At Pharma Day 2026, Genentech’s Aviv Regev presented a six-level autonomy ladder for drug R&D labs — and the most consequential row on the slide wasn’t about machines.
This is an investor-day presentation, not a product launch, a published result or a regulatory filing — nothing below says Roche has deployed an autonomous laboratory, and the paper the autonomy slide cites is marked in review, meaning not peer-reviewed. No ladder level is attributed to Roche below: this reading is from the document’s text layer, which does not capture graphical marks, so nothing here claims a current position and nothing here claims the deck failed to indicate one. The roughly 16 per cent figure is this publication’s approximate multiplication of two disclosed figures and does not appear in the deck, and the 80 per cent attached to TargetNexus is a stated target rather than an achieved number.
On Monday 28 September 2026, Roche held its Pharma Day investor presentation. The research session — titled “Revolutionizing drug R&D with Lab-in-a-Loop” — was delivered by Aviv Regev, Head of Genentech Research and Early Development. The full slide deck is publicly available. Three limits apply before anything else: this is an investor-day strategy document, not a product launch, not a published result, and not a regulatory filing. Nothing in the deck says Roche has deployed an autonomous laboratory.
Page 94, under the slide title “Towards autonomous labs for drug R&D” and the subtitle “Our Autonomous Lab Accelerator to enhance the loop,” sets out a six-level ladder. In the deck’s own words, running from left to right: L0 is Fully manual, L1 is Automated instruments, L2 is Connected workflows, L3 is AI-assisted workflows, L4 is Domain autonomy, and L5 is Lab-wide autonomy. The slide cites a supporting paper — “Cheng et al. Preprints.org 2026” — which is marked “(in review),” meaning it has not been peer-reviewed. The slide also shows a row of partner logos above the phrase “Including collaborations with”; those were images and were not read by the text layer, so no collaborator is named here in connection with that slide. No ladder level is attributed to Roche’s current position anywhere below — this reading comes from the text layer, which does not capture graphical marks, and nothing here asserts the deck omitted such a marking.
The same page includes a second row that runs parallel to the six levels. It reads, in the deck’s own wording: Human role — Operator, Assistant, Approver, Auditor. That four-word sequence is where the structural weight sits, and it is the subject of the analysis below.
The key insight: The human-role row is not a list of jobs getting smaller — it is a list of jobs changing kind. Operator, Assistant, Approver, Auditor describes a shift from execution to accountability, and the last two roles carry a different failure mode and a different liability profile than the first two. That reading comes from the ladder’s own wording and nothing more. No employment effect is predicted here.
The headline figure counts decisions that had a contribution. The same slide grades how much those contributions mattered, which is the more useful disclosure.
The Numbers Slide Grades Itself
Page 73 is where the deck earns credibility, and it does so by disclosing its own grading system. The headline is that 40 per cent of pipeline decisions from Q4 2025 through Q2 2026 had a tracked AI or computational contribution, with coverage reaching 44 per cent in Q2 alone. The slide then breaks those contributions into three tiers: Fundamental or Critical at 41 per cent of contributions, Supportive at 48 per cent, and Informative at 10 per cent. The deck defines its own terms — Critical means “Materially influenced,” Supportive means “Meaningfully impacted decision confidence,” and Informative means “Informed thinking.”
Multiplying the 40 per cent coverage figure by the 41 per cent Fundamental-or-Critical share yields roughly 16 per cent of pipeline decisions carrying what the company itself classes as a fundamental or critical AI contribution. That multiplication is this publication’s own, it is approximate, and the deck does not present it. The fair reading cuts in Roche’s favour: publishing the role mix at all is the opposite of overclaiming. Most companies would have stopped at forty per cent and let the reader fill in the rest. Nothing here calls the 40 per cent misleading or inflated — the disclosure of the grading breakdown is precisely the credible part.
AI Contribution Role Mix — Q4 ’25 through Q2 ’26 (Roche disclosed)
Fundamental or Critical (“Materially influenced”)41%
The six-level ladder is borrowed machinery, and the choice to borrow it is itself the signal. The L0-to-L5 taxonomy is recognisably the shape of the SAE driving-automation levels, transposed onto a laboratory, wrapped around a specific scientific loop the slide spells out: Observe, Learn, Predict, Design, Execute, Analyze, Results, Report. The mapping to autonomous vehicles is not coincidental — it imports a shared language for discussing partial autonomy, handoffs, and accountability that regulators, boards, and engineers already understand.
What the slide does that most automation roadmaps leave implicit is write the human role down. In the deck’s own words: Operator, Assistant, Approver, Auditor. Read left to right across the six levels, the human goes from doing the work to helping with it to signing off on it to checking it after the fact. That is not compression — it is role transformation. Approver and Auditor are accountability functions, not execution functions. They require different training, generate different paper trails, carry different legal exposure, and fail in different ways than Operator or Assistant roles do.
