a Series A startup wince. There’s no AI team. There’s often no data team. There’s an IT director wearing four hats.
Boards are risk-averse. Regulatory oversight makes credit union boards cautious by design — which is generally good governance, but it means AI adoption faces more friction than at a venture-backed fintech that can move fast and ask forgiveness later.
The workforce skews older. The institutional knowledge that makes a credit union’s context layer valuable is concentrated in people who may not be comfortable working alongside AI systems.
These are real constraints. But I’d argue they’re surmountable — and in some cases, they’re advantages in disguise.
Regulatory discipline means credit unions already have the governance frameworks, audit expectations, and oversight culture that make AI deployment safe. A venture-backed startup has to build those guardrails. A credit union already lives inside them.
The thin IT budget means credit unions can’t afford to build bloated, over-engineered AI platforms — which forces them toward the right architecture: purpose-built agents that do specific jobs well, operating under human oversight, with clear audit trails. That’s exactly what the intelligence-native model calls for.
And the older workforce? Those are the people whose institutional knowledge is the world model. Every investigation they’ve run, every exception they’ve handled, every examiner relationship they’ve navigated — that’s the training data. The question isn’t whether they’ll be replaced. It’s whether their knowledge gets captured and compounded before they retire, or walks out the door with them.
What the Architecture Needs to Look Like
If you strip away the company names and the branded terminology, the architecture that Block, Coinbase, and Ramp are converging on has four layers. Any credit union evaluating AI infrastructure should be looking for the same pattern:
A capabilities layer — purpose-built agents with defined skills, not general-purpose chatbots. A BSA compliance agent that knows how to investigate suspicious transactions is fundamentally different from a chatbot that can answer FAQs. The agent needs SOPs, validation gates, and human oversight at every consequential decision point. It needs to produce audit trails that satisfy examiners, not just management dashboards.
A context layer — the institutional world model. This is the hardest to build and the most defensible once built. It requires cross-core data normalization so that agents can understand member data regardless of which processor it lives in. It requires semantic search over institutional knowledge — not just structured queries, but the ability to find the relevant examiner precedent, the applicable policy exception, the similar case from three years ago. And it requires continuous updates — CDC pipelines, change data capture — so the model reflects reality, not last quarter’s snapshot.
This is also where the compounding happens. Every investigation an agent completes, every loan decision that gets reviewed, every member interaction that resolves successfully — these outcomes can feed back into the model. The context gets richer. The agents get more accurate. The humans spend less time correcting and more time judging. Andrej Karpathy recently demonstrated this pattern with AutoResearch, an open-source tool that runs experiments in a loop, keeping only the changes that beat the current best result. The same principle applies to operations: the organization becomes a learning loop, not just a reporting structure.
A control plane — the governance and oversight layer. Every agent interaction needs authentication, rate limiting, audit logging, and the ability to shut things down instantly. You cannot give agents real authority in regulated financial services without infrastructure that lets you revoke it in seconds. This is where credit unions’ existing regulatory discipline becomes an asset — they already think in terms of controls, exceptions, and escalation paths.
A command surface — where your best people observe, direct, and intervene. Think of it as mission control for an AI workforce. Every active task, every decision point waiting for human approval, every agent that flagged something unusual. The intelligence lives in the system. The humans operate at the edge — making the calls that require judgment, cultural context, and ethical reasoning. Exactly where Dorsey says they should be.
The Capital Structure Advantage
I wrote previously in When Markets Stop Funding the Future about how Chamath Palihapitiya’s terminal value thesis accidentally made the strongest case for credit union capital structure in a generation. His argument: AI compresses every competitive advantage to a five-year window, which collapses the terminal value that supports most equity valuations. Patient, long-duration capital becomes the scarcest and most valuable resource.
The same logic applies to organizational transformation.
Block can pursue the intelligence-native model because it has $30 billion in market cap and a visionary CEO willing to cut 4,000 jobs. Coinbase can build multi-agent decision systems because it has thousands of engineers and a $60 billion valuation. Ramp can automate financial operations because it raised at a $32 billion valuation with nearly 2,000 employees.
The average credit union has roughly 50 full-time employees, according to NCUA aggregate data. It doesn’t have hierarchy to dismantle. It has the opposite problem: not enough people to do the work that regulators, members, and market conditions demand. The compliance analyst toggling between six systems isn’t a middle manager routing information — she’s a specialist drowning in work that should have been automated a decade ago.
But credit unions have something those $30 billion companies don’t: a capital structure that doesn’t punish long-term bets. A board that can authorize a five-year technology investment because the people who own the institution aren’t pricing terminal value — they’re pricing whether the credit union serves them well today and will serve them well tomorrow.
The cooperative model was the original rejection of hierarchy-as-extraction. The intelligence-native model is the infrastructure that lets it scale.
Five Questions for Your Next Board Meeting
If you’re a credit union executive reading this, the strategic question isn’t whether the intelligence-native model is real — Block, Coinbase, and Ramp are settling that debate with billions of dollars. The question is whether your institution is positioned to adopt it. Here’s where to start:
-
Can your agents access normalized member data across core systems? If your data is locked in a single processor’s proprietary format, no AI system — regardless of vendor — can build the context layer that makes intelligent automation possible. Cross-core normalization isn’t a nice-to-have. It’s the foundation.
-
Do you have an audit trail for AI-assisted decisions? Examiners will ask. If your AI vendor can’t produce a complete, timestamped record of every decision an agent made, every piece of evidence it considered, and every human who approved or overrode it, you have a compliance gap, not an AI strategy.
-
Is your institutional knowledge being captured, or is it retiring? The 30-year BSA analyst who knows every examiner’s preferences, every edge case, every unwritten policy interpretation — that knowledge is your world model. If it’s not being systematically captured and indexed, it’s a depreciating asset walking toward a retirement party.
-
Does your AI infrastructure have a kill switch? Not a “turn it off and call IT” kill switch. A real-time, per-agent, per-task control plane that lets authorized staff halt any automated process immediately, with full audit logging of the intervention. This is non-negotiable in regulated financial services.
-
Are you buying tools or building capabilities? A chatbot is a tool. A compliance agent with access to your institutional context, operating under human oversight with full audit trails, that gets better with every investigation — that’s a capability. The difference is whether it compounds or just costs.
What Comes Next
The intelligence-native company is not a theory anymore. It’s being implemented across industries, from 6,000-person restructurings at public fintechs to autonomous research loops in AI labs. The pattern is consistent: build a world model from proprietary data, deploy specialized agents against defined workflows, maintain human oversight at the edges, and let the system compound.
The needs of hierarchy aren’t disappearing — they’re shifting. From routing information to setting direction. From managing workflows to judging outcomes. From controlling access to cultivating judgment. The organizations that understand this shift first won’t just adapt. They’ll define what the next era of work looks like.
Credit unions were built for community, not hierarchy. The intelligence-native model is the infrastructure that makes community scale.
This is Article 35 of the Runline Insights series. Previously: When Markets Stop Funding the Future | Your Data Isn’t Your Moat
