AI in Pharma's Quiet Wins: Clinical Trials and Manufacturing

While AI drug discovery chases its first approval, AI is already regulator-qualified and deployed in how trials are designed and medicines are made

Prepared by Richstorm.co

Key Takeaways

• Digital twins and synthetic control arms have received formal regulatory qualification — the EMA has qualified Unlearn’s PROCOVA methodology — and are already reducing control-arm sizes in live trials.

• Sanofi has eliminated entire Phase 2 cohorts using virtual patient models, and Medicenna secured FDA agreement for a hybrid synthetic control arm that saved roughly 100 control patients.

• Quantitative Systems Pharmacology (QSP) modeling has informed FDA regulatory decisions for over a decade, predating the current generative-AI wave by years.

• AI and automation are being built directly into pharmaceutical manufacturing, including AstraZeneca’s planned Virginia drug substance facility within its $50 billion U.S. investment.

• The most-hyped AI application in pharma — molecule discovery — has produced one validated clinical success so far; the least-hyped applications — trial design and manufacturing — are already regulator-qualified and in active use.

• This gap between hype and proof is mostly a story about narrative: discovery offers dramatic before-and-after numbers and fits familiar generative-AI imagery, while trial design and manufacturing gains are incremental and largely invisible to the public.

Two Ways AI Can “Run” a Clinical Trial

When people talk about AI in clinical trials, they are often describing one of two quite different things: simulating what would have happened to a patient without treatment, or simulating how a specific drug will affect a patient given its mechanism of action. Both are real, both are in use today, and both arrived at credibility through very different paths.

Digital Twins: Simulating the Comparator

A digital twin is an individual-level computational model, built from historical trial data, that predicts a patient’s outcome trajectory had they received no active treatment. That predicted trajectory becomes the synthetic control observation, reducing or replacing the need for every enrolled patient to be randomized to placebo.

This has moved well past the research stage. The EMA has issued a formal qualification opinion for Unlearn's PROCOVA digital-twin methodology, with regulators indicating it can reduce control-arm size by roughly 35%. The FDA's CDER, together with the EMA, published ten Guiding Principles of Good AI Practice for drug development in January 2026. In practice, Sanofi has already eliminated entire Phase 2 cohorts using virtual patient models, and Medicenna secured FDA agreement for a hybrid synthetic control arm that saved approximately 100 control patients.

The clearest validation case comes from Alzheimer's research: researchers reanalyzed the donanemab (TRAILBLAZER-ALZ 2) trial using a digital twin generator to quantify how much sample size could have been reduced, or statistical power gained, had digital twins been included from the start — a retrospective proof-of-concept on a real, completed trial.

The honest scope of this method: it primarily shrinks the placebo arm. The treatment arm — real patients receiving the real drug, where safety signals actually emerge — cannot be simulated away. The benefit concentrates in diseases like ALS and Huntington's, where every patient on placebo is someone with a fatal, progressive condition receiving no treatment, and where slow recruitment is often the binding constraint on how quickly a trial can run at all.

QSP and In Silico Trials: Simulating the Drug's Effect

The second method is older than most people realize. Quantitative Systems Pharmacology (QSP) models are mechanistic mathematical representations of disease biology and a drug's mechanism of action, run across a “virtual population” of simulated patients to predict treatment effects.

This is not a new AI fad — 2023 marked the tenth anniversary of the first QSP-influenced FDA regulatory decision, the so-called Natpara case, in which the FDA's Office of Clinical Pharmacology used a calcium homeostasis QSP model to support a request for a post-marketing trial. Since then, the FDA's Model-Informed Drug Development (MIDD) Pilot Program has formally recognized QSP as a tool for dose optimization and trial design, and regulators in 2025 actively expect sponsors to demonstrate how modeling influenced these decisions.

A live example: the SIRIUS program built a mechanistic model of atherosclerotic cardiovascular disease to predict the cardiovascular benefit of the cholesterol drug inclisiran ahead of its Phase III readout — extrapolating from inclisiran's known short-term effect on LDL cholesterol, already measured in real patients, to long-term outcomes like heart attacks and strokes that would otherwise take years longer to observe.

The key constraint, especially relevant for AI-discovered molecules: a QSP model needs real human data on a drug's mechanism and short-term effects to calibrate its predictions. For a brand-new molecule with zero human dosing history, there is nothing yet to calibrate against — so this method works best as a force multiplier after an initial real dataset exists, extrapolating a small amount of real evidence across a larger virtual population and longer time horizon, rather than replacing the need for that initial real data.

AI Is Also Quietly Reshaping Manufacturing

The least visible application of AI in pharma may be the factory floor. AstraZeneca's announced $50 billion U.S. investment includes a new drug substance facility planned for Virginia — intended to be the company's largest single manufacturing investment in the world — where AI, automation, and data analytics are explicitly built into the production process for the company's small molecule, peptide, and oligonucleotide portfolio, including its weight-management and metabolic pipeline.

This kind of investment rarely generates headlines, in part because “manufacturing optimization” doesn't lend itself to a dramatic before-and-after story the way “AI designs a drug in 21 days” does. But it represents real capital committed to AI as production infrastructure, not just as a research tool — and manufacturing reliability directly affects drug pricing, supply continuity, and availability for patients.

Why Discovery Gets the Spotlight — and the Others Don't

The imbalance in attention between AI drug discovery and AI in trials or manufacturing is mostly a story about narrative structure, not actual impact.

Discovery offers a dramatic, easily quotable contrast: a lead compound generated in 21 days versus a traditional two-to-four-year process. Clinical trials don't have an equivalent number, because much of a trial's duration is fixed by biology and ethics — waiting years to observe whether a disease recurs, or whether a side effect emerges, cannot be compressed by better software. AI can remove administrative friction from a trial, but it cannot remove the wait itself.

Discovery also maps neatly onto the generative-AI narrative most people already have from tools that write text or generate images: “AI designs a new molecule” feels like a natural extension of “AI writes an essay.” Trial design statistics and manufacturing processes don't connect to any AI experience most people have had, even before AI enters the picture. And discovery sits at the intersection of Silicon Valley and pharma — DeepMind's Isomorphic Labs, NVIDIA's partnerships with Recursion — which earns coverage from both tech press and pharma press simultaneously. A digital twin reducing a control arm by 35% is a pharma-trade story; it rarely crosses into general technology coverage.

The Scoreboard

Put side by side, the picture is almost the inverse of the public narrative. AI drug discovery — the most-discussed application — has produced one validated Phase IIa clinical success (rentosertib) and zero FDA approvals across the entire field as of mid-2026. AI in clinical trial design — digital twins and QSP modeling — has formal regulatory qualifications from both the EMA and FDA, and is already changing how live trials are run at companies including Sanofi. AI in manufacturing is backed by tens of billions of dollars of committed capital at major pharmaceutical companies.

None of this means AI drug discovery is overhyped in an absolute sense — it remains a genuinely early-stage field with a real, if singular, proof point, and the timeline-compression evidence (discussed in our companion piece on AI drug discovery) suggests a real structural improvement is underway. But for an investor or analyst trying to separate signal from noise, the parts of AI-in-pharma that are quietly furthest along in proof and regulatory acceptance are not the parts getting the headlines — and that gap is, itself, useful information.