Key takeaways

• AI’s generational promise in pharma R&D depends on concentrated bets across data, workflows and roles, not just applications alone.
• AI transformation in pharma R&D requires a unique approach, with a heavily regulated and deeply specialized value chain demanding a purpose-built approach.
• Pharma R&D leaders should direct about 80% of AI investment toward data and workflow transformation, rather than to applications themselves.
• AI creates enterprise value when organizations redesign entire value streams, not simply accelerate individual tasks.

The hope for AI in pharmaceutical R&D is enormous. After two decades of flat-to-declining research productivity, AI is widely seen as the technology that could finally reverse the trend. With many sponsors looking to increase their output of new molecular entities without increasing R&D spending, there’s an expectation that AI will make that ambition attainable. Pharma executives routinely speak about their enthusiasm for AI in quarterly earnings calls, and Pfizer’s Albert Bourla recently published his own R&D-centric position piece on the topic.

The AI industry is reciprocating the sentiment. In the span of a month in April 2026, Anthropic acquired Coefficient Bio and appointed Novartis CEO Vas Narasimhan to its board. This made Narasimhan the first pharmaceutical executive to join a major AI lab’s governing body, and his appointment is a marker of how quickly the two industries are converging.

Why aren't we seeing more value?

Everyone in pharma agrees the opportunity is huge. Even generational. What’s in dispute is why so little of that value is showing up in earnings or pipeline advancements, despite pharma’s well-publicized efforts. Other industries, including much of healthcare, have already seen AI reshape valuations and business models. And yet, pharma has been strikingly insulated:

• Pure-play, AI-native life sciences companies command no valuation premium over their traditional counterparts. In several cases they trade at a discount, even as their pipelines advance.
• AI-driven productivity is essentially unpriced in “traditional” commercial stage pharma equities.
• Surveys of R&D and life sciences executives point, almost uniformly, to a lack of realized value from AI to date.

Data, workflows and roles matter more than the application

So with all the promise, what explains the gap? Our diagnosis is drawn from direct experience helping R&D organizations move AI from pilots to enterprise scale, and we lay out a path forward for successful AI transformations in R&D.

Our central finding is relatively simple: The problem is almost never the AI itself, but where organizations spend their effort and their budgets. In most R&D organizations today, roughly 80% of executive attention and investment flows to the visible, exciting parts: applications, co-pilots and use cases.

But in the transformations that actually create enterprise value, the allocation is close to the opposite, with organizations dedicating most of their focus on elements other than the applications and technology themselves. Let’s set the stage for what makes AI transformation uniquely challenging in pharma R&D.

Why scaling AI is a unique challenge in pharma R&D

Most advice about scaling AI treats it as a technology challenge. You have to pick and validate the right applications, stand up the right platform, roll out co-pilots and finally manage the change. That framing underserves pharma’s value chain, which has five structural features that make R&D a uniquely difficult—and uniquely promising—environment for AI.

1. A very long, and sometimes nonlinear, value conversion cycle. From target identification to launch, a single asset moves through a decade-plus of loosely coupled stages, functions and systems. It may fail in one disease area but succeed in another. Value rarely comes from speeding up one individual step, and no single tool can compress the entire chain.
2. Regulated and deliberately inert processes. Good clinical practice (GxP) frameworks create switching costs in the form of testing and validation to prove consistency and traceability. As a result, clinical trial machinery hasn’t significantly evolved in recent decades. Process change is possible, but it’s slow, expensive and hard-won.
3. The nature and history of the data. Structural impediments prevent R&D teams from creating rapid learning loops with their data. These include patient consents that constrain reuse; fragmented data and systems patched together through years of mergers and acquisitions and tech migrations; inconsistent and pre-standard formats; and in some cases, physical paper forms. The raw material AI needs is not clean or freely usable.
4. Deep specialization of labor. The expertise of R&D professionals is often narrow and deep, cultivated through well-trodden academic and professional training pathways. This dynamic manifests in rigid roles and low labor fungibility, making the activation energy for role and process change uniquely high.
5. A high cost of error. Mistakes carry patient-safety, regulatory, legal and financial consequences. It’s difficult for pharma R&D to fail fast and cheap.

