Back in December I made an argument in two LinkedIn posts that life sciences companies owning their AI infrastructure would pull ahead of those waiting for platforms to mature or buying off the shelf. That argument did not come from nowhere. Having worked in AI for over a decade, I’ve seen how quickly things that look promising run into limits when the infrastructure, governance, and operating discipline are not in place.
Over the past few months, what I’ve been seeing in pharma is a shift from positioning to production specifics. The companies that committed to infrastructure early are now starting to report real operating outcomes.
The next question is what happens when those same organizations start pushing agents into regulated workflows. Once that happens, the challenge is no longer just infrastructure, it shifts to governance, operating model, and accountability.
The early payoff from life sciences AI infrastructure is showing up in manufacturing and operations, but the next bottleneck is not model performance. It is governance for agents operating inside regulated workflows.
What the December Posts Argued
In Part 1 and Part 2, I argued that life sciences data is part of the accumulated intellectual core of these organizations. It reflects years of research, regulatory work, clinical development, and manufacturing learning. Companies that own the models, training pipelines, and compute infrastructure behind that data are in a much better position to build lasting advantage than companies waiting on external platforms to define the terms.
I used Eli Lilly’s AI supercomputer as the clearest early signal. I made a broader point that tech had already worked out much of the production AI playbook through discipline in MLOps, versioning, governance, and monitoring. Life sciences did not need to reinvent these patterns from scratch. The question I left open was whether those infrastructure investments would translate into measurable results, and what new constraints would show up once they did.
Where the Returns Are Showing Up
What stands out is that the first measurable wins are coming from systems that improve throughput, quality, and yield. That matters because it shifts the AI conversation from promise to operating impact.
Eli Lilly’s chief information and digital officer, Diogo Rau, described three production deployments earlier this year: a computer vision system running quality control on autoinjectors at 70 to 80 photographs per device in hundreds of milliseconds, AI-powered demand forecasting across their supply chain, and a digital twin of the GLP-1 manufacturing process that modeled the device, the machine, and all the surrounding inputs. The digital twin found an optimal configuration that Rau says produced materially different revenue numbers.
Roche has deployed more than 3,500 NVIDIA Blackwell GPUs across the US and Europe. This was a substantial investment and bet, the largest GPU footprint in pharma, with integrated AI workflows into roughly 90% of eligible R&D programs.
Pfizer’s CSO Chris Boshoff described a “Golden Batch” manufacturing initiative that uses AI to identify and replicate peak yield conditions, projecting $0.7 billion in manufacturing savings this year. These aren’t aspirational projections. They’re current-year operating figures from companies that made infrastructure commitments early and are now running at scale.
The Agents Are Here
As infrastructure matures, the next challenge is agentic AI. The issue is not only whether companies can deploy agents, but can they run them effectively and responsibly inside regulated workflows, in situations where traceability, review, and accountability are not optional.
The controls that worked for models producing recommendations do not fully cover systems that take actions, sequence decisions, and operate inside processes where regulatory accountability is real. Once agents participate in the workflow, the governance problem changes shape.
Sanofi’s CEO Paul Hudson recently described that their drug development committee meetings now open with an AI agent’s contextual assessment of whether a drug should advance to the next trial phase. It provides context on the decision across the full body of available evidence, and then the human committee makes the call. That is an agent operating inside one of the highest-stakes clinical governance workflows in pharma.
At Lilly, the TuneLab platform runs agentic orchestration across 18 production models trained on more than 500,000 data points, using a federated architecture where computation moves to where data lives rather than the other way around. Aliza Apple, who heads TuneLab, described the implementation as selecting and sequencing different models across a full drug discovery workflow, and was candid that “parts of the implementation are quite complex.”
An example I keep coming back to is Daiichi Sankyo. They deployed an agentic system for personalizing medical affairs responses, integrated into Veeva. Development took six weeks. The compliance and legal review before launch took nine months.
That ratio tells you a lot about where the real challenge lies. It is in the operating model around the build: who approves, what gets documented, how accountability flows, and what monitoring has to be in place before the system touches a regulated process. The planning mistake is treating compliance as a downstream gate instead of designing for it from the outset.
I’ve been working inside regulated organizations long enough to know that the nine month review wasn’t waste. Responsible deployment into a regulated workflow looks slower on paper because the real work is in validation, review design, exception handling, and traceability.
Governance as Operating Model
The governance conversation in pharma AI still gets framed too often as a policy question: what do we need to approve before we launch? The 2026 evidence reframes it as an operating model question: how do we run these systems responsibly once they’re live, and how do we design with that expectation from the beginning?
MSD’s Chief AI Officer Anton Groom has been clear that all of their AI models have scientists in the loop, not as a post-launch review step, but as a key principle of architectural design. Scientists are accountable for outputs, not sitting off to the side as passive reviewers. This matters in shaping how the systems are designed from the outset.
Greg Ulrich, Mastercard’s Chief AI and Data Officer, offered what I think is one of the clearest descriptions of the real governance tension across industries: “Decentralization without standards creates risk, and centralization without proximity slows impact.” That applies directly to pharma. Heavy central governance boards become bottlenecks as use-case volume grows. Loose local experimentation without shared data standards produces exactly the data quality problems that 68% of pharma tech executives are already identifying as their primary failure mode, according to a ZS Associates survey of 115 pharma and biotech technology executives published earlier this year.
J&J’s Ajay Anand put it simply: “AI rarely fails because the models are weak.” There’s rigor and work surrounding the model that needs to be thought through and invested in.
What to Ask Before You Build
We are already seeing signal from recent infrastructure investments. If your company is still in pilot mode, the gap is widening, and it is getting harder to close because the organizations that moved early now have the foundation to layer agents on top of. That foundation takes time to build, and there are steps you cannot skip. At the same time, you do not need the entire design locked down before you start.
If you’re considering near-term ROI, expect these conversations to focus on manufacturing and operations. Early 2026 evidence from multiple companies is consistent: computer vision quality control, digital twins, batch optimization, supply chain forecasting. Drug discovery AI is a longer play, but if you look at steps in the process, you’ll see ROI already emerging there. If you need to show returns this year, the case is being made most convincingly in operations.
For the agentic question, I would focus on three things before starting any deployment in a regulated environment: do you own or control the infrastructure your models run on, or are you dependent on a platform you cannot modify when requirements change? Are your governance controls in code, tests, and runtime monitoring, or are they sitting in a quarterly review document? And when you deploy an agent into a regulated workflow, have you scoped the compliance timeline before the build starts, or are you going to discover it afterward?
Those questions separate organizations that are ready to operationalize agents from those that will spend months discovering what they should have planned for earlier.
What started as an infrastructure gap is turning into a production gap, and increasingly into an operating model gap between companies that know how to run agents in regulated environments and companies that are still treating governance as a policy exercise.
The companies moving now are not pulling ahead because they wrote better AI policies, they’re pulling ahead because they built the infrastructure, operating discipline, and technical capability to put AI into production where it changes outcomes. In regulated environments, governance matters. But it is still in service of the bigger goal: getting better systems into the workflow faster, more reliably, and at scale.