Pharma Put Its AI Money Where the Data Was, Not Where the Trials Fail
Big pharma signed $29.5 billion of AI drug discovery deals last year and paid 2 percent of it up front. The market has already worked out where the value is. Most research budgets have not.
By Mark Rhys Thomas
The Lead
Big pharma put its name to $29.5 billion of AI drug discovery deals in 2025. It paid $0.6 billion of that up front. Two percent.
The rest, 60 to 90 percent of contract value depending on the deal, is contingent on clinical and commercial milestones that have mostly not arrived. McKinsey looked at 38 of these deals signed in 2023 and 2024. By June 2026, less than 1 percent of their maximum milestone value had been paid out.
That is not a market that has decided AI drug discovery does not work. It is a market that has decided exactly where the proof will come from, and has written the contracts accordingly.
McKinsey's new piece on research AI, published on 21 September, explains why. Investor funding in AI-enabled drug discovery reached $8.4 billion in 2025, up from $4.1 billion in 2023. Most of it went to molecule design, hit identification and lead optimisation, because that is where the data is structured, the assay feedback is fast and the models perform reliably. McKinsey rates that stage of the value chain high maturity.
Then it rates the other two. Target identification: medium. Preclinical and translational research: low.
The trouble is that the trials fail in those two. Get the target wrong, the biomarker wrong or the patient population wrong, and a beautifully optimised molecule still fails. Roughly a quarter of Phase II and Phase III failures come from safety problems nobody predicted. McKinsey's own phrase for the mistake is that maturity and importance are not the same thing. The capital followed the tools, not the bottleneck.
Issue 05 closed on a small model trained on measured laboratory data beating two frontier models at lower cost. This is the same finding seen from the funding side. The binding constraint in pharma AI is not the model. It is whether the data exists to train it on the question that matters.
That is why this issue files to Data, the first Batch Record to do so. McKinsey's own recommendation is a data recommendation dressed as a research one: treat target identification as a data generation strategy, move from correlative data to mechanistic evidence, convert unstructured proprietary clinical data into training-ready assets, and fund the linkage between preclinical and clinical results. None of that is a model purchase. All of it is a supply chain for inputs, with provenance, quality and lineage, which is what the Data layer is.
It is also the Pilot Trap, arriving in research rather than in commercial operations. Pilots cluster where success is easy to demonstrate. Molecule design demos well: you can show a binding prediction and a cycle time. Target validation does not demo at all. It takes live cell experiments, multiyear readouts and evidence that one gene causes the disease rather than keeping company with it. So the pilots go where the demos are, the demos go where the data is clean, and the programmes keep dying where the data is not.
I have seen the same instinct in commercial operations. The first automation budget goes to the step that is easiest to show a board, and the delay sits somewhere else entirely.
The obvious rebuttal is that the early results are good. They are, on the measure they are built for. McKinsey reports selected AI programmes reaching first-in-human 15 to 80 percent faster than large-pharma medians. I am not relying on that number, and the discard note below says why. But even taken at face value it measures speed to the start of the clinic, not survival through it. McKinsey says plainly that Phase III efficacy readouts for AI-originated molecules remain years away. The deal terms say the same thing in money. A 2 percent down payment is the industry pricing the claim until the clinic prices it properly.
One more signal, from outside the consultancies. A Stanford team published in Science on 17 September a "virtual biotech" of up to 37,000 agents that catalogued around 50,000 historical trials. Its headline finding was not about molecules at all. It was about which targets succeed. Details are in Also on the Line.
So the question for a head of research is not whether AI works. It is what share of this year's AI budget is aimed at the stage where your programmes actually fail, and how much of it is buying better answers to questions that were already the easy ones. If the answer is mostly molecule design, the portfolio is optimising the step it can see and funding the failure it cannot.
The money went where the data was clean. The trials fail where it is not.
Also on the Line
DATA
Stanford ran a biotech on 37,000 agents, and its best finding was about targets. Researchers led by James Zou built a virtual biotech in which agents, coordinated by a virtual chief science officer, catalogued roughly 50,000 clinical trials in under a week. Drugs aimed at targets concentrated in one cell type, with switch-like rather than dimmer-like gene activity, were 40 percent more likely to move from Phase 1 to Phase 2, 48 percent more likely to reach market and had 32 percent fewer adverse events. These are associations in historical data, not causes, and the team says every prediction still needs the lab. But it is target evidence, which is exactly the input the lead says is missing.
COMPUTE
Your AI supplier may be building the lab next door. Anthropic announced that roughly 950 Claude agents, 21 hours and 210 million tokens had found an uncharacterised enzyme system with CRISPR-like repeats, in its own new biology lab. Scientists were less impressed: Johns Hopkins' Steven Salzberg called the announcement attention-seeking, and Harvard's Philip Kranzusch called it an exciting beginning rather than a breakthrough. The more useful question came from the investor Nic Carter, who labelled it speculation himself: if model margins keep shrinking, a lab has good reason to keep discoveries rather than sell tools. For pharma, choosing a model provider may also mean choosing a competitor in discovery.
