**By Lily Caruso · Wednesday, September 2, 2026** --- Sit down with me, because this is one of those stories where the number everyone quotes is accurate and the conclusion everyone draws from it is not. Here's the number. In 2024, a team at Boston Consulting Group, working with the Wellcome Trust, did something the AI drug-discovery industry had somehow never done for itself: they counted.
They gathered the AI-discovered molecules that had actually entered human trials and asked how they fared. The answer, published in *Drug Discovery Today*, was startling: **AI-discovered drugs succeeded in Phase 1 trials 80 to 90 percent of the time, against a historical industry rate of roughly 40 to 65 percent** (other tallies put the benchmark at 50–60; either way the gap is enormous). Nearly double.
That statistic has since appeared in more investor decks than I can count, usually next to a hockey-stick chart. Now here's the number that appears on the same page of the same paper and almost never makes the deck: **in Phase 2, the AI molecules succeeded about 40 percent of the time — exactly in line with the historical industry average.** A 2025 *Nature Medicine* analysis reached the same conclusion in plainer words: AI-discovered drugs "have experienced similar levels of phase 2 trial failure as non-AI-discovered drugs." The machines doubled the pass rate on the first exam and changed nothing on the second. To understand why that's the whole story, you have to understand what each exam tests. **The chemistry test and the biology test.** Phase 1 is, mostly, a safety and pharmacology exam.
Is the molecule tolerable at useful doses? Does the body absorb it, distribute it, clear it the way the models predicted? Does it behave like a drug?
These are *chemistry* questions — questions about the molecule — and chemistry is exactly what generative models are built to optimize. Give an algorithm a target and a hundred million candidate structures, and it will hand you a molecule with cleaner pharmacokinetics, fewer off-target liabilities, and better drug-likeness than a medicinal chemist could iterate to in the same year. Insilico says its platform now reaches preclinical-candidate nomination in 12 to 18 months on average, a process that historically took four to six years.
So of course AI molecules ace Phase 1. They were designed by a system whose objective function *is* Phase 1. Phase 2 is a different exam.
It asks whether the *target* was right — whether hitting this protein, in these patients, actually changes the disease. That's a *biology* question, and the honest state of the art is that nobody, human or machine, has cracked it. The industry's Phase 2 failure rate has hovered around 60 percent for decades not because the molecules were bad but because our understanding of disease is.
An AI can build a perfect key faster than anyone alive; it cannot yet tell you which lock opens the door. The BCG authors said it themselves, in the polite dialect of consultants: the persistent bottleneck is "target validation and disease biology — not molecular optimization." Read that as: **the part AI fixed was never the expensive part.** **What the funnel looks like in 2026, counted.** A peer-reviewed analysis presented through ASCO this summer tracked **117 AI-enabled therapeutic assets across 63 companies** that had entered interventional human trials. Of those, **60 — 51 percent — had completed Phase 1.
Eight — 6.8 percent — had completed Phase 2.** Median time from a company's founding to its first Phase 1 entry: six and a half years, which is not the timeline the word "AI" implies. Broader tallies count more than 173 AI-designed programs in clinical development and roughly **$60 billion invested since 2019**. And the number at the bottom of the funnel, as of the end of July: **zero FDA approvals.** Not "few." Zero.
Rentosertib, Insilico's drug for idiopathic pulmonary fibrosis, is the furthest along, and it registered its Phase 3 in July. **The one that got to Phase 3, with the humility clause attached.** Rentosertib deserves its headlines, and it also deserves the fine print the desk attaches to every biotech catalyst. It's a TNIK inhibitor — a target Insilico's software nominated, for a molecule its software designed — and the Phase 2a data, published in *Nature Medicine* in 2025, was striking: in the GENESIS-IPF study, 71 patients across 22 sites, the 60-milligram arm showed a **mean forced-vital-capacity change of +98.4 milliliters at twelve weeks against −20.3 milliliters on placebo**, with adverse-event rates similar across arms and a dose-dependent trend. In a disease where the approved standard-of-care drugs *slow* the decline rather than reverse it, a positive FVC number is the kind of thing that makes pulmonologists put down their coffee.
