Sunday, 6 September 2026
Artificial intelligence is dramatically increasing the speed at which pharmaceutical companies can identify potential drug candidates, analyse biological information and automate research tasks. But as the earliest stages of drug development become faster, another problem is becoming increasingly visible: the rest of the pharmaceutical development system cannot necessarily move at the same speed. That emerging imbalance was one of the central themes of a life-sciences panel at AI4 2026 in Las Vegas, where representatives from pharmaceutical companies, clinical research, biotechnology and AI infrastructure discussed how artificial intelligence is changing the journey from laboratory research to medicines reaching patients.
The panel was moderated by Bryson Tombidge, CEO of Tono Health, and included Talia Fakhoury, Chief AI and Regulatory Strategy Officer at Parexel and a former US Food and Drug Administration official; Laura Boykin-Okalebo, a computational biologist working with AI infrastructure provider Nebius; Patrick Lerch, Senior Vice President of Clinical Data Sciences at Gilead Sciences; and an AbbVie executive responsible for clinical systems, digital operations and AI implementation in clinical development. The discussion suggested that AI’s biggest contribution may ultimately come not simply from discovering more molecules, but from reducing the time and resources required to move promising treatments through clinical development.
The starting point was a striking change taking place in pharmaceutical research. AI can increasingly analyse biological targets, screen potential compounds, model proteins and optimise candidate molecules considerably faster than traditional laboratory processes alone. Research teams can therefore produce larger numbers of potential drug candidates in shorter periods. That represents an important scientific advance, but it also risks transferring pressure further along the pharmaceutical development chain. A molecule that appears promising still needs to progress through preclinical testing and several stages of human clinical trials before regulators can consider approving it. Those processes remain expensive and time-consuming, and some elements cannot simply be accelerated through additional computing power. Human biology operates according to its own timetable, while regulators continue to require evidence demonstrating that medicines are safe and effective.
Lerch described the resulting flow of potential compounds emerging from research as a growing wave heading towards clinical development. For pharmaceutical companies, this creates a resource allocation problem. AI may generate more scientifically plausible opportunities, but companies still have finite budgets, clinical-development teams, trial sites and eligible patients. Every decision to advance another molecule therefore remains an investment decision. This means the industry’s next challenge is increasingly about improving the efficiency of clinical development itself.
AbbVie is examining the critical path between the start of clinical development and the eventual submission of a medicine for regulatory approval, looking for individual stages where AI can remove delays. Potential applications include preparing clinical systems faster, accelerating analysis after trials are completed and assisting with preparation of clinical study reports and regulatory documentation. The objective is not necessarily to eliminate large numbers of jobs or simply reduce operating expenses. The more strategically important measurement is time. Removing several weeks from multiple stages of development can eventually shorten a programme by months. Across a large pharmaceutical portfolio, those savings can become commercially significant while potentially allowing successful treatments to reach patients earlier.
Gilead is approaching the opportunity through both scientific and operational applications. On the scientific side, AI is being explored in medical imaging to analyse tumour burden and help researchers evaluate how cancer treatments are performing. Faster analysis could allow development teams to make earlier decisions about whether a programme should progress to another clinical phase. On the operational side, one of the more immediate opportunities involves software development and statistical programming. Pharmaceutical companies produce substantial quantities of specialised code to prepare, transform and analyse clinical data. Much of this work is created for individual studies or analyses. AI-assisted coding can accelerate these processes while retaining validation and human oversight, allowing information to move through the clinical-development system more quickly.
The distinction between speed and useful progress emerged repeatedly during the discussion. AI can produce an answer or complete a task considerably faster, but speed alone does not guarantee that the organisation is moving in the correct direction. Pharmaceutical companies operate in a highly regulated environment where inaccurate analysis can ultimately affect regulatory submissions and patient safety. Human oversight therefore remains particularly important. AI-generated work that feeds into regulatory documentation cannot simply be accepted because it was produced quickly. Companies need quality controls capable of checking outputs while ensuring that automation does not introduce errors or unsupported conclusions into clinical-development processes.
