Brief

Pharma's Primordial Soup: From Functions to Teams

For decades, expertise couldn't be separated from the expert. With AI, that constraint starts to lift.

Di Satty Chandrashekhar, Gopal Pai, Eric Garton, e Pete Hultman

First published on ottobre 21, 2026
Tempo di lettura min.


In evidenza Pharma organized into functions because specialist knowledge lived inside people; AI changes that. When context persists outside of individuals, the work stops having to travel to the expertise, and the team becomes the unit that matters. The companies with the most to gain from loosening functional boundaries are the ones with the most built on top of them. Boil a pharma company down to its components—the science, the regulatory know-how, the manufacturing expertise, the commercial muscle—and rebuild around the work itself. Imagine what emerges. Biology has a name for this: primordial soup. Every ingredient needed for life is present, organizing and reorganizing until something holds. Large pharmaceutical companies today are the opposite of that, by necessity. Functions, defined roles, clear boxes. Everyone fits together, and everyone stays put. Companies were built this way for a reason. The knowledge needed to move an asset forward has always lived inside specialists, so the work had to be routed to them. AI changes that premise. Functions loosen, and the team becomes the unit that matters. What follows is a question about the decade, not a plan for the next year. From functions to teams Consider simplistically how an asset moves today. Discovery hands off to preclinical, preclinical to development, development to regulatory, regulatory to commercial. Each handoff is a design choice rather than a flaw: The next stage needs expertise the last stage doesn't have, and that expertise sits in another department. The handoff is the price of specialization, and pharma has paid it willingly for decades because there were few alternatives in such a highly regulated and complex industry. But with the right use of AI tools, bold reimaginers can find a new way. When the context that governs a decision—such as the data, the prior calls, the reasoning behind them, or a shared sense of priority—persists in a system rather than in people's heads, the work no longer has to travel to the expertise. Instead, the work moves to a team: a standing group of people and AI agents that own an asset and stay with it. The roster changes as people rotate through, but the mission and the accumulated record of what was decided and why persist. A team that's taken an asset through a submission knows things nobody wrote down, and when the team moves forward, the knowledge compounds. Today, most companies are nowhere close to this reality. In Bain's survey of 120 clinical trial professionals, 49% of sponsors and contract research organizations said they're piloting AI tools without having restructured any roles. Just 16% are redesigning roles and workflows around what AI can actually do. The gap between adopting tools and redesigning the organization is the defining feature of where the industry sits. The ones that have redesigned report markedly better gains from AI than the ones running pilots inside the old structure. Scale as a constraint AI-native companies are running discovery as a closed loop, with no functional handoffs where information can be lost. Recursion proposes candidates, tests them in automated labs at scale, and retrains on every result. The building blocks of that model are no longer scarce. Talent moves, technology arrives through partnership or an application programming interface (API), and specialized knowledge is increasingly deployable without the specialist. And the boundaries between discovery, development, manufacturing, and commercialization are already blurring, well ahead of any org chart catching up. When the context that governs a decision persists in a system rather than in people's heads, the work no longer has to travel to the expertise. That cuts against the companies with the most to gain. We assess large pharma will struggle the most—the organizational model is the most entrenched, the teams are the most established, and the number of people whose authority comes from the current structure is the largest. A pre-commercial biotech with 200 people is already closer to an organizational model with looser departmental boundaries: individuals do many things, there are few teams, and nobody has to be talked out of a functional identity. The question for them is whether they can keep growing without growing organizational complexity. The shifts that redefine the organization In technology companies, the composition of teams is already changing, with the roles of product manager and engineer often merging day to day. Pharma’s boundaries were drawn harder and for longer. So, what re-forms when they loosen? Functions become homes for expertise, not owners of work. You'll still have structure. Regulatory affairs will still be where regulatory expertise lives, and it'll still train people and set the standard for how a filing gets done. What changes is what it controls. Today it owns the work, the budget, and the talent, and lends them out to whichever asset needs them. Loosen that, and functions vouch for the expertise while the team does the work. The expertise keeps a home; it just stops having a claim on the work. AI shifts modes across the asset life cycle. In discovery, the work rewards range: wide exploration, inexpensive failure, hypotheses few would have funded. Here, AI runs best with the guardrails loose. In clinical development and manufacturing, that inverts—tolerance for deviation is near zero, and AI earns its place by surfacing the exception rather than generating the option. In commercialization, the work is local and precise, and the value is in calibration to a specific payer, health system, and physician population. The team’s mission doesn’t change across any of it. What changes is how much latitude the AI has. In preclinical, where guardrails are loosest, pharma executives report the strongest returns from AI—3.4 times, compared with 2.3 times in clinical development and 2.6 times in commercialization. Human involvement becomes more nuanced, not a default. Nobody is taking humans out of asset development yet. But the level of human review is inherited from how the process was built, not calibrated to what’s truly at risk in each decision. That gets harder to justify as execution becomes more predictable in certain areas and judgment becomes the scarce skill. In some work, a person approves every action. In other cases, a person reviews exceptions and patterns rather than individual decisions. Elsewhere, the person does the work, and AI assists in real time. The choice among them becomes deliberate rather than inherited. Authority follows scope. A team behind the execution of an asset end to end has to be able to make high-quality, rapid decisions so it can continue to move with speed and avoid current organizational and governance drag. Today, decision authority sits with a senior leader overseeing the functional teams working on specific assets, because that senior leader is the only one with a view of resource deployment across assets. Give one team the broader mandate and authority for an asset, and the visibility shifts, and so does the decision. The hard part is cultural—pushing down decision authority—and it lands hardest on large organizations. Redesign becomes continuous. An operating model change usually comes with a target state and a date to hit it. That won’t hold here. What AI can do keeps moving, and so does what a team needs from it. A group that reorganized around one set of tools is working against a different set two years later. Getting good at reorganizing is the point, and that comes from doing it on real work: a discovery workflow, a submission, a launch. Questions to get started None of this arrives through one reorganization. The companies that try it will start somewhere small and specific, and scale it across the portfolio, likely refining the reorganization at least two or three times along the way. Discovery. Can a single hybrid team carry a target from hypothesis to first filing without handing it off, and what does that do to the timeline? Transitions. Can the context that usually gets lost at a handoff persist well enough that the handoff stops mattering? Launch. Can one team adapt a global strategy across multiple markets, rather than rebuilding it from scratch? Every pharma company is becoming a technology company The easy response to all of this is that it describes a tech business and is impractical in the pharma context—different science, different regulators, and different consequences for being wrong. The harder thing to sit with is that the distinction between a technology company and a biotechnology company is quickly eroding on its own. The ingredients are already in the water. What forms next depends on how much of the company was built to keep them apart.

