The team behind the system tested its performance against standard approaches. In their evaluations, TrialMatchAI identified appropriate trials with higher precision than keyword-based and rule-based matching tools, and it processed patient-trial comparisons far faster than manual screening. The researchers report that the system's recommendations aligned closely with the judgments of clinical experts reviewing the same cases.

The approach relies on recent advances in large language models, which can parse unstructured text — the free-form notes that make up much of a medical record — without requiring that every data point be entered into rigid, predefined fields. This flexibility matters because trial eligibility rules are often written in dense, conditional language that traditional software struggles to interpret. A trial might require, for example, that a patient has received no more than two prior lines of therapy and has adequate liver function, conditions that may be scattered across dozens of clinical notes.

Despite the promising results, the researchers and outside experts urge caution. The evaluations so far have involved retrospective testing — checking whether the system would have identified trials for patients whose records were already available — rather than prospective use in live clinical settings. Real-world deployment introduces complications that retrospective studies cannot fully capture, including incomplete records, inconsistent documentation practices across hospitals, and the need to keep patient data private and secure.

The system also depends on the quality of the data it receives. If a patient record omits a key detail, such as a recent genetic test result, the AI cannot account for it. Clinicians remain responsible for final decisions about trial enrollment, a step the researchers describe as essential to the system's design.

The team has made parts of the system available to the research community and says the next step is testing in prospective clinical settings, where the tool would support — rather than replace — the specialists who currently coordinate trial enrollment. If those tests succeed, the technology could help address the recruitment gap that slows the development of new treatments.

For now, TrialMatchAI represents a concrete example of how language-based AI might fit into clinical workflows: not as an autonomous decision-maker, but as a screening assistant that shortlists options and flags the ones a human expert should examine first.