Roche disclosed decision contribution, not just AI adoption

At its September 28 Pharma Day, Roche said 40% of pipeline decisions from Q4 2025 through Q2 2026 had a tracked AI and/or computational contribution.

It also said it is on track for Target Nexus to contribute to 80% of research portfolio decisions by the end of 2026. The two percentages are not a direct before-and-after comparison: the 40% refers to past pipeline decisions, while the 80% is a goal for a specific tool across research portfolio decisions.

Still, the direction is clear. AI is moving beyond literature review and analysis into the infrastructure used to decide what research to continue and where to allocate resources.

The next step is not a lab that chats with AI — it is a Lab in the Loop

Roche's Pharma Day agenda included a session titled ‘Revolutionizing drug R&D with Lab in a Loop.’ Genentech describes this as a cycle in which lab and clinical data train AI models, models propose targets, molecules or experiments, scientists test them, and the results flow back into the models.

The loop is roughly Data→AI prediction→Experiment→New data→Model update.

Genentech explicitly frames the model as augmenting scientists rather than replacing them, enabling faster iteration while freeing teams to ask larger scientific questions.

Roche says it has started building autonomous AI-driven labs

Reuters reported Aviv Regev saying Roche has recently started building autonomous AI-driven labs.

In a conventional Lab in the Loop, humans still perform much of the physical execution after a model proposes the next experiment. Deeper lab automation can allow automated systems to execute experiments, measure results and return new data into the learning loop faster.

That does not establish a fully human-free laboratory. The more defensible implication is that repeatable execution can become more automated while human work shifts toward problem framing, experimental design, failure interpretation, validation and portfolio judgment.

More than 3,500 GPUs are becoming part of the laboratory infrastructure

In March 2026, Roche announced an additional 2,176 NVIDIA Blackwell GPUs, taking its hybrid-cloud AI factory to more than 3,500 GPUs.

Roche connects that infrastructure not only to molecule design but to Lab in the Loop, clinical trials, manufacturing digital twins, diagnostics and digital pathology.

Pharma R&D infrastructure is therefore expanding from scientists and lab equipment toward a combined system of automation, data, accelerated computing and AI models.

The talent shift is not simply Wet Lab Scientist down, AI Scientist up

As AI connects directly to physical experiments, Biology, Chemistry, Experimental Design, Lab Automation, Data Engineering, AI/Computational Science, Validation and Scientific Judgment meet inside the same workflow.

This does not mean one person must own every capability. It means the interfaces matter more: AI researchers need biological constraints, wet-lab scientists need to validate model outputs, automation engineers need to translate protocols into machine-executable workflows, and research leaders need to make investment decisions from AI plus experimental evidence.

The scarce layer may increasingly be the ability to connect specialist domains rather than the existence of each specialty by itself.

Do not attribute Roche's Phase III improvement directly to AI

Roche also said its Phase III success rate increased from 65% in 2025 to more than 80% year-to-date in 2026 and that it aims to launch up to 20 new molecular entities by 2030.

It also described roughly CHF 2 billion of R&D savings being reallocated to programs and productivity initiatives.

Public evidence does not establish that AI directly caused the Phase III improvement or all of those savings. Portfolio selection, trial design, molecule quality, therapeutic-area mix and other operating changes can contribute.

BANSEOG VIEW | The bigger change may be AI redesigning the research loop

The most useful Roche signal is not how many AI tools it has adopted. AI contribution is already being tracked in pipeline decisions, and the company is now moving toward automating the loop between model prediction and physical experiment.

As experimental execution becomes more automated, human value does not necessarily disappear. Choosing what to test, distinguishing biological signal from process failure, validating recommendations and deciding where to allocate resources can become more important.

For biotech talent intelligence, counting AI Scientist roles is therefore not enough. The more revealing question is which capabilities companies are connecting to build a closed loop across biology, automation, data, AI and validation.

BANSEOG VIEW

Banseog View — watch AI change the research loop, not just replace tasks

AI/computational contribution was tracked in 40% of recent pipeline decisions, while Target Nexus has an 80% research-portfolio contribution goal for year-end.

Roche says it has started building autonomous AI-driven labs, extending the loop from AI prediction toward automated physical experimentation.

The biotech talent shift may be less about replacing wet-lab scientists with AI scientists and more about redesigning the interfaces among biology, automation, data, AI, validation and scientific judgment.

SOURCES

Primary sources and references

The 40% and 80% figures are not the same metric measured at two points in time. The 40% refers to tracked AI/computational contribution in pipeline decisions from Q4 2025 through Q2 2026; the 80% is the year-end goal for Target Nexus contribution to research portfolio decisions. Roche's Phase III 65%→80%+ improvement and CHF 2B R&D savings are not attributed directly to AI in this article.