In evidenza

Pharma organized into functions because specialist knowledge lived inside people; AI changes that.
When context persists outside of individuals, the work stops having to travel to the expertise, and the team becomes the unit that matters.
The companies with the most to gain from loosening functional boundaries are the ones with the most built on top of them.

Boil a pharma company down to its components—the science, the regulatory know-how, the manufacturing expertise, the commercial muscle—and rebuild around the work itself. Imagine what emerges. Biology has a name for this: primordial soup. Every ingredient needed for life is present, organizing and reorganizing until something holds.

Large pharmaceutical companies today are the opposite of that, by necessity. Functions, defined roles, clear boxes. Everyone fits together, and everyone stays put. Companies were built this way for a reason. The knowledge needed to move an asset forward has always lived inside specialists, so the work had to be routed to them.

AI changes that premise. Functions loosen, and the team becomes the unit that matters. What follows is a question about the decade, not a plan for the next year.

From functions to teams

Consider simplistically how an asset moves today. Discovery hands off to preclinical, preclinical to development, development to regulatory, regulatory to commercial. Each handoff is a design choice rather than a flaw: The next stage needs expertise the last stage doesn't have, and that expertise sits in another department. The handoff is the price of specialization, and pharma has paid it willingly for decades because there were few alternatives in such a highly regulated and complex industry.

But with the right use of AI tools, bold reimaginers can find a new way. When the context that governs a decision—such as the data, the prior calls, the reasoning behind them, or a shared sense of priority—persists in a system rather than in people's heads, the work no longer has to travel to the expertise.

Instead, the work moves to a team: a standing group of people and AI agents that own an asset and stay with it. The roster changes as people rotate through, but the mission and the accumulated record of what was decided and why persist. A team that's taken an asset through a submission knows things nobody wrote down, and when the team moves forward, the knowledge compounds.

Today, most companies are nowhere close to this reality. In Bain's survey of 120 clinical trial professionals, 49% of sponsors and contract research organizations said they're piloting AI tools without having restructured any roles. Just 16% are redesigning roles and workflows around what AI can actually do. The gap between adopting tools and redesigning the organization is the defining feature of where the industry sits. The ones that have redesigned report markedly better gains from AI than the ones running pilots inside the old structure.

