Pharma Has a Trial Problem. It Is Not the AI.
Pharma Has a Trial Problem. It Is Not the AI.
Fewer than one in three clinical trial protocols are linked to documented patient data when researchers design them. That is the actual finding at the center of the pharma trial crisis — and it predates the AI hype cycle by years.
Phesi, a clinical trial data company founded in 2007, analyzed more than 600,000 protocols and found that only 29.3% are connected to publicly documented patient data and outcomes at the design stage, according to AZoLifeSciences. The company's blunt summary, as FierceBiotech reported: "flaws are being scaled, not solved." That is not an AI critique. It is a protocol design critique — and it is the more important one.
The amendment data sharpens the picture. Research published in Clinical Trials found that the share of protocols with at least one amendment rose from 57% in 2015 to 76% more recently. Mean amendments per protocol increased 60%, to 3.3. Phase III protocols now average 3.5 substantial amendments each. Each amendment triggers a cascade: a resubmission to a central IRB, weeks of lost enrollment time, budget renegotiations sites never fully recover from. A coordinator who notices a flawed eligibility criterion on day three of a trial typically cannot fix it without starting a process that takes weeks and costs real money. The system was not designed to listen to the people running it.
The pharmaceutical industry has spent the last several years pouring money into artificial intelligence tools for clinical trials. The pitch was clean: AI will find patients faster, cut protocol cycle times, and finally move the needle on the 90% drug candidate failure rate that has defined the industry for decades. The tools got more sophisticated. The budgets grew. The failure rate barely moved.
The reason is becoming clearer now, and it has almost nothing to do with the quality of the AI models. Pharma's clinical trial infrastructure is built on foundations that were never designed to support what anyone is asking of them today. The data used to power AI systems was never designed with AI in mind.
Why does this persist? In part because pharma's data estate accumulated over decades through choices that made sense locally and did not produce a coherent enterprise foundation. Clinical data lives in CDMS systems optimized for trial execution; manufacturing data lives in separate MES and historian systems. Biomarker and imaging data are stored in systems that cannot communicate with EDC or safety platforms. Different study teams, CROs, and geographies used different standards. The result is an environment where even well-intentioned attempts to apply AI end up training models on data that is inconsistent, disconnected, or simply wrong for the question being asked.
Erik Terjesen, managing director at Silicon Foundry (a Kearney company), made the case directly in an essay for GEN News earlier this year. His firm profits from exactly this data infrastructure mess — which is worth noting when weighing his diagnosis. But the underlying observation is not unique to his analysis. "The real bottleneck isn't the biology; it's the lack of standardized, high-quality training data," one industry commentator put it recently.
The regulatory world has noticed. In January 2026, the FDA and EMA jointly published 10 Guiding Principles for Good AI Practice in Drug Development. The principles emphasize human oversight, risk-based approaches, data governance, and lifecycle controls. They are high-level, but their existence signals something concrete: the agencies understand that AI is arriving in drug development whether the data is ready or not, and they are trying to set expectations before the mismatch produces real patient harm. The guidance does not solve the infrastructure problem, but it raises the stakes for any pharma company that planned to bolt AI onto existing workflows and call the problem handled.
There is a class of response that deserves attention. Several large pharmas have begun creating chief AI officer roles and investing in what they describe as data foundation work — curating internal datasets, building governance frameworks, accepting that the models are not the problem and never were. This is harder and less visible than announcing an AI partnership with a major technology company. It does not generate press releases with impressive partner names. It takes years and requires admitting that the previous round of investment did not work. It is also, by most accounts, the right move.
What would actually changing this look like? The more ambitious vision is a shift from trials as one-time experiments to trials as perpetual learning systems — where data from every patient, every protocol, every amendment compounds across time rather than being filed away at study close. That requires solving the interoperability problem at a level most companies have not touched yet. It also requires a different relationship between sponsors and sites, one where the people running the trial have genuine authority to flag problems before they become amendments. None of this is simple. All of it is necessary.
The shorter version: pharma wanted an AI solution to its trial productivity problem. The AI was not the problem. The trials were built wrong, have been built wrong for decades, and adding a more powerful model on top of broken protocol design does not fix the protocol. That reckoning is arriving now, more slowly than anyone in the industry would like to admit, and with more cost to patients than any press release acknowledges.