
[Adobe Stock]
Pharma is investing aggressively in AI to make processes ranging from early stage discovery to clinical trials more efficient. The former area in particular is growing briskly. Venture and private equity funding for AI-enabled drug discovery roughly doubled from $4.1 billion in 2023 to $8.4 billion in 2025, according to McKinsey. Meanwhile, the cost of developing a drug keeps rising. The average cost from discovery to launch rose to $2.67 billion in 2025 from $2.23 billion in 2024, according to Deloitte. For context, the average cost was about $2 billion in 2021.
Among selected AI-enabled programs with enough self-reported and public data, first-in-human or equivalent preclinical milestones arrived 15% to 80% faster than large-pharma medians, according to McKinsey’s September report.
But while agentic loops are steadily improving in software development and AI development, similar loops are far harder to engineer and close in pharma. Mind-boggling complexity runs through every layer, spanning the biology, organizational culture and the clinical trials and post-market signals that follow a drug to patients. If it gets that far.

Guang Yang, McKinsey
Drug discovery, of course, spans a spectrum of distinct problems, each with its own data and its own best-fit model, and a single program can require a molecule that clears several of them at once. “Developing a successful drug candidate requires balancing many different properties, from potency and selectivity to safety and how the molecule behaves in the body,” said Guang Yang, a partner at McKinsey. “Different computational models can help predict and optimize these properties, but the real challenge is finding a molecule that meets the necessary criteria at the same time.”
He described a recent conversation with a biotech company seeking a small molecule that could modulate a specific ion channel, achieve greater selectivity than existing compounds, penetrate the brain, and demonstrate a suitable pharmacokinetic and pharmacodynamic profile.
“They’re essentially looking for a unicorn, a molecule that sits at the intersection of many demanding requirements,” he said. “Optimizing any one property can be done, but achieving the right combination is much harder. That’s one of the fundamental challenges in drug discovery.”
In the lab: Building AI traction meets biological complexity
AI investment has clustered in molecule design because the experiments report back quickly. A model proposes an antibody, scientists make it and within weeks an assay shows whether it binds. Yang contrasts that cycle to a math problem, where a wrong answer shows up immediately and the model can be corrected. “If you go through 1,000 iterations of feedback, the model will unquestionably improve,” he said. The same kind of fast, unambiguous feedback is available in formal theorem proving and competition math, where a proof assistant or answer key catches a wrong step immediately, and in software engineering, where unit tests or a compiler return a pass/fail signal within seconds.
In drug discovery, a given binding result is only one answer among many. The molecule in the aforementioned example also has to be specific for its target, reach the brain, and behave well in the body overall. A tool that helps with one of these properties can leave the others unresolved. Establishing whether a drug ultimately helps patients requires a much longer chain of evidence.