Taken together, these five features make pharma R&D a prime example of the “vertical AI” thesis, which argues that the largest AI value opportunities in complex regulated industries are captured not by horizontal, general-purpose tools but by deeply embedded, domain-specific transformation.

What this demands for pharma R&D: Concentrated bets and the 40/40/20 rule

If R&D value is trapped in a long, regulated, specialized system, it follows that AI value cannot be unlocked in thin slices. Transformational value comes only from large, concentrated bets and deliberate investments that pair AI with getting data ready, reengineering processes and changing roles.

Rebuilding a data layer, deleting steps from an entrenched process and questioning who does what all takes focus and a willingness to expend political capital. No organization can tolerate many of these efforts at once, so it’s critical to pick a few value streams genuinely worth the pain—and then go deep.

In our experience, it’s rarely worth undertaking a transformation unless the expected return is substantial (ideally eight or nine figures for a major pharma company). That discipline forces larger bets in fewer places.

This is where budget and energy allocation becomes the single most reliable predictor of success. We recommend following the 40/40/20 rule. While we aren’t the first to highlight the need to place more effort on data and process transformation, our goal is to focus on the manifestation of this dynamic in pharma R&D.

• About 40% of effort and investment should go to building a contextualized knowledge and data layer
• About 40% should go to redesigning workflows, decisions, roles and operating models
• About 20% should go to the AI capabilities themselves

Most AI programs fail because they invert this ratio, pouring the bulk of their attention into AI capabilities and starving the other two components. Firms around the industry are beginning to recognize that organizations overfund the technology and underinvest in everything around it.

While it’s called the 40/40/20 rule, the exact percentages matter less than the direction. Organizations that fund AI as a plug-and-play technology stay stuck in pilots, while those that budget for the knowledge and operating model work around AI are the ones that scale. Two examples from very different corners of R&D show the importance of this framework in practice.

Rebuilding the knowledge and data layer in pharma R&D: The first 40%

The first job of any serious R&D transformation is to make scientific knowledge usable by machines as structured, curated, governed context—rather than as documents a human must read and rekey. Without this layer, AI outputs tend to plateau at what teams describe as “60% usable,” demanding heavy oversight and frequent rework.  The knowledge foundation simply isn’t there, so every user ends up resupplying the context the application lacks. It’s also why pharma commercial chief information officers, who first built these foundations, have been able to scale AI where R&D has not.

A contextualized knowledge layer for R&D rests on four elements:

Modular, reusable content: Today a safety narrative or an endpoint definition is written, buried in a document and rewritten for the next study. But when captured as standardized building blocks, this same content can be reused across studies, submissions and workflows instead of recreated every time.

Domain ontologies and semantic layers: A shared way to describe scientific concepts and how they connect, so the same trial outcome, molecule or safety signal means the same thing across every system, team and study.

Knowledge graphs and governed repositories: Scientific, operational and trial data connected into a map that AI can follow and reason across, instead of sitting in static folders and files it can’t see.

Traceable, governed context: Every output can be traced back to an approved source and its review history. This is a non-negotiable in a tightly regulated GxP environment, where the work has to be documented and auditable.

The payoff of getting this right is larger than any single use case. A well-governed knowledge layer connects data that normally sits in isolation and makes whole classes of cross-cutting questions answerable for the first time. Forward and reverse translation is one example: with preclinical, clinical and translational data spanning the same layer, it becomes possible to test at scale whether an in vitro signal actually predicted what happened in humans. This is a question R&D organizations have long wanted to ask but rarely could. It’s also why so much of the investment belongs here, as it’s built once and drawn upon many times, making every use case that follows faster, cheaper and better.

Redesigning decisions, workflows and operating models: The second 40%

Making knowledge usable is necessary but not sufficient to substantially move enterprise cycle times. Even with a strong knowledge layer, AI layered onto workflows built for human coordination tends to save time for individuals without moving enterprise cycle times. The second 40% of the equation is the redesign of the decisions, workflows and operating models around the AI.

When AI frees up time in a review-heavy process, organizations tend to spend it on more reviews and more iterations rather than banking it as a faster decision. This causes task-level speed to evaporate before it can become enterprise speed. Drafting a clinical monitoring plan doesn’t create value if the review and approval cycle around it remains unchanged.

