By Omid Toloui, MBA, MPH, Principal, Red Creek Partners, Adjunct Professor, UCLA Anderson School of Management
Most companies are asking how to deploy AI. The harder question is whether the company you built is still the right shape.
Is our company just going to be completely irrelevant in the next coming years?
— Jack Dorsey, Block
That is the question Jack Dorsey has been sitting with, and he is willing to ask it out loud. The same week he said it on a Sequoia podcast, he and Roelof Botha published “From Hierarchy to Intelligence,” explaining why Block laid off 40% of its workforce earlier this year and is rebuilding around what they call an Intelligence Layer.
Most coverage focused on the layoff. The more interesting story might be what they are building toward, and what it implies about leadership for everyone else.
For two years, the loudest AI conversation in healthcare has been about productivity. Will it write our notes faster? Can it summarize our claims? Will it shave a few minutes off prior auth? These are useful questions. They are not the leadership question.
The leadership question is more uncomfortable. Is the company we built still the right shape?
The protocol that ran the world for 2,000 years
The Block essay opens with something I had not fully appreciated. The org chart we use today, the boxes-and-lines pyramid every Fortune 500 company runs on, descends in a straight line from the Roman army. Eight soldiers in a tent, eighty in a century, five thousand in a legion, each layer led by a named commander. The Prussians refined it after Napoleon. American railroads adopted it in the 1850s when informal management started killing people in train collisions. Frederick Taylor optimized what happened inside it.
Two thousand years of iteration on the same problem. How do you coordinate large numbers of people when the only available technology for routing information is other people?
That is what middle management is. An information protocol born of the fact that, until very recently, humans were the only thing that could read a status report and decide what to escalate.
That protocol is what AI now competes with. Not the work. The protocol.
What changes when the protocol changes
When every email, board memo, contract draft, audit finding, claims review, performance review, and meeting recording exists as a structured artifact, an AI system can do what a layer of management used to do, and several things it never could.
The first shift is access. In the old model, the CEO got the freshest picture of the company because everyone reported up. In the new model, anyone can query the company in real time. Dorsey describes board members querying the business directly in board meetings, analysts asking earnings questions on their own timeline, and frontline employees pulling the same context their VP would have had. The asymmetry of information that the org chart was built around stops mattering.
The second shift is direction. In the old model, product managers guess what customers want, debate it in meetings, and ship a roadmap a year later. In the new model, when a customer can describe a feature that does not exist and have it composed on the fly, every gap they hit becomes the roadmap, in real time, generated by usage rather than guessed at by committee.
The third shift is how work gets done day to day. Two months ago at Block, a meeting meant a slide review. Now people show up with a working prototype they built that morning, modified live in the room. The cost of being wrong on a path drops close to zero, because the next path can be tried in an hour. Focus matters less than range. Conviction matters more than process.
I feel this firsthand. Concepts I would have written down in a notebook a year ago now turn into working prototypes the same day. The constraint used to be my technical capability. Now it is my creativity. Technical chops still matter, just less than they used to.
This is where the new roles fall out. Block is normalizing down to three. ICs (individual contributors) who build and operate, augmented by agents so a single person can do the breadth of what a team used to. DRIs (directly responsible individuals) who own a customer outcome end-to-end, assembling capabilities and ICs around it. Player-coaches who develop the people around them by doing the work, not by directing it. The middle management layer whose entire job was routing information is gone.
What does not go away is human judgment, but it concentrates where it actually matters. At the edges, on the things models cannot do well on their own. Ethical calls. Genuinely novel situations. Reading a room. Trust between people. High-stakes decisions where the cost of being wrong is existential.
Even ethical reasoning is starting to get help. Earlier this year I published NAIMA, an open-source moral framework for AI agents that draws 20 principles from seven faith traditions into a consultative sub-agent other agents can query when they hit a consequential decision. The premise is that as we put more agents in front of more decisions, the human work moves upstream. Encode the values once, with care, and the agent reasons against them at scale. Humans still set the principles. The system applies them.
As the Block essay puts it, humans operate at the edge because the edge is where intelligence makes contact with reality.
The diagnostic question for whether any of this is even possible at your company is the sentence I keep returning to.
AI doesn’t augment your company. It reveals what your company actually is.
If your company does not understand something deeply, AI will eventually expose that there was nothing there to understand. If it does, AI compounds the understanding into something nobody can replicate.
