AI and Health Economics — Can the $2.2 Billion-Per-Drug Model Change?── AI shows early evidence of shortening preclinical timelines and improving Phase I success rates, but proven cost reductions remain absent, while health technology assessment frameworks are advancing ahead of the technology itself

In 2024, the FDA authorized 331 AI- and machine-learning-enabled medical devices, bringing the cumulative total past 1,451 by March 2026. How far can AI actually reshape healthcare spending and drug development costs? The honest answer, for now, is that limited evidence exists for preclinical timeline compression and early-phase success rate gains, but no verified reduction in overall development costs has been demonstrated.

01$2.2 billion per drug — the cost estimates and their assumptions

Deloitte's 2024 report estimated average drug development costs across the top 20 pharmaceutical companies at $2.23 billion per asset. DiMasi and colleagues placed the capitalized cost per approval at $2.56 billion (2013 dollars) in 2016. The methodologies differ, but both converge on a figure exceeding $2 billion. Most of that cost accumulates during clinical trials. Phase III cycle times grew 12% year over year, and the average span from Phase I to regulatory filing now exceeds 100 months.

A critical detail: these estimates are averages that include failed candidates. Roughly 90% of compounds entering clinical development never reach approval. The cost of one successful drug carries the accumulated investment in nine that failed. For AI to reduce this figure, it must lower the failure rate, shorten the timeline, or both.

02Healthcare spending at 9.3% of GDP — drug costs in context

OECD member nations spent an average of 9.3% of GDP on healthcare in 2024. The United States stands at 16.7%, while Japan, Germany, and France fall in the 11–12% range. Drug expenditure accounts for roughly 15–20% of OECD average health spending. Lower R&D costs could, in theory, translate to lower drug prices, but pricing is governed by regulation, negotiation, and patent duration — not by production cost alone.

The areas where AI's impact on healthcare spending is measurable remain narrow. Of the FDA's 1,451 authorized AI-enabled devices, 76% are concentrated in radiology imaging. Expansion into drug discovery and clinical decision support is still in early stages.

03High development costs function as both barrier and justification

The $2 billion-plus per-drug figure carries two structural consequences. First, it serves as a barrier to entry. Few organizations possess the capital required to sustain a drug candidate through late-stage clinical trials, reinforcing industry concentration. Second, high R&D costs provide the basis for pricing arguments. The claim that revenue must recoup development investment depends on which cost estimates are used and what assumptions underlie them.

If AI lowers development costs, both structures shift simultaneously. Lower barriers increase competition; weakened cost-recovery arguments strengthen price reduction pressures. But neither shift materializes until AI's cost impact is statistically verified at scale.

04AI acts on three stages — target discovery, preclinical, and trials

AI's potential influence on drug development costs maps onto three stages: target discovery (identifying disease-relevant proteins), preclinical work (compound design, optimization, and toxicity assessment), and clinical trials (patient selection, trial design, and data analysis).

05AI drug discovery success rates — early data and its limits

According to PitchBook's analysis, AI-native biotech companies achieved Phase I success rates of 80–90%, compared to an industry average of 40–65%. Phase II rates reached 40%, above the 29% industry norm. Extrapolating from these figures, one projection places the overall probability of approval at 18%, up from the conventional 8%.

Two caveats qualify these numbers. First, AI-focused biotechs have completed only about 10 clinical trials to date, a sample too small for statistical confidence. Second, as of late 2024, no AI-discovered or AI-designed compound had received regulatory approval. Whether early-phase gains persist through late-stage development and post-market performance remains an open question.

06The conditions under which costs actually fall — timeline compression meets opportunity cost

Drug development costs consist of three components: direct expenditure at each stage, accumulated costs of failed candidates, and the opportunity cost of capital (the time value of money tied up during development). In DiMasi's $2.56 billion estimate, opportunity cost accounts for roughly half. Halving the development timeline, even with unchanged direct costs, substantially reduces the capitalized total.

Reports of AI compressing preclinical timelines from 3–6 years to 12–18 months, if validated broadly, imply significant opportunity cost savings at that stage. However, Phase III trials are constrained by regulatory requirements for patient numbers and endpoints that AI cannot override. The largest cost component sits in the phase where AI has the least leverage.

07HTA frameworks now address AI — CHEERS-AI and reporting standards

In 2024, NICE published CHEERS-AI, extending the CHEERS 2022 health economic evaluation reporting standard to cover AI interventions. The extended checklist contains 38 items: the original 28 CHEERS items plus 10 AI-specific additions. New items address AI training data description, performance changes over time, and the degree of user autonomy — whether clinicians retain final decision-making authority. ISPOR has formally endorsed the standard, and it is registered with the EQUATOR Network.

What CHEERS-AI demands is transparency: when reporting the cost-effectiveness of an AI health technology, authors must disclose how the AI was trained, who retains clinical judgment, and how performance may shift as the system learns. Traditional drug and device evaluations never faced these questions because their performance was static. AI introduces dynamic performance into a framework designed for fixed interventions.

This matters because HTA judgments determine whether AI technologies gain reimbursement and integration into public health systems. Without CHEERS-AI-compliant evidence, AI medical technologies remain in a category of "potentially effective, cost-effectiveness unknown."

08R&D returns tell the story so far — and GLP-1s distort it

Deloitte's 2024 report placed the internal rate of return on R&D investment across the top 20 companies at 5.9%, up from 4.1% in 2023 and 1.2% in 2022. The recovery, however, was driven primarily by GLP-1 receptor agonist revenues. Excluding GLP-1 assets, the IRR falls to 3.8%.

These figures do not yet reflect AI's cost impact. Most AI-native drug development programs have not reached market, and their effect on R&D economics will only become measurable once post-launch revenue data is available. If future Deloitte reports separately track AI-originated assets, the industry will gain visibility into whether AI reduces development costs or merely accelerates timelines without changing the total bill.

09Three pathways by which AI could lower healthcare costs — and the structures that resist each one

AI could affect aggregate healthcare spending through three channels. First, lower drug development costs leading to lower prices. Second, improved diagnostic accuracy reducing unnecessary tests and treatments. Third, clinical decision support improving care quality and reducing readmission rates.

Each pathway faces structural resistance. Drug prices in most countries are set by regulation and negotiation, not by production cost — lower R&D spending does not automatically translate into lower prices. Diagnostic AI remains concentrated in radiology, and expansion to other specialties requires extensive clinical validation. Clinical decision support tools depend on workflow integration and physician trust. Technology alone does not reduce healthcare spending. Institutional, regulatory, and operational changes must accompany it.