Organizations that get this right begin with high-value decisions, not tasks, and redesign the surrounding process and governance to support them. They also recognize that AI doesn’t merely accelerate existing work but can change the nature of the work itself.

• In discovery, computational hypotheses increasingly drive the search while the wet lab shifts toward validating and refining them. Target identification stops being linear screening and becomes a loop between application and experiment.
• In clinical monitoring, protecting data quality and patient safety shifts from uniform manual checks toward targeted, risk-based oversight. As routine verification is automated or risk-targeted, the site monitor’s role moves from line-by-line checking toward relationships, judgment and engagement.

The most valuable move in redesign is often not the most obvious one: Choosing the altitude at which you look at the work. If you frame the problem too narrowly (example: a single document or task like authoring the clinical study report), the most you can achieve is accelerating the existing workflow.

But if you instead frame it around a whole value stream (example: the flow from a trial’s last patient visit through to regulatory submission), more powerful questions come into view: Which steps and documents need to exist at all, which can be collapsed or reordered and where can whole handoffs be removed? That difference in altitude is what separates genuine redesign from contained acceleration. The point of AI is rarely to do the old work faster, but rather to make some of the old work unnecessary.

At the value-stream level, AI can support a continuous cycle of sensing new information, modeling possible outcomes, making decisions, acting and learning from the results. This shifts the focus from producing outputs, such as draft content, to achieving outcomes, such as faster submissions and better asset and study decisions.

AI shifts people toward higher-level work

As the work changes, so do the roles, in two connected ways. Most of the change comes from taking existing roles apart, task by task, and deciding which activities an agent handles, which stay with a person and which they do together.

The person’s role typically evolves to tackle harder, higher-judgement work. Clinical monitoring is a good example. A clinical research associate (CRA) today spends much of the week focused on source data verification and line-by-line checking of trial records against source documents, while also managing queries, visiting sites and engaging site staff. And agents today can oversee routine verification, monitoring the data continuously and surface only what looks wrong.

This is the direction risk-based monitoring has been headed for years. In this new model, the CRA’s job doesn’t shrink so much as move up to interpreting the anomalies the agents surface, judging which sites are genuinely at risk and handling the exception and relationships.

Deploying AI in context: The final 20% for pharma R&D

The 80% we’ve discussed doesn’t diminish the AI itself, which is necessary and far from trivial. But the real work in this final 20% is less about choosing an application than about designing and validating it for the context it will operate in—because in R&D, that context swings more widely than almost anywhere else.

Outside the GxP boundary (examples: exploratory research, literature synthesis or hypothesis generation), the goal is speed and reach. A wrong answer costs a wasted lead or an hour of a scientist’s time, so an application can be deployed quickly, iterated on and governed by light human review.

Inside GxP regulation, where AI helps produce a submission dataset, a safety narrative or a study report, the need is entirely different. Outputs must be reproducible, traceable and defensible to a regulator. The validation effort can exceed the work of building the application itself, because the AI must earn at least the same assurance as the human controls it replaces.

Strong AI can fail on foundations

This calibration cuts both ways and can be costly if done wrong. If you do too much validation on exploratory work, you throw away the speed that justified it in the first place. But if you skimp on it where regulators are involved, you risk building something that can’t make it into a submission.

This is also why the application is rarely what separates one organization’s results from another’s. Capable applications are increasingly available to everyone. What differs is whether that application sits on a real knowledge layer and inside a redesigned workflow validated appropriately for its use or is simply dropped in the old ones. A capable application on a weak foundation still produces the “60% usable” output teams learn to distrust, while a more modest one on strong foundations, validated for its context, can be relied on. The final 20% that’s focused on applications is what converts the other 80% into value—but only once that 80% is in place.

The path to realizing AI’s transformational potential in R&D

In R&D, the applications have never been the hard part. Too many programs pay too little attention to the data underneath the applications and the workflows around them.

The data and the workflows are the 80% that the industry keeps underfunding. To truly show value and scale AI in R&D, plan for the 40/40/20 split to show up in your budget and calendar, not just in your strategy. That’s what will separate R&D organizations that realize AI’s generational promise from those who are still waiting for it.

Want to go deeper? Attend our webinar with leaders from Merck, Johnson & Johnson Innovative Medicine, Amgen and ZS: AI, judgment and the human imperative: The future of biostatistics.