More Than Half The Drug We Make Never Reaches A Patient

By Eric Pittman, Vice President of Quality Assurance & Regulatory Affairs at Project Farma

Somewhere between 12 and 24 months before anyone knows how many patients will actually enroll, someone has to sign a manufacturing order for clinical supply. That signature is the origin of most of what goes wrong later: the overage nobody questions, the emergency couriers, the pallets of kit that expire in a depot without ever reaching a site.

The waste is documented and it is worse than people outside the function assume. Tufts CSDD ran a benchmarking study with 12 sponsor and supplier organizations across 57 studies. Ninety percent of manufactured or procured product got packaged, 74% of that packaged product shipped to sites, and 67.8% of what shipped was dispensed to a patient.¹ Multiply it through and about 45% of manufactured drug reaches the person it was made for.

That data was collected in 2014 and 2015, which is a fair objection and I would not pretend otherwise, though it remains the most complete public benchmark this function has and I have yet to meet a supply lead who thinks the number has improved much since.

AI is being sold hard as the correction, and some of that is warranted. Most of the pitch decks I have sat through, though, solve a part of the problem that was never really the constraint.

The Forecast Is Always Wrong

Commercial supply chains forecast off history. Clinical supply mostly doesn't have any, because every protocol is a new product with a new demand curve dropped into a network of anywhere from 20 to 400 sites whose behavior you are still learning six months in.

Site-level demand comes out of enrollment rate, screen failure rate, randomization ratio, titration behavior, discontinuation, and visit window drift. Six stochastic inputs, none of them well estimated at protocol finalization, and they multiply rather than average. The unit that matters is the site. A trial can sit at 90% supply coverage in aggregate while 14 sites head toward stockout next month without anything on the study-level dashboard changing color.

Then there is the cost function, which is lopsided in a way that never gets written down anywhere. A stockout is a protocol deviation, possibly a lost patient, and a call with the medical monitor that you remember for a while afterward. Overage is a number on a budget line nobody audits at the unit level. I have watched planners get quietly praised for a study that ran heavy overage and never missed a dose, and watched the same organization open a root cause investigation over a single site stockout that cost almost nothing and delayed nobody. Nobody in that building was confused about which outcome to optimize for. So high buffers are not a math failure. Planners are reading the incentive correctly, and until somebody puts a number on the other side of that ledger, they will keep reading it the same way.

The distribution also moves while you are working on it. A country activates, or an amendment changes the dosing schedule halfway through, or a competitor reads out and patient flow at your shared sites redirects inside a month.

These are the conditions where simulation earns its money, and also where a model trained on thin data will hand you a wrong number wrapped in a confidence interval tight enough to be persuasive.

Where AI Is Earning Its Keep

Enrollment-linked demand forecasting is the most mature of these. You train on historical enrollment curves segmented by therapeutic area, geography, and site archetype, then refresh continuously against live screening and randomization data. The first forecast does not get much better, which tends to surprise people. What changes is that the output stops being a single number and becomes a distribution, and that turns overage into a decision somebody can defend in a governance meeting. "Twelve percent overage buys a 97% site-level service level" is a sentence you can take upstairs. "We always run 30" is a sentence people still say out loud.

Most of the value sits in re-forecast cadence. A mediocre model updating weekly against live randomization data will outperform a better model updating quarterly. In a simple dermatology study I inspected, I found several subjects had to wait to get the IP delivered. It actually could have caused protocol deviations the clinician and the sponsor would have had to address in the submission, all because they were using an outdated forecast.

Resupply is the other area with a clear mechanism. Predictive triggers replace fixed min/max par levels at site, and the resupply algorithm gets matched to the trial's risk profile instead of inherited from whatever the last study happened to use. This leads to fewer emergency shipments, smaller site buffers, and less expiry churn at slow-enrolling sites.

Depot allocation, meaning which lots go where against remaining shelf life and projected burn rate, is a well-posed optimization problem, and it attacks the 74% to 67.8% band in the Tufts numbers directly. This is the cheapest win on the list and I have never heard a good argument against doing it. Label extension planning falls out of the same model: whether re-labeling at site beats pulling stock back to a depot, given what you know about that site's burn rate.

I would wait on comparator and co-therapy sourcing. The European Court of Auditors counted 83,266 medicine shortage notifications logged by national authorities across 24 EEA countries in 2022 and 2023, and comparator availability for biosimilar and head-to-head work keeps tightening against that backdrop. Lead time prediction helps at the margin. But the ceiling on any model here is set by external market data you do not own, cannot audit, and mostly cannot obtain, so I would not build a business case on it yet. ²

I had written cold chain off, and I was wrong about that. Route-level risk scoring and customs delay prediction both have working implementations, and IoT-fed anomaly detection is no longer exotic. The Tufts benchmark clocked temperature excursion disposition at 57 hours on average, ranging from 6 to 114, and that window feeds directly into whether a patient at the affected site gets dosed on schedule.¹ Prediction plus pre-staged disposition logic compresses it. Import licensing remains a top disruption source, with Argentina, Russia, China, Colombia, and India named repeatedly in the benchmark data.

One last thing, and it comes before any of the above. Your IRT already holds the best demand data set in the company, and almost nobody mines it across studies. IRTs can be a wonderful tool in clinical trials; however, they must be set up correctly. I have seen too many failures at patient dosing and/or blinding over the years caused by poor UAT, and that setup quality is the quality of the data set you are about to mine. Before evaluating a single external model, go find out what your own randomization history says, and how far you trust it.

References:

1. 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
2. European Court of Auditors, Special Report 19/2025, on EU action on medicine shortages. Source for 83,266 shortage notifications logged by national authorities across 24 EEA countries, 2022-2023. https://www.eca.europa.eu/ECAPublications/SR-2025-19/SR-2025-19_EN.pdf

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.