Building an AI-Native Pharma

Sublinear Scaling and System Design

We tell people all the time that we’re building Formation Bio as an AI-native pharma company. However, the phrase is easily misunderstood. People assume it means we’re unusually good at building AI tools (we are) or that we use LLMs everywhere (we often do). But it’s important to note those are means, not the end. An AI-native company is not one that augments individual experts with clever tools, it is one that re-architects the organization so that entire workflows, spanning many roles and handoffs, are executed by systems.

The goal is to re-architect the enterprise so that growth scales with systems, not labor.

On scaling, simple patterns and irreducible complexity

Scaling is the defining question for any company. When I began my career in Silicon Valley, friends outside the Bay Area would ask, “what is a startup anyway?” There are plenty of joke answers (a group project where the deadline is “before we run out of money”), but the real answer is deceptively simple: a startup is a company designed to scale. A small business grows by adding people and resources roughly in proportion to revenue. A startup bets that technology can break that proportionality.

Technology has always been a multiplier of human effort. With machines, one person could do the work of many. With software, this amplification was dramatically enhanced: a handful of engineers can build a search engine to index the entire web. Beyond that, the move from hardware to software unlocked an explosion in scale by shifting work from the physical world to the intangible, where that scaling and replication can be near-free.

But traditional software has a ceiling. Software fundamentally encodes rules, and rules break at the edge cases. LLMs, however, are the first broadly useful technology that lets us deploy flexible behavior at scale - systems that can generalize beyond what we enumerated in advance. Put differently:

• Software scales reducible complexity (patterns we can codify).

• AI scales irreducible complexity (contexts we can’t exhaustively pre-specify).

That is why there are predictions of the first billion-dollar company with a single employee in the next five years. When we think about how work can get done, AI changes the efficient unit of company execution from a person-plus-tool to a scalable system-with-guardrails.

At Formation, our goal is to leverage this, amplifying expert impact to achieve sublinear scaling: growing our drug asset base faster than our operational footprint. However, we have to solve a problem first: why doesn't augmenting each expert with AI tools achieve sublinear scaling?

Conway's Law and the Pharma Problem

There is an observational adage known as Conway's Law: an organization's products will mirror its own communication structure. Have an iOS team and an Android team? You will ship two different user experiences. Have a frontend team and a backend team? You will create friction at the interface where they meet.

This law can also be exploited. If you are, say, Amazon building a cloud capability ahead of your well-capitalized peers, you want to build a fast-moving collection of disparate capabilities. Conway's Law dictates an org structure of many small, autonomous teams to create that result - which is exactly what AWS created with the "two-pizza team.”

This is the Inverse Conway Maneuver: rather than letting your org chart dictate your product, you intentionally design your org chart to design your product.

Let's apply this framework to pharma. Consider a typical workflow for patient recruitment in a clinical trial (see below). The current process, a direct product of Conway's Law, is fragmented across many specialized roles and handoffs. Each step is locally optimized by an expert, but the integration overhead is severe: coordination meetings, OOO delays, context loss, and interface friction. The product is a patchwork of expert contributions with a high coordination cost.

If we want this system to scale better, Conway's Law states that simply demanding people to "work more efficiently" will fail. An optimized holistic solution must be structural. We must deploy an Inverse Conway Maneuver to engineer fewer handoff points. That is, to get a unified product that scales with systems rather than labor, we must first pull together work into a unified step.

And here, AI can be a solution. The question is not “where can a person use an LLM?” but “where do we genuinely need human judgment, and how do we automate everything between those checkpoints?” Many roles in pharma were codified for a different technological era. We should not knock down Chesterton’s Fence - the old rule that before you tear something down, you'd better understand why it was put up. Some fences are there for safety, ethics, or compliance. But a surprising number persist mainly because the org chart required them.

If we take Conway’s Law seriously, then incremental improvement inside the existing org chart is a dead end. Layering AI tools onto every role in a fragmented workflow preserves fragmentation, and the interfaces, handoffs, and coordination costs remain because the underlying structure is unchanged.

The implicit assumptions baked into decades of pharma org design dictate certain constraints – that work must be decomposed into narrow roles, that accountability must be diffused across functions, and that scale requires adding people, but AI changes those constraints. What once required many specialized humans coordinating on a process can now be executed by systems. Rather than asking how AI fits into the old model, we need to discard the old model and design a new one around what AI makes possible. Here's an example of what that looks like in practice.

