An AI Assistant Improved Clinical Notes Across 16 Kenyan Clinics—at Four Cents in Model Cost

A clinician is halfway through a consultation. The patient’s history is in the electronic record, along with symptoms, vital signs and the clinician’s working diagnosis. Before the visit ends, an AI system offers a second look: possible diagnoses to reconsider, questions that may be missing and a draft treatment plan.

That is the useful frame for AI Consult 2.0, tested in a pragmatic cluster-randomized trial across 16 urban Penda Health clinics in Nairobi and Kiambu counties, Kenya. The system did not replace the clinical officer. It sat inside the existing medical record, while the officer retained authority over diagnosis, prescribing and referral.

The trial’s strongest result was not a dramatic improvement in patient outcomes. It was better clinical documentation and treatment planning, achieved at a reported average model-inference cost of about four US cents per consultation. That figure is not the total cost of building, integrating, supervising or operating the assistant. It is the model cost alone.

Research references: [1]

What the trial tested

Researchers randomized 103 clinical officers—52 to use the AI assistant and 51 to provide usual care—rather than randomizing individual patients. The final paper reports 9,691 enrolled patients and a primary analysis of 9,347 encounters: 4,693 in the AI arm and 4,654 in the control arm. The participant flow also records exclusions and protocol deviations, so the analysis should not be summarized as though every screened or enrolled visit contributed identically to the primary result.

The primary endpoint was treatment failure within 14 days, a composite that included an unplanned return visit, worsening symptoms, hospitalization, death or a need to change treatment. This was the outcome the trial was designed to test; documentation scores, satisfaction, consultation time and spending were secondary or exploratory measures.

AI Consult drew on information already entered during the visit. It generated suggestions, but clinicians could accept, change or ignore them. This distinction matters. The study evaluated decision support under human control—not an autonomous diagnostic service.

Research references: [1] [2]

Better notes, but no proven patient-outcome benefit

In a quality-review sample of 2,000 encounters, the AI arm received higher ratings on three dimensions:

• appropriate diagnosis: adjusted proportional odds ratio 1.74 (95% confidence interval 1.28–2.36);
• comprehensive clinical note: 1.68 (1.24–2.27);
• appropriate treatment plan: 1.71 (1.25–2.34).

Research references: [1]

How to read the quality result

These odds ratios describe shifts across ordered quality ratings. They are not percentages of errors prevented, and they should not be translated into claims that care became “74% more accurate.”

On the primary endpoint, 102 of 4,693 encounters in the AI arm and 94 of 4,654 in usual care met the definition of treatment failure: 2.2% versus 2.0%. After accounting for the clustered design and prespecified covariates, the adjusted odds ratio was 0.77 (95% CI 0.55–1.08; P=0.13). The result was not statistically significant.

A Bayesian analysis estimated a mean risk difference of roughly five fewer failures per 1,000 encounters, with a credible interval spanning from 13 fewer to one more. That is compatible with a modest benefit, no meaningful difference or a small harm. It is not evidence that the assistant improved patient outcomes.

This is precisely why the documentation finding is interesting. Better records and plans may be valuable even when a short trial cannot establish fewer adverse outcomes. Clearer notes can support continuity, audit and future decision-making. But those downstream benefits were not all directly measured here.

Research references: [1]

The cost signal needs careful wording

The paper reports an average AI model-inference cost of about $0.04 per consultation. That is an encouraging engineering figure for a resource-constrained setting, but it excludes the broader costs of product development, electronic-record integration, staff training, monitoring, connectivity and governance.

A post hoc analysis also found antibiotic medication spending about $0.15 lower per encounter in the AI arm (95% CI $0.04–$0.25 lower). This is a narrow category of spending, not proof of net savings for the clinic or health system. It should motivate a prospective economic evaluation, not a headline about an inexpensive system paying for itself.

Patient satisfaction was four out of five in both groups. Median consultation time was 11 minutes in each arm, although the distributions differed slightly. These results argue against a large workflow penalty in this setting, but they do not establish that implementation is frictionless elsewhere.

Research references: [1] [3]

Safety evidence is useful—and incomplete

The trial recorded 33 serious adverse events: 27 hospitalizations and six deaths. Investigators did not attribute any causal link to the intervention. The study, however, was not powered to detect rare harms.

Researchers also reviewed 1,000 outputs selected because the system had raised a “red alert.” Of those, 494 were rated definitely safe and 424 mostly safe; 40 were neutral, 31 somewhat unsafe or inappropriate and 11 unsafe or inappropriate. Because this was an enriched subset of flagged outputs, those counts cannot be treated as the overall error rate for every consultation.

A responsible deployment would therefore preserve the trial’s core safeguards: clinician control, visible uncertainty, audit trails, monitoring for failure patterns and a clear route for reporting harmful recommendations. A cheap inference call is not a safety case by itself.

Research references: [1]

What travels—and what may not

This was a real-world study in a functioning care network, which makes it more informative than a retrospective benchmark. It was also conducted in one private urban network with established quality-improvement systems. Roughly 60% of visits involved febrile or infectious conditions, and follow-up lasted 14 days.

Those conditions narrow the claim. Results may not transfer unchanged to public facilities, rural clinics, different disease mixes, other electronic records or health systems with fewer staff and weaker connectivity. The result is also specific to the tested model, interface, prompts and implementation period. AI systems change quickly; evidence has to follow the deployed version.

The study was funded by the Gates Foundation and sponsored by PATH. Two authors reported Penda Health stock options. OpenAI supplied compute credits and technical guidance but, according to the paper, had no role in study design, data collection, analysis, manuscript preparation or the decision to publish. Those disclosures do not invalidate the results, but they belong in an accurate reading of them.

Research references: [1]

A more human way to read the result

The most compelling version of this story is not “AI outperformed clinicians.” The trial did not test that proposition. It is that a well-integrated assistant gave clinicians a structured second look while leaving judgment and accountability with the person in the room.

For health systems considering similar tools, the next questions are practical:

1. Does the assistant improve documentation under local conditions? 2. Does it preserve clinician agency and patient trust? 3. Can rare but consequential harms be detected quickly? 4. What is the full cost—including integration, training and oversight? 5. Do better notes eventually translate into better health outcomes?

AI Consult 2.0 provides credible evidence for the first question in 16 Kenyan clinics. It offers cautious encouragement on workflow and a useful model for human-in-the-loop design. It does not settle the remaining four.

That is still meaningful progress: not a machine taking over the consultation, but a second look that a clinician can interrogate, amend or refuse.