Each issue, this briefing reads one study closely and passes on what holds up. The first opens with the paper we think sets the bar for evidence in AI-enabled care: a pragmatic, cluster-randomized trial of a generative AI copilot in Kenyan primary care, published 26 June 2026 1. It is among the first large randomized tests of a large language model (LLM)-based clinical decision support system measured against an adjudicated patient outcome — and the primary result is null. That combination is why it leads issue 001.
What the trial did
Between 22 April and 16 July 2025, 103 clinical officers across 16 primary care facilities in Kenya were randomized — 52 to an AI-assisted workflow, 51 to usual care — and 9,691 patients were enrolled 1. The tool, called AI Consult and built on OpenAI's GPT-4o, ran inside the electronic record: as the clinician documented the encounter, it reviewed the case and surfaced tiered, traffic-light guidance — green when nothing needed attention, yellow for minor issues, red for a critical concern 5. Clinicians kept full authority to accept, change, or set aside every suggestion.
The primary endpoint was blunt and clinical: an expert-adjudicated composite of treatment failure events within 14 days of enrollment 1 — in press accounts, events such as unresolved symptoms or death 5. Whether the patient got worse, in other words. Models are usually graded on far easier tests.
What it found
Read the raw rates and the adjusted estimate together, because they pull in different directions 1:
- The primary endpoint was null. Treatment failure occurred in 102 of 4,693 AI-supported patients (2.2%) versus 94 of 4,654 controls (2.0%) — adjusted odds ratio 0.77 (95% confidence interval, CI, 0.55 to 1.08), P = 0.13. The crude rates look marginally worse with the tool; the adjusted estimate points the other way; the interval is wide enough to contain both a real benefit and a slight harm.
- Safety was clean. Independent review found no serious adverse event attributable to the tool 1, and rates of hospitalization and death were similar in both groups 2.
- The process improved. An independent panel of experienced clinicians judged documentation and treatment planning to be of higher quality in the AI-supported arm, and prescribing was more cost-conscious, even though overall antibiotic rates were similar 2.
The authors' own summary is the one to keep: LLM assistance was safe but did not reduce treatment failure within 14 days, and any benefit, if present, is probably modest 1.
"The model is impressive" and "the patients did better" are different findings. This trial is what it takes to tell them apart.
Why this matters for healthcare
The evidence most people cite for clinical AI is an exam score. On the benchmarks vendors quote, models have climbed fast — a widely reported evaluation put Med-PaLM 2 at 86.5% on MedQA, the United States Medical Licensing Examination-style benchmark, against the 67.6% its predecessor system reported two years earlier 34; our benchmark tracker follows that record. But a test score answers a narrow question: can the model produce the right answer in a controlled setting? It says little about the thing that decides value at the point of care — whether real patients, seen by real clinicians using the tool in a real clinic, end up better off.
This trial asked that harder question at scale, with independent adjudication, and reported an honest null 1. Publishing that result — rather than stopping at the flattering process numbers — is the standard this briefing will hold every future paper to. Alongside it sits our clinical trial tracker and the broader state of the evidence, both of which keep telling the same story: effects attach to specific tools and specific workflows, seldom to the category.
The live question the authors leave open — was the ceiling the model, or the workflow wrapped around it? — is where the member discussion continues below, with the methods walked through and the limitations a reviewer would press. The paper itself: doi.org/10.1038/s41591-026-04503-6. Read the methods before the discussion; it rewards the order.