AI was supposed to remap pharma from the molecule up. Instead, the clearest economic return is arriving farther downstream, where the work is less glamorous and easier to measure: manufacturing, quality, supply chain, and administrative operations.
That is the uncomfortable pattern emerging from pharma’s latest AI cycle. Eli Lilly’s digital leadership has publicly said the real gains are not in discovery. Recursion, one of the most visible AI-native pioneers in drug discovery, has spent more than a decade on the thesis and still does not have an AI-developed drug on the market. And despite the broader industry enthusiasm, there is still no definitive evidence that AI improves clinical trial outcomes.
The result is a sharp pivot in where value is actually being captured. AI is saving pharma billions, but the savings are accruing in the operational layers of the business: manufacturing efficiency, back-office automation, planning, and data-heavy workflows where models can be tied to concrete throughput, cycle time, and labor metrics. The lab remains important, but it is not yet where the ROI story is most convincing.
Why the lab is not where the money is
The industry’s original AI narrative treated drug discovery as the highest-value target because the upside seemed obvious: better hit generation, faster lead optimization, fewer wet-lab iterations. In practice, that promise has collided with two realities. First, biology is noisy and data is sparse in exactly the ways that make model validation difficult. Second, the economic bar in drug discovery is not “interesting prediction,” but an actual approved drug that survives the full chain of preclinical work, clinical development, and regulatory review.
That bar is why the discovery ROI gap remains so visible. Recursion’s trajectory is the cautionary example: years of investment, a large AI-forward platform, and still no AI-discovered drug in market. That is not proof that discovery AI is futile. It is proof that the timeline from model to monetization in this setting is long, uncertain, and difficult to benchmark with standard software economics.
By contrast, manufacturing and back-office use cases are easier to operationalize and easier to defend financially. If a model improves batch release workflows, reduces deviations, accelerates documentation, trims planner workload, or flags quality issues earlier, the benefit shows up in familiar enterprise terms. Those gains do not require a drug approval to justify the investment.
What the evidence can and cannot claim
There is a tendency to collapse all pharma AI into a single question: is it working? That question is too broad. The real issue is which workflow is being measured, by what metric, over what time horizon, and with what operational controls.
The current evidence base is much stronger on process optimization than on therapeutic or clinical impact. Analysts continue to project substantial savings from AI in pharma, on the order of tens of billions of dollars, but those forecasts depend on deployment in enterprise systems rather than on speculative breakthroughs in model biology. In other words, the savings case is tied to implementation depth.
What is not there yet is conclusive proof that AI improves clinical trial outcomes. That distinction matters. A model might help identify sites, draft documents, clean trial data, or optimize enrollment operations without changing endpoint success rates. Those are useful improvements, but they are not the same as demonstrating that AI materially increases the probability of trial success or shortens the path to approval.
For technical teams, that means measurement discipline matters more than marketing language. If the value proposition is manufacturing, measure yield, deviations, downtime, release time, and labor hours. If the value proposition is operations, measure cycle time, error rate, and throughput. If the pitch is trial support, measure the workflow, not the hoped-for clinical endpoint.
The stack below the stack
The strongest AI use cases in pharma are also the ones that expose how much work remains in the data layer.
Manufacturing AI only works if the model can see the process. That means integration with MES systems that capture shop-floor execution, LIMS platforms that manage lab data, and ERP layers that hold planning, inventory, procurement, and financial context. These systems are often fragmented, inconsistent, and implemented differently across plants and business units. Without clean interfaces among them, the model becomes a demo rather than a production tool.
This is why the shift in pharma AI is as much an infrastructure story as a model story. Companies need scalable platforms that can ingest structured and semi-structured data across the manufacturing and back-office estate, support low-latency inference where necessary, and produce outputs that are auditable enough for regulated environments.
That regulated setting changes the requirements again. Model governance is not a nice-to-have. Pharma needs validation frameworks, version control, traceability, and clear handling of drift, retraining, and human override. If an AI system influences a batch decision, a quality workflow, or a planning process, operators need to know what the system saw, why it produced a result, and how that result was reviewed.
The Lilly–NVIDIA relationship is useful here less as a product announcement than as a signal about where the industry believes industrial-scale AI will live. Large compute partnerships make sense when the target is not a single model demo, but a sustained platform that can support many workflows across a complex enterprise. The implication is straightforward: pharma is moving toward AI as industrial infrastructure, not AI as a standalone discovery oracle.
What vendors should build, and what they should stop promising
For vendors, the market is clearing a preference signal.
Products that succeed will be deployment-ready rather than aspiration-heavy. They will plug into existing MES, LIMS, ERP, and document systems. They will expose the hooks needed for validation, auditability, and role-based oversight. They will show improvement in process metrics that procurement and operations leaders can verify without waiting for a phase 3 trial.
That also means positioning has to change. The old pitch was: AI will transform discovery, and the rest will follow. The new pitch is more grounded: AI can reduce cost and friction in the industrial core of pharma today, with discovery as a longer-term option value, not the primary ROI engine.
Recursion’s experience reinforces that message from the other side of the market. Being early in discovery AI is not enough if the path to commercialized output remains opaque. By contrast, firms that connect AI to manufacturing execution or operational throughput can point to visible savings and a much shorter feedback loop. That makes the buying decision easier and the deployment case more defensible.
The competitive edge, then, is not merely better models. It is better integration, better governance, and better evidence.
The funding signal changes sentiment, not physics
A new funding moment is reinforcing the idea that AI matters in pharma, but it should not be confused with proof that the industry has cracked discovery or clinical development. Capital tends to follow where there is a credible deployment path, and right now that path is most visible in production-scale operations.
That is why the current wave of investment is best read as a confidence signal for AI relevance in pharma, not as a declaration that the lab breakthrough has finally arrived. Investors and strategists are increasingly rewarding platforms that can show enterprise savings, not just scientific promise. The practical consequence is that vendors will be pressed to demonstrate measurable impact in manufacturing and back-office workflows while maintaining a longer-term discovery roadmap.
If the last decade’s AI story in pharma was about big claims and uncertain timelines, the next phase looks more like industrial software: integrated systems, controlled rollouts, real governance, and incremental but monetizable gains. That is less dramatic than the original hype cycle. It is also, for now, much closer to where the money is.
