Healthcare AI's Pilot-To-Production Problem

Stoyan Mitov is the CEO of Dreamix, a custom software development company helping tech leaders increase capacity without giving up quality.

​A 2025 Bain & Company and KLAS Research survey found that even ambient documentation, the most mature AI use case in revenue cycle management, was still in pilot at twice as many organizations as had it in full production. If the most mature use case looks like that, the newer ones look worse.

A pilot that performs well in a demo rarely predicts whether the same system will ever touch a real patient record, and the reason usually has nothing to do with the model.

Most pilots are built and tested against a clean, static dataset, disconnected from the systems they would eventually need to run alongside. Reviewers see it work and approve it. Only then does the project need to connect to a real electronic health record, laboratory information system or quality management platform, and the work that looked nearly finished turns out to have barely started.

Where The Gap Comes From

Most healthcare and life sciences organizations run on systems bought in different decades, from different vendors and for different purposes—an EHR here, a LIMS there, a QMS and an ERP layered on top, plus a handful of custom applications built over the years to bridge earlier gaps. None of it was designed to share a workflow. In practice, a nurse, lab technician or quality reviewer ends up doing that integration by hand, opening one system to find a result and retyping it into another.

A pilot can bypass all of that by working from a static export of sample data. Production cannot.

The result usually looks like a pilot that looked finished and is sitting unused for months, sometimes years, while someone works out how to connect it to the rest of the organization safely. The AI model itself is rarely what causes the delay.