The most expensive thing in a late stage cancer trial is time, and a surprising amount of that time is spent by people sitting on planes.

On August 12, Medable published an analysis by the Tufts Center for the Study of Drug Development that puts a number on its AI clinical monitoring agent. Expected net present value gains of roughly $7.5 million for a Phase 2 trial, $11.3 million across combined Phase 2 and 3 development, and $21 million for a Phase 3. Direct cost savings of about $4.4 million per Phase 2 study and $5.6 million per Phase 3. Return on investment of 64 times in Phase 2 and 82 times in Phase 3. The figures are benchmarked against Tufts CSDD oncology programme data and Medable’s own contract values.

Every trial site keeps its own records of what happened to each patient. A site could be a hospital, a cancer centre, or a clinic. Those records have to match what ends up in the sponsor’s central database, because regulators will eventually check the two against each other. Traditionally, a person called a clinical research associate flies to the site, sits down with the notes and the screens, and checks line by line. This is called source data verification. It is slow, repetitive work, and once you set the drug itself aside, it makes up a large share of what a trial costs to run.

Multiply that across forty sites, or two hundred, and travel stops being just an expense line and starts showing up in the value of the asset itself. The claimed savings come from three places: fewer on site visits, less travel, and the biggest one, faster enrolment and faster database lock. Database lock is the moment the data stops changing and the statisticians can get started. Everything after it, including the regulatory filing and the patent clock, has to wait for that date.

Eighty two times is a ratio, and the bottom half of that ratio is a software contract. Price the contract low enough and any positive return turns into a huge multiple. The $21 million figure is the one that means more, and even that is a modelled net present value rather than money anyone has watched change hands.

Worth saying plainly: Medable paid for the analysis and supplied half of the inputs, its own contract values. Tufts CSDD supplied the other half, the oncology benchmarks, and is about as credible a source as you can find on the economics of drug development. That doesn’t make the work wrong. But it does mean the study was built to price the product, not to test whether it works.

Most claims about artificial intelligence in medicine are about diagnosis, where the stake is a life. This one is about administration, where the stake is a spreadsheet, and administration is where drug development actually spends its money. If savings on this scale are real, the payoff won’t look like better medicine. It will look like more of it: programmes that stay funded because their numbers crossed a threshold, and trials that report results a few months earlier than they would have. A less romantic story, and probably a bigger one.

Sources: Medable / Tufts Center for the Study of Drug Development analysis, via BioXconomy.