Artificial intelligence and the buzz around it feel like they’re moving at warp speed. As this acceleration continues, especially in the life sciences, we need to pause and understand the vast differences between two uses of artificial intelligence that couldn’t be more different. Drug discovery finds the molecule; drug development gets it to patients faster. AI in drug development deserves far more of the attention.
If the industry wants to unlock true commercial and clinical value from artificial intelligence, it must shift attention upstream: from discovering new medications to developing them smarter.
The great misconception
While the FDA has yet to approve an AI-discovered drug, that doesn’t mean there hasn’t been exciting progress. A study found that AI-discovered molecules pass Phase I at 80 to 90%. The historic rate is 40 to 65%. In Phase II, they pass at roughly 40% [1]. The science is truly amazing, but with some of these therapies years away, we need to ensure AI helps pharmaceutical companies and biotechs move these candidates through the regulatory process as quickly as possible.
It’s not new. Drug development teams have used machine learning and statistical models for decades. “AI” has become shorthand for large language models and agentic workflows, and those are being confused with the discovery models in the headlines. People assume the LLM is the AI. In development, it’s usually a network of AI and big data models working together, and the LLM is one part of it.
This conversation needs to happen early in development, with regulatory, commercial, real-world evidence (RWE) and medical affairs teams included from the start.
Where AI in drug development matters most
Most pharmaceutical development teams bring in AI at the drafting stage, whether for writing the protocol, the submission, or the report. This often comes too late, after strategic decisions have already been made.
By then the useful question is no longer “how fast can we write this?” but “should we have written something else?”, and no language model fixes that. The decisions that matter sit at the front of development:
- What does the landscape look like?
- Which patient population to recruit?
- What to compare against?
- Which endpoint will a regulator and a payer accept?
If an organization gets any of these wrong and only learns of a mistake after designing the Phase II protocol, any change means an amendment or a second study. That makes the year before anyone signs the protocol the most critical time to get these decisions right.
The cost of not moving AI upstream
Using AI in drug development earlier in the process can help get drugs approved sooner. It can also save pharmaceutical and biotech companies millions.
Drug developers know amendments are not cheap. About three-quarters of industry protocols now include at least one amendment [2], and each one reopens the drug to institutional review boards, new protocol rules at approved sites, updated statistical analysis plans, and a longer overall study time.
- A median pivotal trial costs about $19 million, and the average for one with an active comparator is about $49 million [3].
- One measuring a clinical rather than a surrogate endpoint costs an average of about $65 million versus $24 million [3].
Plus, every month delayed is an extra month of lost patent time for the drug [4].
Adding AI sooner can also help answer questions from regulators and payers, including: will this drug be worth it? And will it be commercially viable after navigating the regulatory review process?
The decisions to make before protocol sign-off
So how does an organization take its AI “upstream”? It needs to consider the following before a Phase II or Phase III protocol is finalized:
- Who you compare against: Instead of picking a comparator to pass the trial with the least resistance, AI looks at historical protocols and real-world prescribing to identify the standard of care doctors actually use in each market.
- Who you enroll: The eligibility criteria will become the label. Testing inclusion and exclusion criteria against real-world patient registries before the protocol is final helps target the right patient population to treat.
- What you measure: AI models analyze historical FDA and EMA approval documents and payer coverage decisions to identify primary clinical endpoints alongside secondary health-economics endpoints. This ensures trial design aligns with payer requirements from the beginning.
- How big an effect you expect: Establish a realistic range grounded in existing data, rather than a single speculative number.
- How you count patients who drop out: Define this explicitly, as this statistical choice alone can change the entire trial result. Simulating likely patient attrition in advance helps set this rule with confidence.
How to know if upstream AI is working
Clinical trials take years. How can an organization ensure its artificial intelligence is working in the early stages of development? Test it. Here’s how:
- Time saved: Pick a task the team already does, such as a real-world evidence landscape or an Integrated Evidence Plan, and measure days before and after. Work that once took months can be reduced to days.
- Errors avoided: Count protocol amendments per program. The industry average for protocols that required at least one amendment is 76% [2]; if that number isn’t falling, the AI isn’t far enough upstream in the decision-making process.
- Regulatory alignment: How does the model reach a critical decision? If it can’t show its work, regulators can’t review the information.
- Trained on drug development: All AIs are not the same. If it’s trained on the open internet, it will write a lovely protocol but won’t be able to determine whether the comparator is correct.
Artificial intelligence is already making significant strides in the pharmaceutical and biotech industries. And yes, the fact that this technology is discovering drugs is incredibly exciting, but AI in drug development can make a significant impact right now. And the sooner in that development process AI is applied, the fewer mistakes will be made, the more money can be saved, and ultimately drugs can be provided to the patients who need them most as quickly as possible.
Author Bio