With Muse, our internal patient recruitment platform, we collapse a nine-step org chart into a single accountable owner operating a constellation of agents. We are simplifying patient recruitment from the entire organization shown on the left to the model on the right: a streamlined team of "super-users" who can operate every facet of recruitment. The human’s job shifts from producing artifacts to setting objectives, validating edge cases, and granting approvals. The “team” becomes a system with a human in the loop.
Our goal is not to augment each person in the old workflow with an LLM. It is to reimagine the whole workflow so that the original patient recruiting team, its managers, and its contractors are never built at all.

In essence:

• We scale LLMs to do the work of whole groups of people (scaling irreducible complexity).

• We apply the right guardrails, leveraging existing tools and standard software (managing reducible complexity).

• The entire system is driven by experts (ensuring quality).

It is system design that yields sublinear scaling, dependent on defining where we need human review. The challenge is not identifying which steps can be aided by an LLM. it's determining the critical points where we need expert human feedback, and automating everything in between.

The Goal: Operational Alpha

This solution - one person, many systems - appears again and again in modern companies. The same principle applies to information flow. How many roles exist in a typical company to collate information, synthesize status, arrange slides, and deliver presentations? Is that overhead necessary in a world where all deliverables are digitally created, and an agent can prepare a report on demand?

And let’s be clear: most companies are not well-incentivized to holistically optimize their organization to take maximal advantage of AI systems. Large standing teams, entrenched experts, and existing optimizations all create inertia. Vendors struggle to sell products that require reorgs, and buyers struggle to get internal buy-in for solutions with a long, complex ROI. Instead, vendors focus on easy-to-sell point solutions that ramp existing workflows quickly, but will ultimately fall far short in impact. If you want this level of efficiency, you must build it from the ground up or endure a painful transformation.

In this new model, every employee has breathtaking scope and substantial leverage. Each human expert becomes invaluable as the arbiter of quality.

We seek operational alpha: not just incremental improvement, but a fundamental transformation of how an organization scales. Success means doubling our asset portfolio while only modestly increasing headcount, ultimately driving the marginal operational cost per drug asset toward zero.

And for everyone reading this, that is a good thing because it means medicines will be developed faster and more efficiently.

TL;DR: AI-native pharma means aiming for sublinear scaling and workflow redesign which removes the overhead of teams.

If you want to be an extraordinary point of leverage for a company building the future to improve human health, reach out.

Your Corporate Memory Should Appreciate, Not Depreciate

In Part 1, we discussed how sublinear scaling and system design are twin efforts that can help a company realize benefits by reengineering how companies scale. In Part 2, we look at an even more fundamental change: rethinking what an organization is by how it captures and processes information.

Data Archaeology Tax vs. The Instrumentation Mindset

Picture a familiar Monday: you return from two weeks away to 2,700 Slack messages, 54 document comments, and three conflicting email threads about a decision that happened while you were out. Your morning evaporates reconstructing what happened while you were gone.

This is the data archaeology tax: the hidden overhead organizations pay when they treat information as single-use, letting it sink into drives, chats, and inboxes. The cost is staggering: knowledge workers spend 20-30% of their time excavating information. That's more than one day per week spent on institutional archaeology.

An AI-native company, however, can pay a smaller up-front cost to skip the tax: instrumentation overhead. Rather than spending 20% of our effort excavating context after the fact, I want my teams to spend 5% capturing it prospectively at the source.

Consider how engineers build production systems. They instrument deliberately: requests are logged with sufficient detail to reconstruct what happened; errors route to the right teams with diagnostic context; information flows are traced end-to-end. This instrumentation is not free and it requires discipline and a modest overhead to implement and maintain. But the payoff is overwhelming - because the information must be captured in order to be used.

Consider the return: information that doesn't depreciate.

The Compounding Asset vs. The Wasting Asset

In a traditional organization, knowledge is a wasting asset. The moment it is created, it begins to lose value. Six months later, the document is hard to find. A year later, the context is lost. Two years later, the last person who knew about it moves on, and the knowledge is gone. Amnesia as a business model is just generally accepted as normal and inevitable.

In an instrumented organization, knowledge can be a compounding asset. Each piece of context captured becomes more valuable over time because it can be connected to new contexts, reused for new purposes, and synthesized in ways that weren't anticipated when it was created. Your corporate memory can actually appreciate in value.

