Roche said at its pharma investor day on 28 September 2026 that it has started building autonomous AI driven labs. That got the headlines. The more useful disclosure was a different one: 40 percent of its pipeline decisions between the fourth quarter of 2025 and the second quarter of 2026 included a tracked AI or computational contribution, and its internal tool Target Nexus is on course to contribute to 80 percent of research portfolio decisions by the end of this year. The important word in that sentence is tracked. Almost no other pharma company can produce that number, and the reason is not that they use less AI.
What Roche actually announced
The substance, before anyone's interpretation of it:
- Roche has begun building autonomous AI driven labs, described by Aviv Regev,
- o runs Genentech Research and Early Development, as part of a commitment to
- independence in R&D
- Around 2 billion Swiss francs of R&D savings is being redirected into
- ogrammes and productivity work
- The company is targeting up to 20 new molecular entities by 2030
- 40 percent of pipeline decisions from Q4 2025 to Q2 2026 carried a tracked AI
- computational contribution
- Target Nexus remains on track to contribute to 80 percent of research
- rtfolio decisions by the end of 2026
- Phase III success rate rose to above 80 percent year to date in 2026, against
- percent in 2025
Most coverage led with the labs. The labs are the most visual part and the least immediately transferable.
Why "tracked" is the hard word
Plenty of large pharma companies use AI across discovery and development. Very few can stand in front of investors and state what share of their decisions it contributed to, because saying that requires infrastructure that has nothing to do with models.
To produce that number you need to have decided in advance what counts as a pipeline decision, recorded each one as a discrete event, captured which analyses fed it at the time rather than reconstructed afterwards, and kept that record consistent for three quarters straight.
That is an accounting system for decisions. It is unglamorous, it is expensive, and it is the actual precondition for ever proving AI did anything.
Without it you end up where most companies are: everyone agrees AI is being used widely, nobody can say what it changed, and every claim about its value is a matter of opinion rather than record. That is the position almost every enterprise is in right now, across every industry, and it is why AI budgets keep getting questioned despite obvious activity.
Roche built the ledger. That is the part worth copying, and it does not require a robot.
How to read the phase III number honestly
The jump from 65 percent to above 80 percent in phase III success is the most quotable figure in the release and the one most likely to be misused.
Phase III outcomes in 2026 reflect decisions taken years earlier. A success rate measured year to date covers a small number of readouts, so it moves sharply on a few results. Roche has not claimed AI caused the improvement, and the honest position is that nobody can currently separate AI contribution from portfolio composition, therapeutic area mix, or ordinary variance in a small sample.
It is a genuinely good number. It is not yet evidence of anything in particular. Treat anyone presenting it as proof that AI fixed clinical development as someone selling something.
What an autonomous lab actually is
Stripped of marketing, a self driving lab is a modular experimental platform where a machine learning system selects the next experiment, the equipment runs it, the results feed back, and the loop repeats toward an objective a human set. Nature Synthesis describes it in roughly those terms. The more capable versions now cover most of the scientific method: hypothesis, design, execution, analysis, and revision of the hypothesis.
This is not new in 2026 and Roche is not first. AstraZeneca has iLab. Insilico Medicine runs an autonomous discovery pipeline. Telescope installed an autonomous lab for Pfizer in January 2025 and another for the Korean Pharmaceutical and Biopharmaceutical Manufacturers Association that December.
On the speed claims, one caution. A 2025 preprint found a median acceleration factor of about 6 across self driving lab studies. That is a striking number and it is worth knowing, but it is an unreviewed preprint aggregating heterogeneous studies, and the honest reading is that it is suggestive rather than established. I would not put it in a board deck as a planning assumption.
What this means if you are not Roche
Very few organisations are going to build autonomous labs. Almost all of them have the same underlying problem Roche solved first.
You cannot prove AI worked because you never recorded the decisions. This is the transferable lesson and it costs a fraction of a lab. Define what counts as a decision in your function, record each one as it happens, and capture what informed it at the time. Three quarters later you have something no competitor can argue with.
Automating the experiment is easier than automating the judgement. The robots were never the bottleneck in research, and the equivalent is true everywhere else. What slows organisations down is deciding what to do next and getting agreement on it. Roche's Target Nexus target of 80 percent of portfolio decisions is aimed at that layer, not at the bench.
Savings only matter if they are redeployed. The 2 billion Swiss francs is being moved into programmes, not taken as a margin improvement. An efficiency programme that ends in a cost line looks identical on a spreadsheet to one that ends in more shots on goal, and only one of them compounds.
Instrument before you scale. Roche can expand AI in R&D because it can measure what expansion does. Companies without that measurement are scaling on faith, which works right up until a CFO asks a direct question.
The summary
The autonomous labs are real and they matter eventually. They are also the least copyable part of the announcement for almost everyone reading about it.
The copyable part is that Roche can say what share of its decisions AI touched, with a number, in public, to investors. That capability came from treating decisions as things worth recording, which is a choice available to any company of any size, in any industry, starting this quarter.
Most companies will read this story as being about robots in laboratories and conclude it does not apply to them. It is actually about bookkeeping, and it applies to everyone.
Frequently asked questions
What did Roche announce about autonomous AI labs? At its pharma investor day on 28 September 2026, Roche said it had begun building autonomous AI driven labs as part of a commitment to AI independence in R&D. It also disclosed that 40 percent of pipeline decisions from Q4 2025 to Q2 2026 carried a tracked AI or computational contribution, that its Target Nexus tool is on track to contribute to 80 percent of research portfolio decisions by end of 2026, and that around 2 billion Swiss francs of R&D savings is being redirected into programmes.
What is an autonomous or self driving lab? It is a modular experimental platform where a machine learning system chooses the next experiment, automated equipment runs it, results feed back into the model, and the cycle repeats toward a human defined objective. The most capable versions cover hypothesis generation, experimental design, execution, analysis, and revision of the hypothesis.
Did AI cause Roche's phase III success rate to improve? There is no evidence of that, and Roche has not claimed it. Phase III readouts in 2026 reflect decisions made years earlier, a year to date rate covers few events and moves sharply on small numbers, and AI contribution cannot currently be separated from portfolio mix or ordinary variance. It is a good result, not proof of causation.
Why does it matter that Roche can track AI's contribution to decisions? Because almost no other company can. Producing that figure requires defining what counts as a decision, recording each one as a discrete event, and capturing what informed it at the time rather than reconstructing later. Without that record, claims about AI value stay matters of opinion, which is why AI budgets get questioned despite visible activity.
Is Roche the first pharma company with autonomous labs? No. AstraZeneca operates iLab, Insilico Medicine runs an autonomous discovery pipeline, and Telescope installed an autonomous lab for Pfizer in January 2025 and one for the Korean Pharmaceutical and Biopharmaceutical Manufacturers Association in December 2025. Roche's announcement is notable for its scale and for the decision tracking disclosed alongside it.