Why healthcare incumbents are especially exposed
Three things make the largest healthcare companies more exposed to this shift than most industries, and none of them are anyone’s fault. They are structural.
First, healthcare hierarchies are built for risk management, not velocity. Every layer exists for a reason. Compliance, regulatory, clinical safety, audit trails, board oversight. Stripping middle management is not a memo. It is a multi-year unwinding of operating procedures that protect the company from existential failures: lawsuits, regulatory enforcement, patient harm, and reputational collapse. The protective value is real. The speed cost is also real.
Second, the signal is fragmented in two directions. Externally, no single healthcare player sees the whole picture. A large payer sees claims but not the visit. A health system sees the visit but not the rest of the patient’s life. A producer sees the prescription, not the outcome. A health tech vendor sees the system of record but not how clinicians actually behave inside it. Building a customer world model requires data nobody currently has on their own.
Internally, the same fragmentation shows up. The Block model assumes you can also build a coherent world model of your own company, your operations, performance, priorities, and dependencies. Most large healthcare incumbents grew through acquisition. They run on stitched-together ERPs, regional variations, overlapping ops, legacy systems from companies they bought a decade ago. Knowing what your own company is doing on any given day is itself a real engineering problem before you have done a single thing with AI.
Third, the AI conversation inside most large healthcare organizations is still about copilots. Each employee gets a chatbot. A back-office process gets automated. The new layer of healthcare AI, however, is forming outside these organizations entirely. According to Menlo Ventures’ 2025 Health AI report, 85% of generative AI vendor spending in healthcare flows to AI-native startups rather than legacy healthcare IT platforms. The pattern extends beyond IT vendors. Across the incumbent landscape, AI is being bolted onto a 30-year-old tech stack and a 60-year-old org chart. The two are not competing on the same field.
I want to be careful here. None of this means UnitedHealth, CVS, HCA, or Pfizer is suddenly irrelevant, nor are any of the tech giants moving in from the outside (Google, Microsoft, Amazon) about to replace them tomorrow. These are companies with capital, scale, regulatory licenses, distribution, and relationships nobody can replicate overnight. The point is not that incumbents lose. The point is that the leaders who do not sit with the question Dorsey is sitting with are going to find themselves competing on a board they did not realize was being redrawn.
What AI-native looks like, and why it matters for leadership
The most interesting thing about the AI-native healthcare companies right now is not their valuations. It is how they are organized.
Abridge is in 200-plus health systems including Northwell and UPMC, and is now extending into prior authorization. OpenEvidence is delivering point-of-care clinical answers to roughly 40% of US physicians, reaching 18 million consultations in December 2025 alone, and is valued at $12 billion. Hippocratic AI is in production at WellSpan, UHS, and University Hospitals for patient outreach and post-discharge care, with UHS reporting patient ratings of 9.0 out of 10. None of these companies built a traditional org and then bolted AI on top. They built the intelligence layer first, and the org formed around it.
Fortune ran a piece in back in August describing one AI-first healthcare company that replaced a 10-person engineering team with three people overseeing AI agents. A product owner, a prompt-fluent engineer, and a systems architect. That is the IC and DRI structure Dorsey described, just landed in healthcare.
The startups are not winning because they are smarter. They are winning because they are unencumbered.
This is the asymmetry that money cannot close. An incumbent can buy an AI-native company. An incumbent can hire AI-native talent. What an incumbent cannot do, easily or quickly, is dismantle a 60-year-old hierarchy and rebuild around an intelligence layer while staying compliant, public, and operationally intact.
What this means for Fortune 500 leaders
There is no clean playbook for an incumbent. Anyone pitching you one is not paying attention.
What I would offer instead are four questions I think every senior healthcare leader should be sitting with right now.
What does our company understand that nobody else does, and is that understanding getting deeper every day? If the honest answer is “we have a lot of data and we run a lot of process,” that is a serious problem. Data is not understanding. Process is not understanding. Understanding is a model of your customer that compounds.
Where is our hierarchy actively slowing information rather than protecting it, and would the people closest to the customer have what they need to act if we removed it? Removing layers without empowering the edges produces chaos, not velocity. Empowerment means access to the company’s full context plus tools to actually build, prototype, and ship without waiting for permission. The Block model only works because everyone at the edge can see what the model sees and act on it.