Christian Baber, Pistoia Alliance
Even where scientists can test predictions, they have to trust the tools. Chat interfaces have raised expectations that AI should just work, said Christian Baber, chief portfolio officer at the Pistoia Alliance, a life sciences collaboration organization. Scientific applications demand consistency, and the controls that enforce it can make systems harder to use.
Speaking of LLMs’ potential to, at times, be confidently wrong, Baber said that distrust in them has been an issue among chemists, for instance. “It’s quite difficult to build trust and really easy to lose it,” he said of LLMs. While models’ performance steadily improves with new models, they remain probabilistic. And a confident error backed by an invented reference can end a scientist’s willingness to use a system. Pistoia’s surveys have found AI’s value in document drafting and summarizing, where users expect to check the output. Having AI draft clinical protocols and certain reports can save “weeks or months of people’s time,” Baber said, and the drafts still need checking, as a human’s would.
One workaround is to have the model build deterministic workflows. In the Alliance’s work on instrument data converters, researchers ask an LLM to write a Python script for converting data, then test, validate and reuse the script. “You use something unpredictable to generate something predictable, and that’s what you use,” Baber said. Regulators are more receptive to that approach, he added.
Shared data is another example of a common bench-level problem. Some labs still share reports while keeping the underlying data within their group. Some have a FAIR data system and still keep the data inside the lab, he said. AI’s demand for accessible data is pushing companies toward central storage, with reports generated from that shared resource. Baber also wants researchers to capture why they ran an experiment, what they expected and what they concluded, so the next scientist can understand the question behind a measurement. Getting that context into a shared system requires people to change their daily work. Some companies now count only what reaches the electronic lab notebook, and Baber sketched how that conversation goes: “I did 30 syntheses last month.” “Your ELN says you didn’t do any.” “Oh, I just didn’t put them in my ELN.” “Well, then you didn’t do them.” What’s still missing, he said, “is the carrot for people.”
Culture change is tricky from the lab up to the executive ranks
As has long been the case for digital and process changes in R&D, the people are the bottleneck. “The issue with digital transformation is not the technology; it’s the people. They’re too busy,” Baber said. Scientists have experiments to finish while their employers introduce new software, data requirements and ways of working. Baber traced the Alliance’s change-management community, formed about a year to 18 months ago, to an idea from Novo and that recurring complaint.
One common response has been to borrow agile methods from software development, short cycles of iteration and continuous adjustment. “Agile is great in theory. In practice, it means never-ending change for the people in the lab who are trying to get work done,” he said.
The Alliance’s January white paper, Elevating Organizational Change Management in Pharma and Life Sciences, draws on five structured sessions and survey data from seven pharmaceutical companies and two technology partners. Participants described “change theater”: elaborate communications, meetings and rollout plans that leave users’ needs unresolved. Most respondents rated several practices as unmet, including preparing users before rollout, assigning clear accountability, updating operational documents and securing advocacy from project or department leaders.
The problem of change management extends throughout organizations, including to the executive ranks. Some leaders “will quite often talk the talk but not walk the walk,” Baber said, asking employees to help design and adopt a technology without lightening their workload. “You’re punishing people for implementing and adopting these new technologies, because you’re giving them extra work and no extra time to do it in,” Baber said. A trained chemist, he compares digital transformation to a reaction that needs energy up front before it settles at a lower state. “There’s work at first, but in the end it takes less work to get the same amount done. Quite often it doesn’t work like that,” he said. His alternative to carrot and stick is time: “allowing time to do what is necessary.”
The Pistoia Alliance recommends involving users early, rewarding feedback and peer mentoring, and pairing central change-management standards with practitioners embedded in project teams. In practice, that means time for adoption, someone responsible for helping users, and a route for their feedback to change the implementation.
Incentives need a similar type of grounding. Counting ideas or candidates handed to the next team encourages employees to optimize for volume. Baber argues that rewards should reflect developability, clinical endpoints and translation, even though those outcomes take longer to assess. “If your carrot is how many things are thrown over the fence, all you’ll do is barely clear the fence and not get the best candidate,” he said.
Yang sees companies pursuing two parallel tracks. The first is decentralized experimentation, giving individual researchers the freedom to adopt tools best suited to their work. “You have 16,000 individual employees, and they all need to experiment with whatever tool is out there, and that may change every three months,” Yang said.
The second track involves institution-wide big bets on problems that require coordinated teams, substantial infrastructure or strategic partnerships. Unlike individual experimentation, these investments demand executive conviction years before results become visible.
Pharmaceutical companies are taking different approaches. Some are investing heavily ahead of proven returns, including building or scaling supercomputing modeling capabilities. Others are holding back, risking a late start in developing critical capabilities if the technology delivers on its promise around 2030. A third group is pursuing partnerships that tie payments to development milestones, shifting much of the early financial risk to partners and their investors. Headline deal value among the top 20 pharma companies doubled to $29.5 billion in 2025, but up-front cash came to $0.6 billion, or 2%, and partners in 38 deals signed in 2023 and 2024 had collected less than 1% of maximum milestone value by June 2026, the McKinsey report says.
“R&D is always about this: whatever you invest in now will only materialize in seven years. It requires people to have conviction before they actually have the data,” he said.
Waiting months or even years
Feedback slows dramatically once a candidate reaches people. Whether a molecule binds to a target can often be tested within weeks. Whether a drug is safe and effective in humans may take years to establish. That difference matters for AI, which improves by learning from experimental results. “Imagine every time you tell the model it’s wrong, you have to wait five years, versus just a month,” Yang said.
Earlier prediction of human toxicity is one of the biggest prizes. Roughly one-quarter of Phase 2 and Phase 3 failures stem from unpredicted safety issues, according to McKinsey’s report, and animal livers differ enough from human ones that a drug can pass preclinical tests and still fail. “Better predictions could help companies identify problematic candidates earlier, but whether AI can meaningfully improve overall clinical success rates remains an open question,” Yang said.
For a biotech startup, the wait for feedback can determine whether the company survives. A failed lead program can cut off funding before it gets another chance. “If you’re looking at just one drug candidate, statistically, it’s much more likely to fail than succeed,” Yang said. “Historically, roughly 90% of assets entering Phase 1 never make it to approval.” The industry’s learning curve can outlast the companies generating the evidence.
“The inherent problem is you need to pool multiple trials to get a statistically robust number,” Yang said. The industry has encountered similar challenges before. Evidence accumulated across many development programs has helped establish that drugs targeting well-validated causal human biology tend to have higher success rates. For AI-originated drugs, McKinsey doesn’t expect that critical mass of readouts until around 2030; most AI-native clinical assets remain in Phase 1 or Phase 2, and definitive Phase 3 efficacy readouts are years away, the report says. Until then, each readout is a one-off. “It’s like a weather forecast predicting a 60% chance of rain. If it doesn’t rain that day, it doesn’t mean the forecast was wrong,” he said. “You need to look at hundreds of forecasts to know whether the predictions are accurate.”
Trial operations offer a faster loop inside the slow one. Operational data is well structured, Yang said, and enrollment shows within about a month whether a team picked the right sites, so AI can learn while a study runs toward its clinical endpoints. “Right now, Phase 1 to approval takes roughly seven years. Imagine squeezing that down to four. That’s going to be worth billions and billions,” he said.
For now, the push to reduce experimentation through AI may actually be increasing it. Companies are investing in large-scale biological data generation to train and improve their models, creating additional demand for laboratory experiments, consumables and equipment. Yang calls it a “training bump.” “Of course, the question is how long the training bump will last,” he said. “When will AI be good enough that the balance tips over?”




Tell Us What You Think!
You must be logged in to post a comment.