This distinction has profound implications. The returns to instrumentation are nonlinear and back-loaded. The first month of capturing structured meeting notes feels like pure overhead. By month twelve, you have a searchable, interconnected repository of decisions. By month twenty-four, patterns emerge that no individual could perceive. By year three, your institutional memory recalls better than your longest-tenured employees. The system gets smarter every single day.

Recomposability: From One-Time Use to Infinite Applications

Why instrument so extensively? Recomposability is the payoff - and where genuinely new capabilities emerge.

In traditional companies, most information has a single use case. Meeting notes are shared, briefly reviewed, then forgotten. A market analysis is read by the strategy team, then never touched again. This isn't because people are lazy or disorganized - it's because the archaeology tax makes reuse prohibitively expensive. By the time someone needs that analysis again, recreating it from scratch is often faster than finding and contextualizing the original.

The math here is brutal. If information is expensive to retrieve (high archaeology tax), it's only worthwhile if the use case is high-value and imminent. Most information fails this test, and most people anticipate this, so the content gets created and stored as informally as possible and retrieved never.

However, if information is nearly free to retrieve and recompose, then the start-up cost for reusing prior information drops precipitously. A single piece of well-instrumented context can serve dozens of applications, many unconceived at the time of capture.

At Formation, we're beginning to see this realized. Meeting notes from a business development diligence session might be consumed initially by the deal team. But because they're captured in structured, machine-readable format, they simultaneously:

• Feed into our pattern recognition framework (Atlas) to identify recurring signals about promising therapies

• Provide real-world ground-truth corroboration for datasets

• Update therapeutic area-wide risk-adjusted NPV models with new cost or timeline assumptions

• Flag causal biology insights relevant to other portfolio assets

• Answer a new analyst's question 18 months later about why we passed on a similar opportunity

Consider a more intricate example. We recently purchased an expensive proprietary dataset - MRI images and clinical outcomes for thousands of patients with knee osteoarthritis - to de-risk a therapeutic drug we own. We built an AI model predicting disease progression from MRIs, demonstrating our drug candidate could meaningfully reduce total knee replacement risk: a clinically significant and economically valuable endpoint.

That's valuable. But here's where recomposability creates leverage: we realized we could use the same dataset to de-risk an entirely different asset with a different mechanism of action (MoA) targeting a similar patient population. Because an approved drug existed with a similar MoA, we filtered the dataset to patients taking this medication and examined in silico whether these drugs, via a different MoA, meaningfully impacted disease progression.

A data scientist completed this analysis in roughly one week during acquisition diligence. At a traditional pharmaceutical company, this would require months - assuming it happened at all. More likely, the relevant people wouldn't even know the dataset existed and information silos would prevent the connection from being made.

This is recomposability in action: expensive assets paying off their return on investment with each reuse, knowledge compounding rather than dissipating, and insights connecting across domains in ways that create genuine competitive advantage.

And notice the connection to Part 1: this kind of recomposability is only possible in organizations designed for sublinear scaling. In a company of 50,000 people, spanning dozens of divisions and geographies, the coordination costs would overwhelm any potential benefit. But in a deliberately small, cross-functional organization where the data scientist working on osteoarthritis can easily talk to the team evaluating a new acquisition, these connections happen naturally - if the information is instrumented properly.

Remember Conway’s law from the first part? Your org structure shapes your products. But people rarely consider why: it’s because information architecture shapes your decisions. If your information is siloed in department wikis and individual heads, then your decisions will be siloed too. Recomposable context enables recomposable thinking.

The Organizational Exocortex

Taken to its limit, this creates a fundamentally different kind of organization. When organizational data doesn't evaporate but instead artfully accumulates, you invite a new form of intelligence to your company.

Steven Johnson, author of You Exist In the Long Context and other works on innovation, articulates this vision eloquently:

In a matter of years, I suspect it will seem bizarre to draft the specs for a new feature or a company initiative or a grant proposal without asking for feedback from a long-context model grounded in the organization’s history. (And perhaps the public history of its competitors.) It wouldn’t be a replacement for the expertise of the employees; instead, the model would occupy another seat at the table, adding a new kind of intelligence to the conversation, along with a vastly superior recall.

