AI and the Pharma Business Model── AI has punched a hole in Eroom's Law at the preclinical stage, but clinical trials and pricing regimes are slowing the transformation of the cost structure

01$2.2 billion per drug — pharma R&D has been losing efficiency for seventy years

Pharmaceutical R&D has a long-run trend that no other industry shares. In the 1950s, $1 billion of inflation-adjusted R&D spending yielded roughly 40 to 50 approved drugs. By the 2010s, that same billion yielded fewer than one. The number of approvals per billion dollars halves approximately every nine years. Jack Scannell and colleagues named this pattern "Eroom's Law" in 2012 — Moore's Law spelled backwards.

Deloitte's annual report, published in March 2025, estimated the average cost per approved drug across the top 20 companies at $2.2 billion in 2024. The internal rate of return on R&D rose to 5.9%, but this was driven by the revenue expansion of GLP-1 receptor agonists; excluding them, the figure drops to 3.8%. The pharma business model remains one in which a few blockbusters subsidize a large number of failures. Historically, about 90 percent of candidates that enter clinical trials never reach patients. The cost of those failures is folded into the price of the drugs that do succeed.

02Seventy percent of costs sit in clinical trials

Breaking down drug-development costs, preclinical work — target identification, compound optimization, and toxicity testing — accounts for roughly 20 to 30 percent of the total. Clinical trials from Phase I through Phase III absorb 50 to 70 percent. Regulatory submission and post-launch activities make up the remainder. Scannell identified four drivers of the decline: rising regulatory hurdles, the depletion of easy targets, the "better-than-the-Beatles" problem of beating existing therapies, and low translational accuracy in basic research.

This cost distribution is decisive for evaluating AI's impact. Even a dramatic reduction in preclinical time and expense changes the total bill only modestly if the clinical-trial phase remains untouched.

03AI's impact is concentrated in the preclinical stage and Phase I

As of 2026, AI's demonstrated achievements in drug discovery are clustered in preclinical timeline compression. The traditional 4-to-6-year journey from target discovery to preclinical candidate selection has been shortened to 13 to 18 months in multiple reported cases. One AI-native drug-discovery company reports that none of its AI-designed preclinical candidates have been terminated before reaching clinical trials.

A 2024 BCG analysis estimated that AI-discovered molecules achieve Phase I success rates of 80 to 90 percent, well above the industry average of approximately 52 percent. This suggests that AI outperforms conventional methods in molecular safety engineering and ADMET prediction.

Phase II tells a different story. AI-discovered molecules succeed at roughly 40 percent, statistically indistinguishable from the conventional rate of about 37 percent. Phase II asks whether a drug actually works against a disease, and here the biological complexity of the target — not the precision of molecular design — determines outcomes. As long as AI cannot bypass the biology, Phase II and later costs and timelines will not shrink. This is not a temporary gap that better models will close; it reflects the fundamental difference between predicting molecular properties, where AI excels, and predicting therapeutic outcomes in complex human physiology, where the data needed for reliable prediction does not yet exist at sufficient scale.

04The AI pipeline has surged but no drug has been approved

The number of AI-designed drug programs in clinical development surged from roughly 24 in late 2023 to over 173 by early 2026. Yet as of September 2026, no AI-discovered drug has received regulatory approval. The most advanced programs are entering Phase III, with initial results expected in late 2026 or 2027.

One AI-native company reported that its platform identified development candidates using roughly 330 compounds in 17 months, compared with an industry average of over 2,500 compounds in 42 months. The speed is real, but speed through the preclinical stage does not guarantee speed through the clinic.

This fact carries two implications. First, AI drug discovery is transitioning from "research efficiency" to the harder test of delivering approved medicines. Second, until approvals materialize, there is no way to quantify whether AI has actually changed the cost structure of drug development at the portfolio level.

05Three structural forces keep clinical-trial costs high

The persistence of clinical-trial costs has causes that lie outside AI's current reach.

06Drug pricing regimes are compressing R&D cost recovery

The pharma business model rests on recovering R&D investment through high drug prices. In the United States, the Inflation Reduction Act of 2022 gave the federal government the authority to negotiate prices for Medicare-covered drugs for the first time. The negotiated prices for the first ten drugs, announced in August 2024, represented discounts of 38 to 79 percent off list prices. CMS estimates these will save Medicare approximately $6 billion and patients $1.5 billion in 2026.

Outside the U.S., the UK's NICE and Japan's Central Social Insurance Medical Council have long incorporated cost-effectiveness assessment into pricing. ICER, the U.S. health-technology assessment body, lowered its annual budget-impact threshold from $880 million to $821 million in October 2025. These moves compress the revenue side of the pharma equation. The second cycle of Medicare negotiations will cover 15 additional drugs, and subsequent cycles will add up to 20 more each year. For the industry, this represents a structural shift — not a one-time event.

If AI reduces preclinical costs but pricing regimes adjust downward in parallel, the savings do not accrue to the company. The relationship between pricing regimes and R&D productivity is a decisive variable for the future of the pharma business model.

07AI improves early failure detection, not the cost structure itself

The evidence assembled here points to a more precise conclusion than the broad narrative of AI slashing drug-development costs. AI's demonstrated contribution is to eliminate failing candidates earlier and raise the quality of molecules entering clinical trials. The preclinical timeline compression and Phase I success-rate improvements are expressions of this mechanism.

Three practical implications follow. First, investment in AI is rationally concentrated in preclinical and Phase I pipeline strengthening. Second, clinical-trial costs are governed by regulatory structure, disease complexity, and patient-recruitment logistics — factors that AI alone cannot change. Third, pricing-regime shifts determine whether AI-driven cost savings remain with the company or transfer to patients and payers.

The question facing pharmaceutical companies is not whether to adopt AI but where in the development chain to deploy it, and the answer must account for a pricing environment that is itself in motion. A company that uses AI to shorten preclinical timelines by three years saves time and capital, but if payers treat that efficiency as evidence that lower prices are sustainable, the financial benefit migrates away from the firm. Strategy must address both sides of the equation simultaneously.