PROCESS
Two thirds of biopharma plants are still exploring. McKinsey's companion piece on manufacturing says about two thirds of manufacturers remain in exploration or targeted AI implementation, and names the reason: pilots are built onto existing handoffs, documents and spreadsheets, so they improve a step without changing the system. Productivity rose about 25 percent from 2014 to 2018 and has fallen 5 percent since. The worked example to steal is deviation management, where agents assemble the evidence and QA stays accountable at four named decision points: classification, root cause, the corrective action plan and closure. Research and manufacturing, same trap, same week, same firm.
What Did Not Meet Spec
Thirty-five 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. Twenty-two emails were read in full, all ten TLDR issues among them, and thirteen sources were read at primary. Rejected by category: every sponsored placement in the TLDR editions, eighteen advertisers in all; the general edition's rocketry, robotaxis, consumer hardware and career essays; model tooling with no strategic signal; and every Accenture hub item, because none of it was newer than May.
Three named rejections.
**The 15 to 80 percent, from this issue's own lead source.** McKinsey reports selected AI programmes reaching first-in-human 15 to 80 percent faster than large-pharma medians. The sample is programmes that chose to disclose, with sufficient self-reported data. A range that wide, from a self-selected sample, is a marketing number whatever the intent. I have cited the article for its deal data and its maturity ratings and left this figure out.
**A caveat on issue 05.** Issue 05 reported Anthropic's claim of a fourfold speed-up across more than thirty biomolecular models and said to treat the multiple with care. This week the same lab's enzyme announcement drew public criticism from working scientists for overselling a preliminary result. That does not make the fourfold figure wrong. It does mean a vendor's claims about its own scientific output should now carry an independent reading before I lean on them, and issue 05's item should be read with that attached.
**The two 93 percents.** Deloitte reported this month that 93 percent of organisational AI investment goes to technology and 7 percent to people. McKinsey is promoting, again, its finding that 93 percent of organisations exceed their AI budgets. Same number, unrelated measures, and the McKinsey figure rests on 75 respondents. The Deloitte split is interesting and its underlying study is not yet traced. Neither goes in until it is.
The Line
Pharma paid 2 percent up front for AI drug discovery. That is the market telling you where the data is missing.
Sources
- 1. McKinsey & Company, Life Sciences Practice, "AI drug discovery: Focusing on what matters most", 21 September 2026. Alex Devereson, Erika Stanzl, Guang Yang, Alex Peluffo, Alex Zhao, Mai Wang, Xin Wang. Full article read. Source of the $8.4 billion and $4.1 billion funding figures and 43 percent CAGR (the article's footnote 1 cites the underlying dataset, not traced), the $29.5 billion headline deal value and $0.6 billion up-front payments for top 20 pharma in 2025, the 60 to 90 percent milestone-contingent share, the fewer-than-1-percent payout across 38 deals signed 2023 to 2024 as of June 2026, the high, medium and low maturity ratings, the roughly one quarter of Phase II and III failures from unpredicted safety issues, and the recommendation to treat target identification as a data generation strategy. The 15 to 80 percent first-in-human figure is from the same article and is rejected in this issue.
- 2. Zou, Zhang et al., Stanford Medicine, "virtual biotech" study, published in Science, 17 September 2026. Figures taken from Business Today's report of 20 September 2026; the Science paper itself has not been read. Source of the 37,000 agents, roughly 50,000 trials, and the 40, 48 and 32 percent associations. Historical associations, not causal findings.
- 3. Anthropic, "Claude discovers a novel enzyme system with CRISPR-like repeats", 23 September 2026. Primary read. Vendor announcement of its own result; the agent, hour and token counts are Anthropic's own.
- 4. Bloomberg, Robert Langreth, "Anthropic biology discovery draws cautious notes from scientists", 24 September 2026. Paywalled; the Salzberg and Kranzusch remarks were confirmed through Bloomberg Law and syndicated coverage.
- 5. Nic Carter's remarks, as relayed by Nathaniel Whittemore on the AI Daily Brief podcast and summarised in Brightbeam's SNI Week 39, 26 September 2026. Not heard at source. Described by Carter himself as speculation.
- 6. McKinsey & Company, Life Sciences Practice, "Backbones and networks: A blueprint for the biopharma plant of the future", 15 September 2026. Full article read. Source of the two thirds figure (the article's footnote 1, underlying survey not traced), the 2014 to 2018 and post-2018 productivity figures, and the deviation management workflow. McKinsey co-founded the Lighthouse Network the article cites.
- 7. Deloitte Center for Integrated Research, "How will AI agents fit into the human workplace?", 10 September 2026. Primary read. The 93 percent and 7 percent split is attributed to "a recent Deloitte study", not traced. Rejected in this issue.
- 8. McKinsey, "The cost of intelligence: How CIOs can manage AI demand at scale", Enterprise AI FinOps Survey, fielded May 2026, 75 respondents, re-promoted in The Five Fifty, 23 September 2026. Rejected in this issue.
- 9. Anthropic, "How Claude is uplifting biomolecular modeling", 18 September 2026, as cited in issue 05. Caveated in this issue.
The pharma application of source 1 to the Data layer, and the link to the Pilot Trap, are mine rather than McKinsey's. McKinsey describes a capital allocation problem in research; the claim that it is at root a data supply problem, and the same mechanism as the Pilot Trap in commercial operations, is my inference.
The list
Subscribe to Batch Record
Weekly, every Monday. One issue, one layer, one line worth quoting.