Scale as a constraint

AI-native companies are running discovery as a closed loop, with no functional handoffs where information can be lost. Recursion proposes candidates, tests them in automated labs at scale, and retrains on every result. The building blocks of that model are no longer scarce. Talent moves, technology arrives through partnership or an application programming interface (API), and specialized knowledge is increasingly deployable without the specialist. And the boundaries between discovery, development, manufacturing, and commercialization are already blurring, well ahead of any org chart catching up.

When the context that governs a decision persists in a system rather than in people's heads, the work no longer has to travel to the expertise.

That cuts against the companies with the most to gain. We assess large pharma will struggle the most—the organizational model is the most entrenched, the teams are the most established, and the number of people whose authority comes from the current structure is the largest. A pre-commercial biotech with 200 people is already closer to an organizational model with looser departmental boundaries: individuals do many things, there are few teams, and nobody has to be talked out of a functional identity. The question for them is whether they can keep growing without growing organizational complexity.

The shifts that redefine the organization

In technology companies, the composition of teams is already changing, with the roles of product manager and engineer often merging day to day. Pharma’s boundaries were drawn harder and for longer. So, what re-forms when they loosen?

Functions become homes for expertise, not owners of work. You'll still have structure. Regulatory affairs will still be where regulatory expertise lives, and it'll still train people and set the standard for how a filing gets done. What changes is what it controls. Today it owns the work, the budget, and the talent, and lends them out to whichever asset needs them. Loosen that, and functions vouch for the expertise while the team does the work. The expertise keeps a home; it just stops having a claim on the work.

AI shifts modes across the asset life cycle. In discovery, the work rewards range: wide exploration, inexpensive failure, hypotheses few would have funded. Here, AI runs best with the guardrails loose. In clinical development and manufacturing, that inverts—tolerance for deviation is near zero, and AI earns its place by surfacing the exception rather than generating the option. In commercialization, the work is local and precise, and the value is in calibration to a specific payer, health system, and physician population. The team’s mission doesn’t change across any of it. What changes is how much latitude the AI has. In preclinical, where guardrails are loosest, pharma executives report the strongest returns from AI—3.4 times, compared with 2.3 times in clinical development and 2.6 times in commercialization.

Human involvement becomes more nuanced, not a default. Nobody is taking humans out of asset development yet. But the level of human review is inherited from how the process was built, not calibrated to what’s truly at risk in each decision. That gets harder to justify as execution becomes more predictable in certain areas and judgment becomes the scarce skill. In some work, a person approves every action. In other cases, a person reviews exceptions and patterns rather than individual decisions. Elsewhere, the person does the work, and AI assists in real time. The choice among them becomes deliberate rather than inherited.

Authority follows scope. A team behind the execution of an asset end to end has to be able to make high-quality, rapid decisions so it can continue to move with speed and avoid current organizational and governance drag. Today, decision authority sits with a senior leader overseeing the functional teams working on specific assets, because that senior leader is the only one with a view of resource deployment across assets. Give one team the broader mandate and authority for an asset, and the visibility shifts, and so does the decision. The hard part is cultural—pushing down decision authority—and it lands hardest on large organizations.

Redesign becomes continuous. An operating model change usually comes with a target state and a date to hit it. That won’t hold here. What AI can do keeps moving, and so does what a team needs from it. A group that reorganized around one set of tools is working against a different set two years later. Getting good at reorganizing is the point, and that comes from doing it on real work: a discovery workflow, a submission, a launch.

Questions to get started

None of this arrives through one reorganization. The companies that try it will start somewhere small and specific, and scale it across the portfolio, likely refining the reorganization at least two or three times along the way.

Discovery. Can a single hybrid team carry a target from hypothesis to first filing without handing it off, and what does that do to the timeline?
Transitions. Can the context that usually gets lost at a handoff persist well enough that the handoff stops mattering?
Launch. Can one team adapt a global strategy across multiple markets, rather than rebuilding it from scratch?

Every pharma company is becoming a technology company

The easy response to all of this is that it describes a tech business and is impractical in the pharma context—different science, different regulators, and different consequences for being wrong. The harder thing to sit with is that the distinction between a technology company and a biotechnology company is quickly eroding on its own. The ingredients are already in the water. What forms next depends on how much of the company was built to keep them apart.