This is creating a different challenge from the one facing many other industries adopting generative AI. Pharmaceutical companies need to balance automation against scientific evidence, regulatory requirements and patient risk. An AI system capable of reducing a four-week analytical task to a day may create enormous value, but only if the resulting work meets the same or higher quality standards. Clinical research organisations are already reporting measurable improvements in some administrative processes. Fakhoury said Parexel has been using AI to reduce the time required to prepare regulatory documentation and regulatory-grade datasets. The broader significance is that these relatively routine processes sit between important milestones in drug development. Accelerating enough of them could collectively reduce the time between clinical phases.
The more ambitious opportunity involves changing the structure of clinical trials themselves. One area receiving increasing attention is the use of computational models to estimate how patients might have responded without receiving an experimental treatment. In carefully designed circumstances, these approaches could supplement information from traditional control groups and potentially reduce the number of patients required to receive a placebo. This could be particularly valuable for rare diseases, where finding sufficient numbers of eligible patients is difficult and allocating some participants to placebo groups can make recruitment even harder. AI combined with historical clinical information and statistical modelling may eventually allow researchers to extract more information from smaller patient populations.
However, the term “digital twin” is increasingly being used to describe several different technologies. The panel cautioned against treating it as a single concept. In some applications, it refers to statistical models used to estimate a patient’s likely response in a control group. Elsewhere, it can refer to synthetic or external control groups built from historical health records, while other organisations use the term for virtual representations of biological systems or operational processes. The underlying opportunity is nevertheless significant. Better modelling could help pharmaceutical companies estimate treatment effects more accurately, determine appropriate trial sizes and potentially reduce uncertainty before expensive late-stage studies begin.
AI can also address another persistent clinical-trial problem: finding patients. Matching patients to studies currently requires analysing complex eligibility criteria against fragmented healthcare information. AI can potentially search electronic health records more efficiently and identify patients whose medical profiles make them suitable for particular trials. Yet this immediately exposes one of the industry’s largest structural problems. Healthcare information remains fragmented across hospitals, insurers, laboratories, research organisations and different technology systems. In the United States in particular, data required for sophisticated patient matching may exist across numerous organisations that cannot easily exchange it. AI therefore does not eliminate the pharmaceutical industry’s data problem. In many cases it makes the importance of that problem more visible.
The same applies to the transition between animal research and human clinical trials. Despite advances in laboratory science, predicting how findings from animal models will translate into humans remains difficult. AI could potentially improve estimates of appropriate first-in-human doses, likely treatment effects, safety risks and other biological responses, but this remains one of the hardest problems in drug development. The panel identified this translational stage as potentially more important than simply designing molecules faster. If AI can substantially improve predictions about which treatments are most likely to work safely in humans, pharmaceutical companies could avoid committing years of development and substantial capital to programmes that eventually fail.
Until that improves, clinical development is likely to remain slower than AI-driven discovery. Researchers may become capable of generating candidate molecules at unprecedented speed, but each candidate still enters a system requiring scientific validation, clinical trials, regulatory scrutiny and real patients. This creates the possibility of a pharmaceutical pipeline bottleneck. Rather than suffering from too few promising compounds, companies could increasingly face more candidates than their clinical-development organisations can realistically process.
Compute infrastructure represents another emerging constraint. Modern biological AI models can require substantial GPU capacity, particularly when researchers are working with genomics, medical imaging, molecular simulations and increasingly sophisticated foundation models. Boykin-Okalebo said demand for advanced computing infrastructure continues to exceed available supply in parts of the market. For biotechnology start-ups without the financial resources or technology infrastructure of global pharmaceutical companies, access to suitable computing capacity can determine whether research progresses at all.
Research computing is consequently becoming part of biotechnology business planning. A start-up may have promising science and financing but still struggle if it cannot secure the computing resources required to train or operate its models. This means founders and investors increasingly need to consider compute requirements alongside laboratory space, clinical strategy, intellectual property and capital requirements. The infrastructure challenge is not simply obtaining GPUs. Healthcare and pharmaceutical workloads frequently involve sensitive patient information, creating additional requirements around storage, encryption, cybersecurity and data location. Computing capacity and data storage therefore need to be designed together.
If sensitive information must remain in a particular jurisdiction while the available computing infrastructure is located elsewhere, the theoretical performance of the AI hardware becomes less relevant. The architecture needs to satisfy privacy and regulatory requirements while allowing data to move efficiently between storage and compute. This creates an important infrastructure investment angle around the expansion of AI in life sciences. Data centres supporting pharmaceutical and healthcare workloads may need configurations specifically designed for regulated information, high-performance computing and secure connections to large datasets. As biological AI becomes more sophisticated, proximity between compute and data could become increasingly important.