Are we doing AI-as-copilot or AI-as-architecture? Most enterprise AI rollouts give every employee a chatbot, automate a process or two, declare efficiency gains, and move on to the next initiative. That can produce real productivity wins and a quarter or two of clean earnings calls. It will not change your competitive position in five years.
What would our company look like if we were starting today with 100 people? When Dorsey’s leadership team ran this exercise after the holidays last year, the answer was uniformly that the company would not look anything like it does today. Asking that question of your own organization is uncomfortable and clarifying.
A reasonable counter-argument deserves attention as well. BCG’s Nick South recently argued that the orchestration layer of large companies will grow, not shrink, because someone has to manage the human-agentic workforce. Yale’s Tristan Botelho expects middle managers to redefine their role rather than disappear. They may be right. The honest read is that nobody knows yet, and the leaders making confident proclamations in either direction are mostly selling something.
What I do think is settled is this: The cost of doing nothing while you wait for clarity is going up every quarter.
What I am doing about it
The pace of change is neck-breaking. People ask me how to keep up. My answer is the only one I have. Build. Experiment. Learn the capabilities by doing.
For the past several months I have been running a small intelligence layer of my own.
The setup is unremarkable. A Mac Mini in my home office runs local models. For fast, private inference I rotate between deepseek-r1:14b and qwen3:14b. For harder reasoning I run larger models locally, and reach for cloud models when the task warrants it. On top of that sits a recursive knowledge base in the spirit of the LLM wiki pattern Andrej Karpathy described, where every new input (an article, a meeting transcript, a paper, an email thread, a piece of my own health data) gets ingested, cross-referenced against everything already there, and woven into a graph of entities and concepts that gets denser and more useful over time. The 50th source is meaningfully more valuable than the first.
A small set of narrow agents runs on a daily cycle. One processes my inbox and surfaces what actually needs me. One runs a morning briefing that synthesizes my health data, the day’s priorities, and what is moving in healthcare and AI. One does a weekly review across the projects I am tracking and flags what I have not touched. None of these agents tries to do everything. Each does one thing reliably, and they get better as the knowledge base they share gets richer.
It is, structurally, a tiny version of what Dorsey is describing for Block. A world model of one person’s work and life. An intelligence layer that composes answers from it. Interfaces I can query in plain language. Edges, where I do the actual judgment.
It is tinkering, not production. I break it constantly. The point is not the system. The point is that once you have built one, even badly, you stop confusing AI deployment with AI architecture. You feel where the data gaps are. You see how much of the value lives in the world model and how little of it lives in the chatbot on top. You understand why Dorsey calls his company an intelligence rather than a hierarchy, because you have built one for yourself.
It is hard to lead an organization through a structural shift you have not felt firsthand.
I think most senior healthcare leaders have not built one. Not even a small one.
I am also a professor at UCLA Anderson, where I teach the Business of Healthcare and get to watch the next generation of healthcare leaders make sense of all this. The MBAs who will run these companies in ten years are not going to compete on traditional management skills alone. They will compete on AI fluency layered on top of managerial judgment. The tools and frameworks I am working with personally show up in the projects we run together, because abstract slides about AI are not going to prepare anyone for what is coming.
Every quarter the AI conversation in the boardroom and classroom gets less abstract. The students who will be running these companies in ten years are not asking whether AI matters. They are asking which incumbents will still exist for them to work for.
That is the real question.
Close
Leadership in the age of AI is not about having the loudest AI strategy memo. It is not about giving everyone a chatbot, automating a few processes, and calling it transformation. It is about whether you can sit with the question Dorsey is sitting with, in public, and let the answer reshape your company before someone else’s answer does.
The org chart that ran the world for two thousand years is not going to run the next ten.
Omid Toloui, MBA, MPH, is principal at Red Creek Partners, a healthcare and AI strategy advisory practice, and adjunct professor at UCLA Anderson School of Management. He has spent more than two decades building and scaling AI, digital health, and care delivery models across payer, provider, and healthtech, most recently as VP of Innovation at Elevance Health, where he restructured the enterprise innovation operating model and scaled AI-powered consumer engagement and care delivery solutions to millions of members. His career spans management consulting, founding team roles in healthtech, and enterprise innovation leadership across health systems, health plans, and technology companies.
Recent Comments