AI Clinical Support Tool Improved Primary Care Decisions

Key Takeaways

• AI support was associated with better documentation and treatment planning, while short-term treatment failure did not differ significantly from standard care.
• Hospitalisation, death, and patient satisfaction were similar between groups, and the investigators reported no evidence of harm.
• Clinicians kept full responsibility, patients did not see the interface, antibiotic-related costs were lower, and the authors said larger studies and evaluation beyond Kenya are still needed.

In a pragmatic cluster-randomized trial in Kenyan primary care, an embedded generative AI support tool improved clinician decision quality and documentation, yet treatment failure within 14 days was 2.2% with AI support versus 2.0% with standard care, according to a June 2026 University of Birmingham news report. The study included more than 9,600 patients across 16 clinics, placing the evaluation in routine workflow rather than a simulated test environment. The main finding paired better decision processes with no statistically significant change in short-term patient outcomes.

Researchers tested AI Consult, a large language model-based clinical decision-support tool embedded in the existing electronic medical record system. Clinicians were assigned to use the record with or without the integrated tool across 16 primary care clinics in Kenya, covering more than 9,600 patients. The comparator was the same electronic medical record without AI Consult. Working in the background, the system generated real-time diagnostic and treatment suggestions and used colour-coded alerts to flag potential concerns. Patients could not see the interface, and clinicians kept full responsibility for diagnosis, prescribing, and referral decisions, keeping the tool within routine workflow rather than in front of patients.

An independent blinded panel of experienced clinicians assessed the quality of documentation and treatment planning. Panel members did not know whether clinicians had used AI support when they reviewed the records. Both measures were judged better in the AI-supported group than in the standard-care group. The investigators also reported lower antibiotic-related costs with AI support, linking this to more cost-conscious prescribing choices rather than higher prescribing rates. The release described this as among the first randomized evaluations of whether generative AI can influence patient-level outcomes, but the clearest gains in this trial were clinician-facing process measures.

Short-term patient outcomes were similar between groups, with no statistically significant difference in treatment failure during the 14-day follow-up. Hospitalisation and death rates were also similar, and the investigators reported no evidence of harm. Patient satisfaction was the same in both groups, and the release said patient trust was not undermined. In this readout, measured safety signals and patient experience were unchanged.

The authors said larger studies may be needed because serious outcomes such as hospitalisation or death are rare in primary care. They estimated that detecting modest patient-level effects could require studies involving more than 100,000 patients. The trial was published in Nature Medicine, with Bilal Mateen and Richard Riley as senior authors and Alastair Denniston as a co-author, and the study was funded by the Gates Foundation and sponsored by PATH. The researchers also suggested broader relevance, while stating that generalisability to higher-income settings still needs evaluation. Overall, the trial showed better clinician decision processes without a measurable short-term patient-outcome difference during the reported follow-up period.