Discussions of artificial intelligence in healthcare tend to centre on model performance. Yet the more consequential constraint on deployment is not accuracy but the cost of applying these models at scale. This distinction matters, because cost - not capability -increasingly determines how much of the clinical record can realistically be analysed.

‍

The scale of the challenge is illustrated by the processing of large patient populations. Industry estimates suggest that analysing a cohort of one million patients can cost in the region of $1M on dedicated infrastructure, rising to an estimated $13-30M when frontier-model APIs are used. At LynxCare, the same volume can be processed at approximately thirty times lower cost than the $1M benchmark.

‍

The significance of this reduction lies in its effect on analytical scope. When each additional document, clinical note or model call carries a material cost, organisations are compelled to make trade-offs: sampling from the patient population rather than analysing it in full, restricting the range of documents processed, or running analyses only once. A substantial reduction in marginal cost changes these constraints. It becomes feasible to process every relevant document across an entire patient population, to re-run analyses as models improve, and to examine the record in greater depth where the clinical question warrants it.

‍

This efficiency is achieved not by applying a large language model to every task, but by combining efficient data processing, targeted AI models and human-in-the-loop workflows, reserving computationally expensive methods for the tasks in which they add genuine value. The outcome is an approach to AI that is both accurate enough for clinical application and economically viable at the scale healthcare demands. Ultimately, the cost of computation should not determine which clinically relevant data is analysed.

‍