The Map of AI framework is the right lens here. Roche is not building foundation models — it is operating in the integration and application layers, assembling an AI-native research process on top of external model infrastructure. Page 71’s journey slide lists partner-or-acquisition items including an AI factory (NVIDIA), an AI partnership pilot (OpenAI), and an MHS arrangement (Anthropic). Page 100, titled “AI independence: owning the loop, learnings, improvements,” makes this architecture explicit: independence means owning the data and the iteration loop, not disengaging from those infrastructure suppliers. Nothing here says Roche is moving away from, replacing, or reducing reliance on any named AI company. The general pattern recurs across comparable decks: own the thing that compounds — the proprietary data and the feedback loop — while renting the thing that depreciates, which is raw model compute.
Map of AI — Integration Layer
Own the Loop, Rent the Compute
The durable competitive position in applied AI is not in model weights — it is in proprietary data and the feedback loops that improve on it. Roche’s “AI independence” framing is a statement about where the compounding value sits: inside the Lab-in-a-Loop iteration cycle, not inside any single model provider. Infrastructure is rented; the scientific memory that accrues through the loop is owned.
One further figure from the deck, stated precisely as what it is. TargetNexus — described on page 80 as the company’s AI agent for target assessment — is “on track to be part of 80% of research portfolio decisions by Q4.” That is a company target, not an achieved number, and nothing here predicts whether it is met.
The scoping is the structurally interesting detail. Target assessment is the earliest stage of the loop the deck describes — the point at which a wrong decision is cheapest to make and most expensive to keep. Concentrating an AI agent at that stage is a leverage decision: shifting probability distributions on target selection compresses error cost earlier in the pipeline than any downstream intervention could.
Roche Pharma Day 2026 — Page 80
“TargetNexus on track to be part of 80% of research portfolio decisions by Q4.”
Three Implications
THE ACCOUNTABILITY SHIFT IS THE REGULATORY STORY
When the human role moves from Operator to Auditor, the governance question changes completely. Auditing an autonomous system requires different institutional competencies, different documentation standards, and different liability frameworks than operating one. Drug regulatory agencies have not yet standardised on how to treat AI-contributed pipeline decisions. The ladder Roche published is a roadmap — and regulators will eventually need one of their own to match it.
GRADED DISCLOSURE SETS A NEW BENCHMARK
Publishing a 40 per cent AI-contribution figure alongside a breakdown of what “contribution” actually means — Fundamental/Critical versus Supportive versus Informative — is materially more rigorous than the headline-number disclosures that have become standard investor-day practice. If this format propagates to peers, it raises the floor for what AI impact claims have to demonstrate, and it makes comparability possible for the first time.
FRONT-OF-LOOP CONCENTRATION IS THE REAL ARCHITECTURE CHOICE
Deploying TargetNexus at target assessment — the earliest and highest-leverage decision point in the pipeline — rather than at later, more visible stages signals a deliberate architectural preference. The compounding effect of better target selection is larger and slower to attribute than a model that accelerates a late-stage assay. That makes it strategically harder to copy and harder for competitors to benchmark against.
Everything above is drawn from a single investor-day presentation. It is a strategy document, not a product launch, a published result or a regulatory filing, and nothing above says Roche has deployed an autonomous laboratory.The paper the autonomy slide cites is marked “in review”, which means it has not been peer-reviewed.No ladder level is attributed to Roche above. This reading was taken from the document’s text layer, which does not capture graphical marks, so nothing above claims a current position on the ladder and nothing above claims the deck failed to indicate one. The collaborators shown on that slide appear as logos and were not read, so none is named. The figure of roughly 16 per cent is this publication’s own approximate multiplication of two figures the deck discloses separately — 41 per cent of 40 per cent — and the deck does not present that product.The 80 per cent attached to TargetNexus is a stated company target rather than an achieved number, and nothing above predicts whether it is reached. “AI independence” in this deck refers to owning the data and the iteration loop. The same slide carries a Partnerships column, and the journey slide names partner or acquisition items including MHS (Anthropic), an AI factory (NVIDIA) and an OpenAI pilot partnership — nothing above says Roche is moving away from, replacing, or reducing its reliance on any of them.The reading of the Operator, Assistant, Approver and Auditor sequence is a reading of the slide’s own wording. Nothing above says jobs are lost or scientists replaced, and no employment effect is predicted. Any count of autonomous laboratories, go-live date, cost, capex or savings figure, headcount effect, attributable molecule, and the phase-level percentages on the same slide are not established and do not appear — a limit of this reporting rather than evidence that none exist. Nothing above predicts Roche, drug discovery, laboratory automation or anyone’s employment.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
