Artificial intelligence (AI) is one of the most discussed technologies in pharmaceutical research. Many of those conversations focus on AI-driven productivity: automating workflows, generating reports, or accelerating routine tasks.

Those applications matter. But they may not represent AI's greatest opportunity for drugmakers.

For decades, scientific and pharmaceutical innovation followed a familiar pattern. Researchers develop a hypothesis, generate evidence, run studies, and gradually determine whether a program deserves continued investment. The process works – but it’s slow, expensive, and remarkably unforgiving.

The primary challenge facing pharmaceutical research and development (R&D) is the amount of time required to determine which ideas deserve continued investment. That challenge becomes particularly costly in an industry where roughly 90% of drug candidatesentering clinical development never reach approval.

That failure rate statistic often gets framed as a discovery problem. But what if we viewed it as a learning problem instead?

The Cost of Learning Too Slowly
Too Much Time: Drug development operates under a very different set of constraints than most industries.

In software, a new hypothesis can be tested in hours. In manufacturing, process improvements can often be evaluated in days. In pharmaceutical research, generating enough evidence to validate or reject a hypothesis can take years.

The bottleneck begins at the very first step. Target identification has so far operated as an elaborate guessing game: form a hypothesis around a single gene, spend years running experiments, and find out whether you were right. With upwards of 40,000 possible candidates in the human genome, this one-hypothesis-at-a-time approach is a primary reason why so many complex diseases remain uncured. Breaking through it requires changing the operating system of how drugs are developed.

Yes, every failed molecule contains valuable information. Every unsuccessful trial teaches researchers something about biology, disease pathways, patient populations, or treatment response. The problem is that those lessons often arrive after years of investment.

As one industry leader noted during a recent symposium, "Two and a half years to build a hypothesis is crazy."

For patients, it’s the speed of treatments that matters. Decades of organizational milestones means nothing for a child with a yet-uncured, debilitating disease. Every additional month spent pursuing an ineffective target or molecule carries a real cost for both companies and patients.

Too Much Money: The economic stakes are equally daunting.
By the time a drug reaches late-stage development, companies may have invested hundreds of millions of dollars in research, clinical trials, and supporting activities. In an interview with CNBC, AstraZeneca’s CEO, Pascal Soriot, said, “We spend $300 million, $400 million, $500 million on a trial.”

The Congressional Budget Office reported:

• Total clinical trial spending for a drug completing the first three phases averaged about $375 million.
• Phase III spending alone averaged about $282 million.
• Average spending on clinical trials per approved drug was about $1.1 billion, including the cost of failed programs.

The challenge is not simply finding successful therapies. It is identifying unsuccessful ones sooner.

AI may dramatically increase the number of successful drugs. But a more immediate opportunity may be helping organizations learn – and fail – faster.

The Goal Isn't Faster Success. It's Faster Failure.

AI-powered modeling, simulation, and prediction offer a different approach. Rather than waiting for every experiment to play out in the real world, researchers can evaluate more possibilities, test more assumptions, and identify weak signals earlier in the development process.

In the same CNBC article, Soriot noted that AI is helping the company make better decisions about which drugs to advance through the development process. “The value of AI in our industry [is] in the way you design a new medicine, a new drug, you can actually do it faster, do it smarter.”

AI is not replacing scientific rigor. It’s compressing learning cycles.

None of this means that AI eliminates uncertainty or guarantees clinical success. But it may be about reducing the time and cost spent pursuing losing ideas. Failing fast may be the best use case for AI yet.

Can drug development find a way to fail faster?