Cool Does Not Scale: Novartis Says What Every Pharma Steering Committee Needs To Hear
Novartis's research data chief describes the Pilot Trap from the inside. A week later, Roche, its neighbour in Basel, published the first metric I have seen for what getting out of it looks like.
By Mark Rhys Thomas
The Lead
Christian Diehl runs data and digital for biomedical research at Novartis. In a McKinsey interview published on 21 September, he described a meeting most people reading this have sat in.
An impressive demo. A steering committee that likes what it sees. Six months later, someone checks whether it was rolled out, industrialised, made generally available. Usually it was not.
His summary was three words. Cool does not scale.
That is the Pilot Trap, described by the person at one of the world's largest drugmakers whose job is to get out of it. Issue 01 of this newsletter put a number on the trap. This issue has the operator's account of how it forms, and, from a rival headquartered in the same city, a public metric for what the exit looks like.
Diehl's account has three parts. Each one is something you can check in your own organisation this week.
The demo is built for the wrong audience. A demo exists to win a steering committee. A production system exists to change how people work on an ordinary Tuesday. Diehl's word for what scales is "boring": AI embedded in a process people already run, until it is simply the way things are done. His example is not a moonshot. It is hit-finding and molecule optimisation in generative chemistry, a core process that exists with or without AI, rebuilt around it. He says he would rather have AI embedded in real processes than many proofs of concept.
The data work is longer than anyone budgets for. Novartis's data42 platform combines more than 20 years of clinical trials with animal toxicology and real-world data. Diehl calls the investment in cleaning, curating and harmonising it to be AI-ready "huge", and says it has been running for many years. The platform went through several technology pivots before it showed a return. One curated data set found no real takers, while scientists asked for others nobody had touched. His lesson was to start from the scientist's question and back-solve the data product. That is last week's issue seen from inside a research organisation. The money goes where the data is clean. The value goes where someone has asked a question.
Scale is a governance decision before it is a technical one. Diehl wants agents governed "almost like associates", with explicit rights to do some things and not others. He names three positions for the human: a circuit breaker who must make the decision, an observer who intervenes when needed, and no human at all where a process is validated and fully automated. He also wants an audit trail replicable enough to defend to a health authority. That is the Permission and Proof layers, specified by a practitioner.
Now the contrast. On 28 September, at Roche's Pharma Day in London, Aviv Regev, who heads Genentech Research and Early Development, gave a number. Over the past three full quarters, AI and computation contributed to 40 percent of Roche's portfolio decisions. Roche grades each contribution. Critical means the AI output was material or decisive. Supportive means it meaningfully changed confidence in the decision. Informative means relevant but peripheral. About 40 percent of the contributions were critical. In lead optimisation, almost 70 percent of decisions carried one.
Read the measure, not just the number. Roche does not count pilots, use cases or tools. It counts decisions in its portfolio process and asks whether AI moved them, and how much. A demo cannot score on that metric, because a demo never sits inside a portfolio decision. That is what stage three of four looks like: AI measured as a share of how the business decides, not as a list of things it has tried.
Roche's deck also writes down the human role as the lab moves towards autonomy: operator, assistant, approver, auditor. Diehl's circuit breaker and Roche's approver are the same control, arrived at separately. Both companies also use the same phrase for the architecture, lab in the loop, in which models predict, the lab tests, and the results retrain the models.
The rebuttal writes itself. These are two of the richest research organisations in the world, and this is drug discovery, not commercial operations. Both points are fair. Roche's figures come from an investor day. They are self-reported and unaudited, and I would not quote the 40 percent as an industry benchmark. Nor am I attributing Roche's improved pipeline results to AI, because Roche has not.
But the mechanism transfers. The commercial organisations I have worked inside run a version of the pattern Diehl describes: a pilot per brand, a demo per quarter, a data set curated before anyone asked for it. McKinsey's survey this year puts AI high performers at 6 percent of companies. The trait that most clearly separates them is that nearly three quarters have fundamentally redesigned workflows around AI, against one quarter of everyone else. That is "make it boring", with a sample size.