Imagine an AI system with instant recall of every significant decision, experiment, analysis, and lesson - not as static files, but as an interconnected web of context. It synthesizes, identifies cross-domain patterns, and flags connections no single employee could track. What does this enable?

• Historical pattern recognition: Propose pursuing Indication X, and the system surfaces that you assessed it in 2023, summarizes the rationale for deferring (insufficient patient stratification biomarkers, prohibitive trial costs, competitor had a five-year head start), and flags what's changed since (new biomarker discovered, regulatory path clarified, competitor trial failed).

• Cross-domain synthesis: A legal teammate asks about liability exposure for a trial design. The system doesn't just fetch precedents; it links contract clauses to R&D feasibility constraints, adverse event patterns from past trials, and finance models your company has developed - highlighting risks visible only at the intersections.

• Constructive devil's advocate: Before committing to a major bet, you ask the model to argue against it using your own historical reasoning. It finds analogs, tallies successes versus failures, and isolates the relevant decision points with perfect recall and no ego.

This differs fundamentally from document management or even sophisticated search tools. It is a reasoning system grounded in your organization's lived experience, one that improves as it continually ingests more context from ongoing work.

In a few years, it will seem profoundly myopic for companies to forfeit these advantages because enabling auto-note-taking seemed too difficult.

Network effects and compounding

The exocortex represents a new kind of competitive advantage, deeply tied to organizational network effects. We're familiar with product network effects where each additional user makes the platform more valuable for all others. But network effects operate at the organizational level too: each additional instrumented decision makes the entire system more valuable.

This creates genuinely unusual compounding returns. Most competitive advantages are either one-time (patents, regulatory moats) or are fragile and require continuous investment to maintain (brand, customer relationships). But instrumented institutional knowledge compounds effortlessly. Every documented decision, every captured meeting, every structured analysis makes the system more cumulatively capable.

Consider the trajectory:

• Year 1: You instrument your workflows. This feels like overhead. The system has limited data, so its suggestions are modest. ROI is negative or barely positive.

• Year 2: The system has meaningful historical depth. It starts catching duplicated work, identifying risks earlier, and making connections across domains. ROI turns positive but still modest.

• Year 3: The system has seen enough patterns to make non-obvious predictions. It's preventing mistakes before they happen, identifying opportunities that would otherwise be missed, and making every new employee productive faster than human onboarding alone could achieve. ROI accelerates.

• Year 5: The system knows your organization better than any individual. It identifies patterns no human could perceive, connects dots across years of work, provides analysis impossible without this historical depth. Competitors see the value and want to catch up - but they're starting from zero, and the gap is widening.

Here's the crucial compounding with Part 1: a small, focused team with deep institutional memory can outmaneuver a much larger organization perpetually reconstructing context. This can be a critical advantage for companies navigating times of change.

The Future of Work: From Wasting Knowledge to Compounding Memory

The natural objection: "Won't the change management be painful?"

Yes - and that is part of the benefit. This isn't a feature you can copy in a quarter. It's a capability that compounds over time, which means the winners are decided by when you start, not how fast you move once you realize you're behind.

Consider the asymmetry: an organization that began instrumenting three years ago has three years of accumulated context, pattern recognition, and institutional memory that a latecomer simply cannot purchase or replicate. The gap doesn't close with effort - it widens with time. Every day of delay is a day of compounding advantage surrendered to whoever started first.

This brings us back to Part 1: sublinear work scaling only becomes possible with sublinear context scaling. You can collapse workflows and eliminate handoffs, but if every decision still requires excavating lost context, you've just shifted the bottleneck to retrieval or additional manual input. The two transformations are inseparable.

In fact, all four elements - sublinear scaling, system redesign, instrumentation, and recomposability - form a single reinforcing loop. Instrumentation captures the context that makes recomposability possible. Recomposability makes your systems smarter, enabling bolder system redesign.

System redesign collapses handoffs and consolidates workflows, delivering sublinear scaling. And sublinear scaling keeps your organization small enough that instrumentation remains feasible and coordination costs stay low. Each turn of the cycle accelerates the next. Miss one element and the others stall; get all four spinning and you have a flywheel that compounds.

This is what "AI-native" truly means. Not augmenting individual experts with clever AI tools, but building an organization where every employee is amplified by the total, recomposable memory of an entire enterprise. A company that, by design, gets smarter every single day.

That's what we're building. That's the future of work.