Pharma's Biggest Cost Isn't Research

Jack Dorsey co-founded Twitter and built Block. A few months ago, he and Sequoia's Roelof Botha argued that the shape of the modern company is obsolete. Unless you live in the Valley, you probably aren't thinking much about it. That's the problem I want to write about.

The AI conversation runs on the loudest signal. One week it's a warning that the technology is too dangerous and must be slowed. The next it's a tool that does what we already did, faster, better, cheaper. The news decides what we discuss, and the news rewards alarm and novelty. What it skips is the quiet, structural idea that changes how the work gets done. Dorsey's is one of those, which is why it's worth going back for.

I want to look at it through the lens of the innovative medicines industry, because that's where his argument has the most value to unlock. As I'll argue, his essay, "From Hierarchy to Intelligence" isn't a blueprint for pharma. It's a telescope, showing the industry where to look for its next big leap.

Dorsey asks organizations to rethink what AI is for. He's forecasting a revolution in organizational design, the first real one in two thousand years. So the question worth asking is what that means for pharma, one of the most matrixed industries on earth.

Dorsey's argument reduces to three ideas.

First, to date, AI has mostly revolutionized productivity. A copilot sitting beside workers and speeding up tasks. The gains are real, but they're incremental: the same work, done faster.

Second, organization and information structures haven't changed. Information still travels from the top of an organization to the bottom through matrices of eight to ten people. That system is human-sized and roughly unchanged since Ancient Rome. Dorsey's alternative is to let AI hold the company's context, rather than relying on people to carry it. Everyone gets what they need to know directly, without waiting for it to pass down through layers of managers.

Third, that changes AI's role. It stops being a tool people use and becomes part of how the company runs. It does the job middle managers have done since the Roman army, carrying context between the people who decide and the people who act. So the next advantage won't go to the company with the best tools. It'll go to the company that redesigns itself around them.

Block is trying to build it. In February 2026, Block cut roughly 4,000 employees, about 40% of its total workforce. The essay describes what replaces them. Block runs on two models. One is a company world model: how the company understands itself, continuously built from code, decisions, workflows and performance metrics. The other is a customer world model: a picture of each customer and merchant, built from Block's own transaction data. An AI layer sits on top of both and does the coordinating: tracking projects, identifying issues, assigning work and sharing critical information faster than humans can. The people who remain do one of three jobs: individual contributors, who do the work in a defined area; directly responsible individuals, who own a single cross-company problem for a short sprint; and leaders who still do the work themselves while coaching others.

The appeal is easy to see. Picture a company that knows itself in real time. Anyone, from an analyst to a board member, can ask it what's happening and why, and get an answer without waiting for the next meeting. On Sequoia's podcast, Dorsey went further. In the ideal version, he said, there are no layers at all: all 6,000 people report to him directly, because most of the work runs through the intelligence layer. He admits that sounds ridiculous under the old structure.

Pharma can't build that, and the reasons matter.

The cost of being wrong is different. Block says it's early, and early means mistakes. When Block gets it wrong, a merchant doesn't get a loan. When a pharma company gets it wrong, the cost can be a patient's life or a regulatory catastrophe. An industry pledged to preserve life doesn't get to break things to build them.

Pharma's feedback loop is different. Block's AI learns from its own transactions. It makes a call, sees the result almost immediately, and corrects. That loop is what makes it fast and trustworthy enough to stand in for managers. In healthcare, no such loop exists. The outcome that matters, whether a patient got better, happens outside the company, scattered across a siloed healthcare system and walled off by privacy law. A model that can't see the full picture can't be trusted in the same way.

Some of pharma's hierarchy is there on purpose. Medical, safety, legal and regulatory review. The named person who signs off. Regulators require a human to be accountable. In pharma, some of those layers aren't overhead. They're the control.

So the lesson for pharma isn't in the machine Dorsey built. It's in the problem he named: the cost of getting what one part of the company knows to every other part that needs it.

Once you start looking at that cost, a lot of what pharma does to save money looks different. Most of the industry's recent cost moves aim at the cost of people: cheaper hours, fewer heads, leaner teams. That's a reasonable thing to aim at. But it means the cost Dorsey is talking about never shows up in the business case. And many of the moves that cut the cost of people quietly raise the cost of what happens between them.

Offshoring is the clearest example. Move a function to a low-cost center and every hour of work gets cheaper. But now the work crosses time zones, handoffs and gaps in context, so you spend to close them: new processes, new coordinating roles, months of transformation. The task got cheaper. Getting it done got more expensive. And every dollar of that transformation goes into making the matrix run better.

None of the moves in the industry's standard playbook are bad decisions. That's the trap. Each one is easy to justify and easy to count. Together, they mean pharma is spending real money perfecting a structure Dorsey says AI has the potential to replace.

That doesn't mean pharma is behind. Outside the AI-native companies, it may be the most aggressive adopter of AI anywhere, using it from discovery to omnichannel targeting to creative development. Some of it is remarkable. Anthropic's Claude Science, launched this summer, takes a research goal and does much of the computational work on its own. At launch, it was shown finding drug candidates for a rare genetic disease.

But however far these tools go, they do the same kind of thing: the work, done faster and with less help. That's the copilot, Dorsey's first idea. Vendors today sell it as revolution. It's mostly acceleration. None of these tools changes the shape of the organization. The work gets faster, but what it produces still travels through people, layer by layer, the way orders moved through a Roman legion.

The most expensive thing pharma does isn't research. The industry spends billions creating knowledge. It spends far more making that knowledge usable across the organization. And it's the one cost nobody is selling a tool for, because it's too big to be a product.

Dorsey answered the question for Block. Pharma hasn't asked it yet: what does a pharma company look like built on intelligence instead of hierarchy?