Five Places Where AI Stalls In Clinical Supply
By Eric Pittman, Vice President of Quality Assurance & Regulatory Affairs at Project Farma
AI adoption is earlier than the noise suggests. A peer-reviewed Tufts CSDD and DIA survey of 302 respondents, published in 2025 and averaged across the 36 clinical development activities assessed, found 36.9% reporting no AI or machine learning use at all, 30.3% beginning or piloting, 22.1% partially implementing, and 10.7% fully implemented.1
The failure mode is almost never the model. In a 2026 survey of 200 senior life sciences decision makers commissioned by Medidata, the top three barriers to scaling AI came out as integration complexity (79.5%), model accuracy concerns (77.5%), and weak data foundations (75%).2 Two of those three are plumbing.
The data problem in supply is specific. IRT, ERP, the depot warehouse systems, packaging and labeling, EDC, CTMS: six systems that rarely reconcile down to a single physical unit of drug. Lot genealogy tends to break at the depot boundary, so you can usually trace a kit from manufacture to depot receipt, and separately trace site dispensing back to a shipment, but the join in the middle is often a manual reconciliation somebody performs in a spreadsheet at month end. That does not survive contact with a model trying to learn the relationship between what shipped and what got dispensed.
Blinding is a harder constraint than vendors want it to be. A model forecasting demand by arm needs unblinded data, and the supply planners who need that forecast are not permitted to see it. Firewalled unblinded forecasting teams work. So do aggregate-only outputs back to the blinded side. But this is a governance problem, and most sponsors have not solved it.
Validation under GMP/GCP is mostly settled, with one exception. The framework is all there: GAMP 5 Second Edition has Appendix D11 on AI/ML, ISPE published a dedicated 290-page guide on AI-enabled systems in GxP environments in July 2025, FDA's January 2025 draft guidance introduced a seven-step risk-based credibility assessment framework in which model risk derives from model influence and decision consequence against a defined context of use, and EMA's reflection paper from September 2024 takes a risk-based and human-centric line.3, 4, 5
What none of them resolves cleanly is continuous learning against change control. A model that retrains every week on new enrollment data is, under a conservative reading of computerized system validation, a system changing without a change record. The workaround most sponsors land on is to freeze the model at qualification, and most of the adaptive behavior they paid for goes out with it. I have witnessed workflow validations freeze process because it didn’t meet the predefined criteria to move forward. A frozen model does the same thing more quietly. This can have impact on subjects.
Cold start is the fourth barrier. Autologous cell therapy, radioligand, and first-in-class mechanisms in new indications have no analogue history to train on, and these happen to be the small (n) programs where a forecasting miss does proportionally more damage. AI contributes scenario simulation here rather than prediction, and it should be judged on whether it improves the quality of the risk conversation. I would be skeptical of anyone quoting predictive accuracy on a 12-patient cohort.
The fifth is not technical at all. Forecast ownership sits in the gap between clinical operations, supply chain, and manufacturing, and no model resolves who signs the batch order. Service-level targets routinely go undeclared. Ask three people in the same governance meeting what per-site stockout probability the program will accept and you will get three answers – or, more often, a pause followed by a question about who owns that decision.
What The Ones That Work Have In Common
This is not the best model, in my experience. It’s mostly a few dull habits.
Declare the service level before evaluating anything. Put a number on acceptable per-site stockout probability, because that number is the objective function and without it you are shopping without a spec.
Fix lot-level data lineage first. IRT-to-ERP reconciliation is a prerequisite rather than a parallel workstream. It is tedious and it takes the better part of a year.
Start with re-forecasting rather than initial forecasting: best data, lowest regulatory exposure. Write the context of use in the regulators' own language at kickoff, not when an inspector asks for it. Measure dispensed-to-manufactured ratio, emergency shipment count, and expiry write-off value rather than forecast accuracy.
And keep a human in the disposition loop. FDA and EMA are both explicit on human oversight, and an automated error at this point in the chain lands on a patient. 4, 5
The Number Hasn't Moved
The Tufts benchmark found that 67.87% of the drug shipped to clinical trial sites was ultimately dispensed to patients. That figure is more than a decade old, and it remains roughly where it was.1,6 It’s not for want of mathematics; deterministic supply simulation has been sitting in this function for 20 years and nobody ever needed a neural network to run a Monte Carlo on enrollment. It stuck because the operating model rewarded buffer over precision and nobody carried the difference on a P&L they owned.
That is the part AI does not touch. A better model sitting on top of unreconciled lot data, an undeclared risk tolerance, and an unnamed forecast owner will produce a more expensive version of the answer you already had, and I have watched sponsors spend eighteen 18 validating exactly that.
The tools have got good enough. What is missing is all unfunded, and it is unfunded because none of it demos well.
References:
- Lamberti MJ et al. "The Adoption and Use of Artificial Intelligence and Machine Learning in Clinical Development." Therapeutic Innovation & Regulatory Science 2025;59(5):1074-1086. Tufts CSDD with DIA and 16 biopharma/CRO companies; 302 complete responses across 36 activities. Source for the adoption maturity split. https://link.springer.com/article/10.1007/s43441-025-00803-0
- Medidata, The State of AI in Clinical Trials, second annual report, released May 18, 2026. Survey of 200 senior decision makers, fielded by Everest Group. Source for the scaling-barrier percentages. Vendor-commissioned; attribute in text.
- ISPE. GAMP 5 Guide, Second Edition, Appendix D11 (AI/ML); and ISPE GAMP Guide: Artificial Intelligence, July 2025 (290 pp.). https://ispe.org/news/ispe-announces-availability-ispe-gampr-guide-artificial-intelligence
- FDA. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, draft guidance, published in the Federal Register January 7, 2025; comment period closed April 7, 2025. https://www.fda.gov/media/184830/download
- EMA. Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle (EMA/CHMP/CVMP/83833/2023), adopted by CHMP and CVMP, published 30 September 2024. https://www.ema.europa.eu/en/use-artificial-intelligence-ai-medicinal-product-lifecycle-scientific-guideline
- Getz K, Lamberti MJ, Mahon C, Hsia R, Milligan C. "Assessing Global Clinical Supply Logistics." Applied Clinical Trials, Vol. 25, Issue 10, October 2016. Tufts CSDD benchmarking study; data collected fall 2014 to early 2015. Source for the 90% / 74% / 67.8% figures (the 67.8% based on 12 companies and 57 studies), the 57-hour mean excursion disposition (range 6 to 114), and the import-license delay countries. https://www.appliedclinicaltrialsonline.com/view/assessing-global-clinical-supply-logistics
About The Author:
Eric Pittman is a vice president of quality and regulatory affairs at Project Farma, bringing over 27 years of leadership experience across regulatory compliance, inspection readiness, and global quality systems. He specializes in developing and executing regulatory strategies, preparing organizations for health authority inspections, and building scalable quality frameworks across pharmaceuticals, biologics, and medical devices.
Eric has led global inspection readiness efforts across more than 25 sites, supporting teams through complex regulatory interactions and driving consistent, compliant operations.
Prior to joining Project Farma, Eric held senior leadership roles in both industry and federal regulatory agencies, where he directed large teams responsible for inspections, investigations, and compliance programs spanning clinical, non-clinical, and pharmacovigilance activities. He has also contributed to regulatory policy development, international collaboration initiatives, and industry training programs.
Eric holds a BS from Purdue University and an MBA from Youngstown State University and is recognized for his ability to align quality, compliance, and business strategy.