So here is the test for a steering committee. Stop asking how many AI use cases are live. Ask what share of this quarter's decisions in one core process carried an AI contribution, and how many of those were critical. If nobody in the room can answer, the organisation is still running demos.
Count decisions, not demos.
Also on the Line
COMPUTE
Your pilot price was never your run rate. McKinsey's public sector team describes US agencies coming off introductory AI deals, some priced at $1 per agency, onto consumption billing. Its framing applies well beyond government: the cost of a fixed level of AI capability is falling roughly tenfold a year, yet total bills rise, because agentic workflows use five to 30 times the tokens of a chat query and usage grows faster than prices fall. The recommendations are short contracts, exit clauses, and tracking cost per useful outcome rather than token volume. Pharma's version is the enterprise copilot bought on introductory terms. The parallel is mine, not McKinsey's.
DATA
AI-designed proteins now carry a signature. Google DeepMind released SynthID Bio on 30 September, a watermark embedded in the amino acid sequence and predicted structure of AI-designed proteins. In its lab tests, watermarked binders against PD-L1, VEGF-A and a coronavirus protein domain performed in line with unwatermarked versions. The stated purpose is biosecurity: giving DNA synthesis screeners a way to tell where a design came from. The pharma implication is provenance. As more candidate molecules start life in a model, where an input came from becomes part of the record. This is vendor research about its own tool.
PROCESS
Confidence up sixteen points, controls down thirteen. KPMG's US AI pulse (314 leaders, fielded July and August) found confidence in AI governance rose from 57 to 73 percent in one quarter. Over two quarters, the share building controls into their agents fell from 43 to 30 percent, as firms leaned on guardrails already in place. Multi-agent development jumped from 6 to 25 percent. The constructive finding: 49 percent have written down high-risk cases where agents may not decide alone. KPMG's separate global sample adds that only 12 percent consistently assess AI's value against its cost.
What Did Not Meet Spec
Thirty-one newsletter emails came through this week's pull, across five consultancies and two TLDR editions, plus three research hubs checked directly and Brightbeam's SNI roundup read on the web. Sixteen emails were read in full, all ten TLDR issues among them, and eleven sources were read at primary, including a Roche investor-day transcript and McKinsey's full survey text. Rejected by category: about thirty sponsored placements across the TLDR editions; rocketry, robotaxis, consumer hardware and developer-culture essays from the general edition; McKinsey digests on energy, beauty, sport and healthy living; and EY's only email, a repeat of one already discarded last week.
Three named items.
A caveat lifted, on issues 02, 04 and 05. Issues 02 and 04 used McKinsey's figure that 37 percent of companies report AI contributing to EBIT, and issue 02 also used its 6 percent of high performers. Issue 05 suspended both, because a McKinsey newsletter described the 6 percent three different ways in one email. I have now read the survey text. The 37 percent is the share reporting any positive EBIT contribution from AI. The 6 percent is the share of high performers, defined as attributing at least 5 percent of EBIT to AI and calling the value significant. They measure different things and do not conflict. The newsletter's third description, 6 percent as the share of firms redesigning workflows, was wrong: workflow redesign is a trait of the high performers, not their definition. Both figures, as issue 02 used them, stand. The filter runs backwards as well as forwards, and sometimes it clears a figure.
Roche's Phase III success rate. Reports of Roche's AI announcement also carried a Phase III success rate above 80 percent this year, against 65 percent in 2025, along with Roche's claims of a 2.5-fold rise in early pipeline entries on flat resources. Roche presented these at the same event. It did not attribute the Phase III figure to AI, and one year of late-stage outcomes cannot carry that weight. Left out of the lead on purpose.
McKinsey's 25,000 agents. In the Novartis interview, McKinsey's host says the firm has 60,000 people, of whom 25,000 are agents. It is a striking line in a McKinsey interview, about McKinsey, with no definition of what counts as an agent. It is not evidence of anything, and it is not in the lead.
