Roche Is Giving AI a Job Review. It Wants AI in 80% of Drug Decisions.

London, late September. At Roche’s annual investor event Pharma Day, Aviv Regev, head of Genentech’s Research and Early Development, opened with a slide that had nothing to do with the pipeline or sales. It was a set of statistics on AI contribution.

Investors and analysts in the room were used to hearing pharma companies say AI sped up a project. Regev described something else. Roche now tracks AI’s contribution to R&D decisions the way it tracks sales. From Q4 2025 to Q2 2026, 40% of pipeline decisions had traceable AI or computational contributions. In Q2 2026 alone, that figure reached 44%. Roche breaks those contributions into three tiers. 41% were critical and changed the direction of a decision. 48% were supportive and gave decision-makers more confidence. 10% were informational and only fed discussion.

Roche has spent more than five years building this system. Regev has led the effort since joining Genentech in 2020. AI now touches every eligible large-molecule program and nearly half of small-molecule programs at Genentech. The next target is higher. TargetNexus, a target-assessment agent, is supposed to join 80% of research portfolio decisions by Q4 this year.

When a drug company starts scoring AI like a sales team, the tool becomes part of the floor. AI is no longer a lab assistant. It is infrastructure. Infrastructure raises a question: who owns it?

1. Moving targets upstream: pushing AI to the most dangerous place

The most expensive mistake in drug development is picking the wrong target. Once the target is wrong, every later dollar makes the loss bigger. Roche’s decision to put AI at the front of the chain follows that lesson.

TargetNexus reviews candidate targets across three levels: target biology, drug discovery, and clinical and portfolio. It scores 11 dimensions. The company also changed how projects start. Any researcher can nominate a target. After the nominator prepares a data package, a target assessment portfolio team gives a recommendation in about two weeks on whether to start.

That changes who decides. Target selection used to depend on senior scientists’ intuition and experience. Discussions could run for months. Now an agent produces a structured review in two weeks. The human role shifts from decision-maker to reviewer. Regev did not pretend otherwise. She said AI cannot replace human clinical trials, and confidence in model reliability has to build over time. She also gave a harder target: TargetNexus should join 80% of research portfolio decisions by Q4.

Roche has numbers to support the push. With resource input flat and the lead identification cycle cut in half, the number of projects entering lead identification rose 155% from 2022 to 2025. Internally generated development candidates rose 50%. Those figures appear in investor materials.

The numbers need context. Roche has not disclosed how it judges AI’s contribution to a decision. FourWeekMBA, a business analysis site, made a rough calculation: multiply 40% by 41%, and AI critically influenced about 16% of pipeline decisions. The Phase III success rate also needs unpacking. It rose from 65% in 2025 to more than 80% now. In late 2023, Roche introduced a project evaluation framework called The Bar, which raised the standard for advancing projects. Drug development cycles last years, so most projects now in Phase III began before the lab-in-the-loop system was widely deployed. Roche has not separated how much of the success rate comes from AI picking better molecules and how much comes from killing weak projects earlier.

Those caveats do not change the near-term goal. Roche wants AI in more decisions. Going from 40% to 80% requires infrastructure.

2. The four layers under the loop

Regev calls that infrastructure Lab-in-the-Loop. The idea is simple. Train models on experimental or clinical data. Let the models predict the next experiments worth running. Feed the results back into the models. Repeat. Each turn gives a project key predictions and makes the model better across all projects.

The loop is already part of daily work. Small-molecule optimization runs weekly. Models propose and score molecular designs. Selected molecules go to synthesis and testing. The data returns to the model. Roche also described a biologic project against a new target. It went from lead identification to clinical candidate selection in 18 months, about twice as fast as before. It completed first patient dosing this year.

Data, compute, and experimental capacity keep the loop moving.

Data comes from two sources. Internally, Roche started a foundation dataset program to produce data for training foundation models. Externally, it works with Recursion. In August 2026, the two said they were advancing the first AI-discovered target, for an undisclosed neurodegenerative disease.

Compute comes mostly through partnerships. In March 2026, Roche added 2,176 GPUs through a partnership with Nvidia. Its total on-premises and cloud GPUs now exceed 3,500, one of the largest clusters in pharma.

Experimental capacity is harder to buy. At the end of 2025, Genentech partnered with Medra, a robotic biological experimentation company. The two connected Genentech’s lab information management system and machine learning platform with Medra’s autonomous execution layer. Prediction, experimentation, validation, and improvement now feed one another automatically. Roche then proposed the Autonomous Lab Accelerator. It borrows from self-driving cars: six levels of lab autonomy, from L0, fully manual, to L5, fully autonomous.