The Phase 3, registered July 7: 320 patients, 47 centers, 52 weeks, primary endpoint the annual rate of FVC decline. Now the clause. Twelve weeks is not fifty-two.
Seventy-one patients is not three hundred twenty. And all 47 centers are in China, which means the U.S. regulatory path will want its own evidence. This is the house rule from our catalyst-ladder coverage, unchanged: *biotech's only currency is data*, and the data that counts hasn't been generated yet.
The readout is a 2027–28 event. **The names, sorted honestly.** The pitch decks show the wins; the desk shows the whole board. *Recursion* — which absorbed Exscientia in a $688 million all-stock deal — has genuine Phase 2 signals (REC-4881 cut polyp counts 43–53 percent in familial adenomatous polyposis; REC-617 posted a confirmed partial response in platinum-resistant ovarian cancer) and, in May 2025, discontinued three programs *despite* some of them showing positive signals, which is what portfolio triage looks like when the capital markets stop paying for optionality. *Schrödinger's* SGR-1505 posted a 22 percent response rate in B-cell lymphoma; its SGR-2921 was discontinued last August after two treatment-related deaths — a reminder that the 80–90 percent Phase 1 statistic is an average, not a guarantee, and that "safe by design" has a body count too when it misses. *Relay's* zovegalisib showed an 11-month median progression-free survival in PI3K-alpha-mutated breast cancer. *Generate:Biomedicines* has an anti-TSLP antibody in a 1,600-patient Phase 3 for severe asthma. And *Isomorphic Labs* — the DeepMind spinout, the purest bet on the biology problem rather than the chemistry problem — raised a record **$2.1 billion** in May, has **no disclosed clinical candidate**, and has pushed its first trials from end-2025 to end-2026. The most-funded company in the field has not yet taken the first exam.
And the ones the decks skip: Exscientia's DSP-1181, the first AI-designed drug to enter humans in January 2020, quietly abandoned two years later; BenevolentAI's BEN-2293, which missed its efficacy endpoints in 2023 and cost 180 people their jobs; BergenBio's bemcentinib, discontinued last June with no responses among ten evaluable lung-cancer patients. Each of them, note, was a Phase 2 story. The chemistry worked.
The biology didn't. **What the regulator thinks, which matters more than what the models think.** The FDA's January 2025 draft guidance acknowledged more than 500 submissions with AI components between 2016 and 2023 and said the only thing that matters: approval "depends on demonstrated clinical safety and efficacy, not discovery method." In January of this year the FDA and EMA jointly issued ten principles for AI across the drug lifecycle. And in April 2025 the agency announced it would move away from animal testing toward "new approach methodologies" — computational models and organoids — which is the single most consequential policy tailwind the field has, because it moves the *regulator's* evidence standard toward the tools the field is built on. None of it lowers the bar at Phase 2.
All of it lowers the cost of getting there. **So what is the 80–90 percent actually worth? Do the arithmetic the decks won't.** Suppose a conventional program costs X to reach a Phase 2 readout and passes Phase 1 half the time. Suppose an AI program reaches the same readout for a third of X and passes Phase 1 85 percent of the time — and then faces the *same* 40 percent Phase 2 odds.
The expected cost per Phase 2 success falls by something like 80 percent. That is a real, large, economic advantage — **a productivity story for the front of the funnel, not a hit-rate story for the whole funnel.** It changes who can afford to take shots. It does not change how many shots score.