Regulation itself may not be as significant an obstacle as some companies assume. Drawing on her previous experience at the FDA, Fakhoury argued that many operational uses of AI within pharmaceutical companies sit outside direct regulatory oversight. AI used to help prepare documents or improve internal processes is different from a system whose output becomes evidence supporting the safety or effectiveness of a medicine. She cautioned companies against automatically applying validation procedures created for older deterministic software systems to every modern AI application. Excessively conservative interpretations can increase costs and delay projects without necessarily improving patient protection.
The more appropriate approach is to evaluate systems according to their actual risk and intended use. AI directly influencing clinical evidence or patient safety requires far greater scrutiny than technology automating an internal administrative task. The European environment adds another layer through data-protection requirements such as GDPR, particularly when patient information is involved. As clinical research becomes increasingly international, pharmaceutical companies and their technology providers need infrastructure capable of operating across different regulatory and data-governance regimes.
This reinforces the importance of global clinical research. AI may accelerate the creation of new drug candidates, but trials still require patients. Access to diverse patient populations, healthcare systems and clinical sites could therefore become even more strategically important as the number of potential therapies entering development increases. The panel also highlighted the importance of including more geographically diverse researchers and datasets in AI development. Models trained predominantly on information from limited populations may perform differently when applied elsewhere. Expanding participation in AI-enabled medical research could improve both scientific representation and the ability to develop treatments for wider populations.
Another obstacle is organisational rather than technological. Pharmaceutical companies cannot simply provide employees with AI tools and expect productivity to improve automatically. Existing processes were generally designed around human workflows and older technology, meaning many need to be redesigned before AI can deliver its full benefit. Lerch argued that implementation and change management are becoming some of the biggest challenges. Employees need to understand why workflows are changing, where AI fits into their responsibilities and how their expertise remains essential. Without that engagement, even technically successful systems may struggle to achieve widespread adoption.
The same lesson applies to pharmaceutical executives. Ordering an organisation to use AI without identifying specific problems can create numerous disconnected experiments with little measurable impact. Scientists, clinicians, engineers and data specialists need to work together to identify tasks where AI can genuinely improve outcomes. Productivity may also initially decline. Employees need time to learn new systems, organisations must redesign processes and AI-generated work requires validation. The benefits may only emerge after those initial implementation costs have been absorbed.
The economics of running AI are becoming another consideration. Companies that initially focused almost entirely on model performance are increasingly paying attention to the computing resources consumed by large-scale AI applications. As usage expands across thousands of employees and complex scientific workloads, the cost of inference and token consumption can become material. This introduces another layer to the investment calculation. Pharmaceutical companies need to decide not only whether an AI application works, but whether the value it produces justifies the infrastructure and computing resources required to operate it at scale.
Despite these challenges, the panel suggested that the industry’s direction has changed fundamentally. AI is already producing measurable value in specific areas of pharmaceutical research and clinical development. The next question is whether those improvements can be connected across the entire drug-development chain. If research becomes dramatically faster while clinical trials improve only incrementally, the bottleneck simply moves downstream. The pharmaceutical industry could find itself with more promising compounds than it has patients, trial sites, regulatory capacity or development resources to process.
The larger opportunity is therefore to apply AI across the complete journey from scientific discovery to clinical evidence. Molecule design, translational biology, trial design, patient recruitment, clinical monitoring, data analysis, regulatory preparation and post-market evidence all represent potential areas for improvement. The biggest gains may come when those individual applications begin operating as part of a connected development system rather than separate AI projects.
For pharmaceutical companies, that could fundamentally change the economics of innovation. Reducing development by several months can have significant commercial value. Improving the probability that a candidate entering an expensive clinical programme will ultimately succeed could be considerably more valuable. For patients, the measurement is simpler. The value of faster molecule discovery remains limited until the resulting medicine reaches the people who need it.
AI has begun accelerating the front end of pharmaceutical research. The next stage of the transformation will be determined by whether the industry’s clinical, digital and physical infrastructure can accelerate with it.
Source: CIJ.World Research & Analysis Team
