Do AI-Native Biotechs Need Departments? Benchmarking Company World Models for AI-Driven Drug Development

Imagine a world where building a new medicine isn't just about mixing chemicals in a lab, but about making thousands of high-stakes decisions under pressure. It's like running a massive, complex video game where you have to choose which characters to upgrade, which levels to skip, and when to sell your inventory, all while the rules keep changing. For decades, human companies have solved this by splitting the work into "departments"—a team for biology, a team for rules, a team for sales, and so on. They pass notes back and forth, hoping the final boss (the FDA or the market) is happy. But what if we could build a company run by AI? Should we just copy-paste those human departments into robot roles, or is there a smarter way to organize a digital brain?

This is the big question tackled in a new study called "Do AI-Native Biotechs Need Departments?" The researchers are exploring a concept called a "World Model." Think of a World Model like a super-powered crystal ball that doesn't just guess the future, but constantly updates a live map of "what is happening right now" and "what will happen if I do X." Instead of just looking at a single piece of the puzzle, a World Model keeps a running score of how every action changes the company's chances of success, money, and getting approved. The study asks: If we build an AI company, should we force it to act like a human corporation with different departments, or should we let it run on a single, shared "live map" that everyone updates in real-time?

The Experiment: A Time-Traveling Decision Game

To find the answer, the researchers didn't build a real drug company or test new medicines in a lab. Instead, they created a "dry-lab" benchmark, which is like a massive, 45-level video game level designed to test decision-making. They took 45 real historical moments from the drug industry—times when a company had to decide whether to keep funding a drug, stop it, or sell it. They gave these scenarios to different AI teams, but with a catch: the AI teams could only see information available up to that specific date in the past. They couldn't peek at the future to see if the drug actually worked or made money.

The researchers tested four different "team structures" to see which one made the best decisions:

1. The Human Copy: AI agents acting exactly like human departments (Biology, Clinical, Sales, etc.), passing memos to a committee.
2. The Human Copy Plus: A smarter version of the above, where the departments share a better memory and argue more effectively.
3. The Asset-Centric AI: A team organized around a single, live "asset record" (a digital file tracking the drug's progress) rather than departments.
4. The "Company World Model" (Value-Conversion): The star of the show. This team didn't have departments. Instead, it had a shared "Live Asset Value Record" and four special "rooms" that acted like operators: a Deal Room (checking if partners want to buy), an Approval Room (checking if regulators will say yes), a Revenue Room (checking if people will buy it), and an Investment Arbiter (the boss who decides the next move based on the live map).

The Results: It Depends on What You're Trying to Win

The results were surprising and nuanced. When the researchers judged the AI teams based on how well they could convert a drug into a business success (getting a deal, approval, or sales), the Company World Model architecture was the clear winner. It scored a 4.68 out of 5, beating the "Human Copy" (4.33) and the "Asset-Centric" version (4.26). When blind judges (who didn't know which team was which) looked at the decisions, they strongly preferred the World Model team, with one judge picking it 42 times out of 45 against the original human copy.

However, the story gets more interesting when you change the rules. The researchers realized they might have been cheating by asking the AI to play a game specifically designed for the "Value-Conversion" team. So, they added a "neutral judge" that didn't care about money or deals, but only about general decision quality, risk awareness, and logic. Under this neutral test, the World Model team did not dominate. The "Human Copy Plus" team actually did slightly better (25 wins vs. 20), and the difference wasn't statistically huge.

This tells us something crucial: The "Company World Model" isn't a magic bullet that wins at everything. It is a specialized tool. If your goal is to maximize business value (deals, approvals, revenue), organizing around a shared, predictive state of "asset-to-value" is superior. But if your goal is just to be a good, neutral scientist, a traditional department structure (even a smart one) works just as well.

The "Rooms" vs. The Departments

One of the coolest findings came from breaking the winning team apart. The researchers removed one "room" at a time to see what happened.

• When they removed the Revenue Room, the team's ability to think about money and sales dropped significantly.
• When they removed the Deal Room, their ability to find partners crashed.
• When they removed the Approval Room, their regulatory thinking got worse.

This proved that the "rooms" weren't just fancy names for departments. They were active "operators" that constantly updated the shared map. The study suggests that in an AI-native company, you shouldn't just copy human departments. Instead, you should have a shared, living state (the Live Asset Value Record) that everyone updates, and specific "operators" that simulate different futures (Will this get approved? Will this sell?).

The Bottom Line

So, do AI-native biotechs need departments? The paper suggests the answer is "not as the core brain."

Departments are fine for human governance and for organizing human meetings, but they are a bad blueprint for an AI's internal logic. The best design seems to be a Company World Model: a system where a single, auditable "live map" of the drug's value is constantly updated by different operators (Deal, Approval, Revenue, etc.), and a planner decides the next move based on that map.

The study is careful to note that this is a simulation. It proves that this structure makes better decisions in a computer game about the past. It does not prove that an AI using this method will actually discover a new cure or make billions of dollars in the real world. But it does suggest that if we want AI to run a drug company, we shouldn't just build a robot version of a 1990s corporate org chart. We should build a system that constantly simulates the future of value, rather than just passing memos between departments.