The companies that win under those economics are the ones with the most cheap shots and the best biology to aim them with, which is why Recursion's "maps of biology" and Isomorphic's structure-prediction lineage are the bets that matter, and why a company whose only edge is faster chemistry is selling a commodity that big pharma's own platforms will match. **The desk's rows — placed in public, graded on data.** *One:* **65% that the Phase 2 success rate for AI-discovered molecules remains within five percentage points of the industry base rate through 2027** — the biology wall holds until something in the field demonstrably improves target selection; falsifier is a peer-reviewed cohort update showing a Phase 2 rate above 50 percent. *Two:* **60% that the first FDA approval of an AI-discovered drug lands no earlier than 2028** — rentosertib's Phase 3 is a 52-week study registered in July with China-only sites, and every other program is behind it; falsifier is an accelerated or conditional pathway for rentosertib or a surprise from the Generate asthma program. *Three, a watch, not a row:* Isomorphic's first clinical candidate by year-end — if the best-funded biology bet in the field misses its second deadline, the market will start pricing the biology problem the way this desk does. **The honest fine print.** Success-rate figures are per the 2024 BCG/Wellcome analysis in *Drug Discovery Today* and the 2025 *Nature Medicine* commentary, as cited; the historical Phase 1 benchmark varies by source (40–65 percent in the paper's framing, 50–60 in others) and both are carried. The 117-asset funnel, the 173-program count, the $60 billion figure, and the zero-approval status are per the cited July 2026 pipeline review of the ASCO/JCO analysis. Rentosertib's Phase 2a and Phase 3 details are per Insilico's own release and the *Nature Medicine* publication; the standard-of-care comparison is characterized loosely and is not a head-to-head claim.
Program outcomes for Recursion, Schrödinger, Relay, Generate, Isomorphic, Exscientia, BenevolentAI, and BergenBio are per the cited pipeline review. The expected-cost arithmetic is this desk's own illustration with assumed inputs, labeled as such. Every number carries its stamp.
One image before I let you go. Picture a school that suddenly produces a class of students who score in the ninety-ninth percentile on the entrance exam — and then perform exactly like everyone else once they're inside. You'd be right to be impressed by the entrance exam.
You'd be wrong to predict the graduation rate from it. The AI drug-discovery industry has spent $60 billion building the best entrance-exam students in the history of medicine, and the graduation rate has not moved. The scientists know this; it's in their own paper.
The question for the next two years isn't whether the machines can design a molecule. It's whether anyone — Recursion with its maps, Isomorphic with its structures, Insilico with its 320 patients in 47 Chinese clinics — can teach them which door to open. The chemistry is solved.
The biology is the business. The first diploma is a 2028 story. — Lily **Tickers in play:** RXRX · SDGR · RLAY · ABCL · XBI · LLY · NVO · SNY · NVDA · (Isomorphic, Insilico, Generate: private or non-U.S. listed) --- *This is TrendyVest's analysis and opinion — for informational purposes only, not investment advice or a recommendation to buy or sell any security or commodity. Sources: "How successful are AI-discovered drugs in clinical trials?
A first analysis and emerging lessons," Jayatunga et al. (Boston Consulting Group, with the Wellcome Trust), Drug Discovery Today, 2024 — the 80–90% Phase 1 and ~40% Phase 2 findings and the target-validation interpretation, as summarized in the cited narrative review; the 2025 Nature Medicine commentary on Phase 2 failure parity, as quoted in the cited pipeline review; the ASCO/JCO analysis of 117 AI-enabled assets across 63 companies (51.3% completing Phase 1, 6.8% completing Phase 2, 6.5-year median founding-to-Phase-1), the 173-program and ~$60B tallies, the zero-approval status as of July 31, 2026, and the program-level outcomes (Recursion's REC-4881, REC-617, and May 2025 discontinuations and the $688M Exscientia merger; Schrödinger's SGR-1505 and SGR-2921; Relay's RLY-2608; Generate's GB-0895 SOLAIRIA program; Isomorphic's $2.1B Series B and trial timing; Exscientia's DSP-1181; BenevolentAI's BEN-2293; BergenBio's bemcentinib) per IntuitionLabs' July 31, 2026 pipeline review; rentosertib's GENESIS-IPF Phase 2a (71 patients, 22 sites, 60 mg, +98.4 mL vs. −20.3 mL FVC at 12 weeks, similar adverse-event rates) per the Nature Medicine publication as cited and the Phase 3 design (320 patients, 47 centers in China, 52 weeks, annual FVC decline primary) and the 12–18-month PCC claim per Insilico's July 7, 2026 release; ISM6331 Fast Track per the cited review; the FDA's January 2025 draft guidance (500+ AI submissions), the April 2025 animal-testing transition, and the January 2026 FDA–EMA principles per the cited narrative review; market-size estimates per the cited review; the two rows and the watch item logged to the master tracker. Every number carries its stamp. Do your own research.* *Markets.
Tech. The Edge. Research with receipts.*