Can AI Make Drug Development Cheaper? Follow the Costs

Imagine a research team with enough money to investigate several treatment ideas, but too little to pursue all of them. The team considers an AI tool that could help choose which ideas to test. Running predictions might cost less than making and testing every candidate. The question is whether those predictions lead to better choices.

That sounds like a saving. Whether it becomes one depends on what happens next. Does the team avoid work that would have failed? Spend more time checking predictions? Reject an idea that would have worked? Those outcomes have different economic consequences.

The cost of producing a prediction is only one part of the cost of making a good research decision. To understand what AI might change, we need to follow both the spending and the results.

Why is developing a drug so expensive?

Drug development combines discovery, experiments, manufacturing work, clinical studies, and regulatory review. The FDA’s overview of the development process separates discovery, preclinical research, studies in people, review, and safety monitoring after approval. Each stage asks questions that earlier work could not settle.

A promising molecule may turn out to be difficult to make, behave poorly in the body, or fail to provide enough benefit in patients. The spending on those unsuccessful programs still happened. Looking only at the program that reached approval leaves much of the research effort out of the picture.

Published estimates also differ in what they count. Some describe cash spent on a particular program. Others estimate spending across successful and unsuccessful programs, then include the economic cost of committing money for years.

The Congressional Budget Office’s 2021 report on pharmaceutical research explains how the treatment of failed programs and capital costs affects estimates. A capitalized estimate includes the cost of tying up money over time, under an assumed rate of return. That component is different from a bill paid to a laboratory or trial site.

One example shows why the details matter. A 2020 JAMA study , subsequently corrected , examined available expenditure data for 63 medicines approved between 2009 and 2018. Its corrected base-case median estimate was about $1.14 billion in 2018 dollars, allowing for failed trials and capital costs. That is a historical study estimate with a particular sample and method, not a universal budget for a new medicine today.

Where AI could change the spending

AI can affect several kinds of work. Their value needs to be measured at the point where a team uses the result.

Finding evidence. An agent can help gather papers and connect results to a research question. The saving depends on how much searching it replaces and how much source checking remains.

Choosing molecules to make. Predictions can help narrow a set of proposals before synthesis and testing. A shorter list is useful if it retains candidates worth investigating. A model that confidently removes the wrong candidates can make the search less productive.

Investigating a possible problem earlier. A warning about poor solubility or an unwanted interaction can prompt a focused experiment while a molecule is still easy to change. The warning has value when it leads to a better decision, even if the immediate result is spending money on another test.

Reusing analyses and documentation. Shared tools can reduce repeated setup, copying, and formatting. The team still needs to validate the method for its new use and review the output.

These are plausible routes to savings. They do not establish a percentage reduction in the cost of delivering an approved treatment.

A worked example: spending less can still cost more

Consider a deliberately simplified research portfolio. A team starts 100 exploratory programs at $1 million each. It advances 20 programs to a later study costing $10 million per program. Five eventually meet the same predefined research goal.

Total spending is $100 million plus $200 million, or $300 million. Dividing by five successful outcomes gives $60 million per successful outcome.

Now add a screening and review process costing $30 million across the portfolio. Suppose it reduces the number of later studies from 20 to 15. That removes $50 million in later spending, so the net saving is $20 million.

What happens to cost per successful outcome depends on which five programs were rejected. Explore the comparison below.

In the second scenario, screening removes five programs that would have failed and preserves all five successful outcomes. Spending falls to $280 million and cost per success falls to $56 million.

In the third, the filter removes four programs that would have failed and one that would have succeeded. The same 15 later studies cost the same $280 million, but only four programs succeed. Cost per success rises to $70 million.

That is the risk of a false negative: rejecting something that would have been useful. The example does not predict how a real AI system will perform. It shows why a lower total budget, on its own, cannot establish better research productivity.

This model leaves out time, financing, differences between programs, and the value of their outcomes. Its “success” is an agreed research goal, not an assumed drug approval. Real development budgets need much richer evidence.

An earlier stop can be a valuable result

A team that abandons a weak program after a clear experiment may have made an excellent decision. The useful saving is the avoidable future work, after accounting for the cost of reaching that conclusion.

That does not mean every stopped program should count as money saved. A program may have been stopped by the existing process anyway. Another may have been worth continuing. Those possibilities make the comparison important.

A practical evaluation asks what the team would otherwise have done, which costs changed, and what evidence supports the change. It also follows the candidates that were rejected when that is feasible. Studying only the selected candidates can hide the filter’s mistakes.

Three levels of evidence for an economic claim

A tool can perform well on a narrow task while having little effect on a program’s total cost. Following the result through the workflow makes the contribution easier to assess.

For example, generating a report faster can free a scientist’s time. If the scientist then spends the saved time repairing errors, the net gain may be small. If the report helps the team identify an unnecessary experiment, the downstream saving may be larger than the drafting benefit.

The right measure depends on the claim. A trial is not needed to establish that a calculation runs faster. A faster calculation is not enough to establish that a development program is cheaper.

Count the work around the model

A useful budget includes the effort needed to make an AI result usable. That can involve preparing data, checking compound identities, connecting tools, paying for compute, reviewing outputs, and investigating errors.

Some of those costs are paid once and shared across projects. Others recur for each new dataset or research question. Mixing them together can make an early pilot look unusually expensive, or make a later deployment look artificially cheap.

Separate those costs, then compare like with like. The existing process may already use scripts or predictive models. A fair comparison should reflect that starting point and include the same standard of scientific review on both sides.

For a small pilot, useful questions include:

• How many usable decisions did the team reach, after checking the work?
• What did each decision cost in staff time, tools, and follow-up experiments?
• Which costs were avoided, and which were simply moved to another part of the team?
• What evidence is available about missed opportunities and incorrect recommendations?
• Can another team reproduce the benefit on a comparable task?

These questions are more informative than counting generated molecules or completed tool calls.

Would cheaper development mean cheaper medicines?

For patients, the price at the pharmacy matters more than the cost of a prediction. A reduction in research spending does not automatically determine that price.

The CBO’s 2024 report on prescription drug prices describes the roles of competition, exclusivity, insurance, and negotiation in the United States. Research costs and the price of a marketed medicine are related economic questions, but they are not interchangeable measures.

AI could help a team investigate ideas it could not previously afford to explore. That would be meaningful progress. Claims about lower patient prices need evidence about pricing and access as well as research productivity.

How we approach research efficiency at AgentCures

At AgentCures, we use shared software and AI agents across research tasks. Reusable models, analysis workflows, and documentation templates can reduce the setup work for another project. Keeping research records in version control also helps teams find earlier work and see why a decision changed.

Our focus is on making proposals and their evidence easier to inspect. A useful result might be a clearer comparison between candidates, a well-supported reason to stop, or an experiment that resolves an important uncertainty. Any claim about savings should include the cost of producing and reviewing that result.

For more context, read our guides to AI in drug discovery , how research agents work , and assessing ADMET properties earlier .

What better use of a research budget looks like

The goal is to learn more useful things with the resources available. Sometimes that means running fewer experiments. Sometimes it means paying for a better experiment earlier, or exploring an idea that a narrow ranking would have missed.

AI earns its place when the evidence supports those decisions. Following costs through to outcomes gives researchers, funders, and readers a better way to judge the promise.