The organizational layer shows up in project review and budget. Roche plans to reallocate about CHF 1.96 billion saved from R&D efficiency gains between 2024 and 2026. CHF 486 million of that goes to future capabilities such as lab-in-the-loop and organoids.

Regev calls the strategy AI independence. Roche’s partner list is long: Nvidia, OpenAI, Anthropic, Medra, Recursion, AWS. Compute and general-purpose models can be bought. Roche wants to keep the data and experience created inside the loop. In Regev’s words, algorithms and compute can be bought. The data and validation capability produced by the loop must stay in Roche’s hands.

3. Who controls which layer

Many drug companies use AI. They do not agree on what to own.

Eli Lilly is closest to Roche. It also wants a loop where wet labs and dry labs push each other. In January 2026, Lilly said it would build an AI co-innovation lab with Nvidia in the San Francisco Bay Area. The two plan to invest up to $1 billion over five years. Autonomous wet labs and computational dry labs are supposed to update each other around the clock.

Novo Nordisk chose a different route. It has not announced plans to build its own autonomous lab. It has partnered with OpenAI, Anthropic, and AWS. Novo wants Big Tech AI across drug discovery, manufacturing, supply chain, and commercial operations. CEO Mike Doustdar said he expects AI to accelerate everything from basic biology and pharmacology to drug development and commercialization. The goal is to become the world’s most AI-driven healthcare company.

The split comes down to which layer is worth owning. Roche and Lilly bet on the loop between wet labs and computational models. They think real experimental feedback is something models cannot learn from public data. Control the wet lab, and you control the entry to the loop. Novo bets on breadth at the application layer. It spreads AI across as many business links as possible and trades partners’ model capability for speed and coverage.

AI companies are splitting too. Alphabet’s Isomorphic Labs went furthest. It develops its own pipeline, raised a $2.1 billion Series B in May 2026, and aims to enter clinical trials by the end of 2026. It controls prediction, validation, and the drug itself. Anthropic bought Coefficient Bio, a biotech startup founded eight months earlier, for about $400 million. The deal gives Anthropic wet-lab capacity. The company says it will not run clinical trials or compete with pharma. Robotics companies like Medra sit between the two. Medra deploys autonomous experimental systems for Roche and other drug companies. Their loops depend on it. Medra is also building its own labs and accumulating training data.

Drug companies are growing AI outward from R&D. AI companies are extending from models into labs. Robotics companies are moving into both. The choices reflect a dispute over who owns data and validation. Clauses about data ownership and model improvement will likely become the next fight.

4. Humans sign, but AI’s role grows

The faster the loop turns, the harder responsibility becomes to assign. If an AI-assisted decision fails, who answers for it?

If TargetNexus reaches 80% coverage and a decision where AI played a critical role fails in later clinical trials, does responsibility fall on the scientist who approved it or the management that designed the system? Roche’s lab autonomy levels offer a partial answer. Alongside the six levels, the human role shifts from operator to assistant to approver to auditor. Scientists move from doing experiments by hand to signing off and reviewing afterward.

Regulation supports human sign-off. In December 2025, the FDA qualified the first AI drug development tool, AIM-NASH. The agency required pathologists to review AI output, decide whether to adopt it, and take full responsibility for the final interpretation. A human signs. But AI now carries more weight in what that human signs. Roche has not explained how responsibility is divided inside the company.

Regev also said confidence in model reliability has to build over time. AI cannot replace human clinical trials. The lab-in-the-loop system currently compresses the path from target to clinical candidate. No approved drug has come out of Roche’s loop yet. AI is already in Roche’s operating metrics for investors. Clinical data will determine its final grade.

5. Where the loop ends

Roche putting AI into KPIs sends a signal. Drug R&D is moving from scientist intuition to model-driven work. The core of model-driven work is the loop: data, compute, experiments, and decisions feeding one another. Whoever controls that cycle controls the infrastructure for the next generation of drug R&D.

The loop has limits. It compresses the path from target to clinical candidate. Whether a drug reaches the market depends on clinical data. Regev did not avoid that point. The 18-month case she showed had only completed first patient dosing. Approval is far away.

The larger question is what happens when the loop becomes a drug company’s core asset. How will AI companies, robotics companies, and drug companies rearrange control over data and validation? Data ownership and model improvement clauses will become the focus of the next round. The most basic question, who is responsible for the loop, has no clear answer yet.

Roche’s lab-in-the-loop keeps running. The first approved drug to come out of it will show whether the system works.