The Line
Novartis says cool does not scale. Roche shows what does: a count of the decisions AI actually moved.
Sources
- 1. McKinsey & Company, Life Sciences Practice, "Novartis's Christian Diehl on scaling AI beyond the demo", Eureka! podcast, edited transcript, 21 September 2026. Hosts Navraj Nagra and Valentina Sartori. Full transcript read. Source of "cool does not scale", the "boring" prescription, the data42 description and history, the "no real takers" data set, governing agents "almost like associates", the three human positions, and the audit-trail requirement. Interviewee opinions are his own. The 25,000 agents remark is the McKinsey host's and is rejected in this issue.
- 2. Roche Holding, Pharma Day 2026, London, 28 September 2026. Research session "Revolutionizing drug R&D with Lab-in-a-Loop", Aviv Regev. Transcript read (StockAnalysis). Source of the 40 percent of decisions over the past three full quarters, the critical, supportive and informative definitions, about 40 percent of contributions critical, and almost 70 percent of lead optimisation decisions. The operator, assistant, approver, auditor row is from the published slide deck, page 94, as reported by FourWeekMBA, 2026; the deck was not read directly. Company presentation to investors, self-reported, unaudited. Roche's own reporting of the Phase III rate is rejected as an AI claim in this issue.
- 3. Reuters, "Roche outlines plans to move towards autonomous AI labs", 28 September 2026, via Euronext. Corroborates the 40 percent and the 80 percent Target Nexus target, which is a company target, not an achieved figure, and is not used in the lead.
- 4. McKinsey & Company, QuantumBlack, "The state of AI in 2026: On the road to ROI", Tinkoff, Van der Veken, Chui, Balakrishnan, 25 August 2026. Online survey, 4 May to 8 June 2026, 1,719 participants in 97 nations, weighted by each nation's share of global GDP. Survey text read. Source of the 37 percent (any positive EBIT contribution), the 6 percent (high performers, at least 5 percent of EBIT and significant value), and nearly three quarters of high performers fundamentally redesigning workflows against one quarter of others.
- 5. McKinsey, The CEO Shortlist, "From productivity to returns", 18 September 2026. The email whose three descriptions of the 6 percent prompted the caveat in issue 05, now resolved.
- 6. McKinsey & Company, Public Sector Practice, "The next phase of government AI economics", Spencer, Shenai, Bailey, Mandel, Martin, 29 September 2026. Full article read. Source of the $1 introductory pricing, the roughly tenfold annual fall in the cost of fixed capability (the article's footnote 2; originator not traced), five to 30 times the tokens for agentic workflows, and the contract recommendations. The pharma parallel is my inference.
- 7. Google DeepMind, "Introducing SynthID Bio", 30 September 2026, Google blog. Primary read. Source of the watermarking method, the three binder targets and the biosecurity purpose. Vendor research about its own tool; the provenance implication for pharma is mine.
- 8. KPMG LLP, Quarterly AI Pulse Survey, Q3 2026, United States, press release, 24 September 2026. 314 US C-suite and business leaders at organisations with $1 billion or more in revenue, fielded 24 July to 25 August 2026. Primary read. Source of 73 and 57 percent confidence, 30 and 43 percent building controls into agents, 25 and 6 percent multi-agent, and 49 percent defining no-autonomy cases.
- 9. KPMG International, Global AI Pulse Survey, Q3 2026, press release, 24 September 2026. 2,131 senior leaders across 20 countries. Primary read. Source of the 12 percent consistently assessing AI value against cost. A different sample from source 8; the two are not combined.
- 10. Batch Record issue 02, 31 August 2026, issue 04, 14 September 2026, and issue 05, 21 September 2026, for the 37 and 6 percent figures and their caveat.
Applying Novartis's and Roche's research experience to pharma commercial operations is my inference, not either company's claim, as is the reading of Roche's metric as stage three of a four-stage model.
The list
Subscribe to Batch Record
Weekly, every Monday. One issue, one layer, one line worth quoting.