Pregnant women gain knowledge and receive more recommended care

PROMPTS. [1]

An earlier Google Cloud classifier preceded TRIM-AI. The later system adapted XLM-RoBERTa-large as XLM-RoBERTa-JH. [2]

The exact deployment release and each trial participant’s exposure to it are not reported. [1] [2]

Editor’s note. The trial ran from November 2021 through August 2022, while the technical paper dates TRIM-AI deployment to June 2022. Its use during the latter part of follow-up is plausible, but the trial does not report participants’ exposure by model.

Jacaranda Health. [1]

Natural language processing classified messages for routine replies or clinical helpdesk review. [1]

The later TRIM-AI classifier adapted a multilingual model using 939,819 messages, then trained for 58 intents on 107,714 labeled messages. [2]

Women sent questions by SMS. The platform classified their intent and urgency, then provided guidance or routed the message to clinicians. [1]

In the documented TRIM-AI workflow, sufficiently confident classifications triggered vetted replies. Failed or uncertain exchanges went to staff. [2]

Women received scheduled messages and reminders and could text the helpdesk. Clinicians handled questions that required human review. [1]

Clinical staff prepared responses and handled escalated messages. The technical evaluation asked users whether automated answers were helpful. [2]

The helpdesk used natural language processing to sort incoming questions. [1]

The trial evaluated the complete platform. It did not measure the separate effect of the NLP helpdesk. [1]

Not assessed

The invitation to use PROMPTS improved several measures of knowledge, preparation, and care. [1]

Improved health knowledge and use of recommended maternal and newborn care. [1]

Observed in a randomized evaluation. [1]

Pregnant and postpartum women and their newborns. [1]

Women attending participating facilities who had access to a mobile phone. [1]

The trial enrolled 6,139 women: 3,140 at treatment facilities and 2,999 at control facilities. [1]

The trial covered public and faith-based facilities in eight counties. It excluded women without phone access. [1]

Researchers followed participants from pregnancy into the early postpartum period. [1]

Better knowledge and care were the intended benefits. [1]

The service depended on phone access and did not replace clinical care. The trial did not establish effects on deaths or serious illness. [1]

AIBD-0100-E01, AIBD-0100-E02 [1] [2]

Twenty facilities introduced PROMPTS and twenty continued usual care. Assignment was stratified within counties. [1]

Telephone surveys measured knowledge and reported care. Analysis compared assigned groups and accounted for clustering by facility. [1]

There were 6,139 participants at 40 facilities. Researchers completed 5,509 of 6,128 eligible postpartum interviews. [1]

Knowledge, birth preparation, routine and danger-sign care, newborn care, and postpartum care content. [1]

Women at control facilities were not systematically introduced to PROMPTS, although they could join independently. [1]

Five indices improved by 0.06–0.09 standard deviations. Care-seeking for danger signs did not differ significantly. [1]

The evaluation did not separate automated triage from information, reminders, or clinician support. [1]

Outcomes were self-reported. The study was not powered for clinical outcomes, and a parallel nurse-training program could have influenced care. [1]

Among treatment participants surveyed during pregnancy, 1,453 of 1,700 reported joining PROMPTS. [1]

The public repository contains analysis code and documentation. Participant data remain restricted under the study’s approvals. [1]

A provider estimate of enrollment cost and a separate estimate of NLP operating costs. [2] [3]

Jacaranda estimated US$0.74 per enrollee over their time on PROMPTS. The technical paper reports monthly NLP costs falling from US$819 to US$632. [2] [3]

The sources report U.S. dollars. The technical paper also gives about KES 24,000 for its rounded monthly saving of about US$200. [2] [3]

Per enrollee over their time on the platform, and per month for the separate NLP estimate. [2] [3]

Both figures are provider estimates. The trial did not audit them or measure a net monetary benefit. [2] [3]

Jacaranda and its partners fund delivery. Participants use PROMPTS free of charge. [2] [3]

The enrollment figure is a cost. The technical saving covers model and translation services, not the full health program. [2] [3]

The enrollment estimate covers messages, technical infrastructure, enrollment work, and county partnerships. [2] [3]

The technical comparison is the earlier Google Cloud classifier plus translation versus TRIM-AI. [2] [3]

Jacaranda supplied the operating estimates. Subtracting US$632 from US$819 gives US$187 per month. [2] [3]

No full financial return or payback period is reported. [1] [2] [3]

Jacaranda Health ran PROMPTS with Kenyan health facilities. Academic researchers and Innovations for Poverty Action evaluated it with the provider. [1]

Forty facilities participated across Kajiado, Kilifi, Kirinyaga, Kisii, Meru, Narok, Nyeri, and Siaya counties in Kenya. [1]

Jacaranda Health operated the platform. Participating facilities helped enroll women. [1]

Women needed access to a phone that could receive texts. The service also needed mobile networks, facility enrollment, and clinical helpdesk staff. [1]

The technical work involved Kenyan and U.S. institutions and Google Cloud services. The papers do not establish where participants’ messages were stored. [1] [2]

Not assessed

Women need access to SMS, facilities must support enrollment, and clinicians must be available for escalated questions. [1]

The later report documents expansion, rather than a new randomized test of the Kenyan trial result. [4]

For July–September 2025, Jacaranda reported cumulative reach of 3,580,095 mothers across 23 Kenyan counties and 1,500 facilities. [4]

Jacaranda reported continued operation in Kenya and an agreement for national expansion in Ghana during 2025. [4]

Not assessed

Not assessed

Not assessed

Not assessed

AIBD-0100-source-0001, AIBD-0100-source-0002, AIBD-0100-source-0003, AIBD-0100-source-0004 [1] [2] [3] [4]

The trial provides the benefit findings. Its references lead to the original NLP paper and provider cost estimate. A later provider report covers expansion. [1] [2] [3] [4]

English. [1] [2] [3] [4]

The benefit study includes both outside researchers and Jacaranda. The technical and cost claims come from the developer partnership and provider. [1] [2] [3] [4]

The trial cites the technical paper for NLP and the provider’s brochure for costs. Those original sources are retained. [1] [2] [3] [4]

Original papers, relevant supplements, technical appendix, and provider reports were inspected. Model exposure by trial participant remains unresolved. [1] [2] [3] [4]

Offering the complete platform improved several reported measures of knowledge and care. [1]

The sources are linked to the findings that they support. [1] [2] [3] [4]

A peer-reviewed cluster randomized trial supports the platform’s effects. The evaluation included the provider and outside researchers. [1]

Not assessed

Women’s health knowledge and reported care practices. [1]

Personal: 2 [1]

Women had measurable improvements in knowledge and reported care practices. [1]

Observed in a randomized evaluation. [1]

Not assessed

Not assessed

Not assessed

The deployer's account [1] [2] [3] [4]

A trial that randomized health facilities to introduce a digital health platform. [1]

Both programs used automated health-message support. This is an editorial comparison of a shared function, not evidence that their systems or clinical safeguards were equivalent.

Chatbot Tessa gives unauthorized diet advice to users seeking help for eating disorders

Maternal and newborn health. [1]

PROMPTS provided pregnancy information, reminders, and help with questions through text messages. [1]

Recruitment ran from 8 November to 10 December 2021. Follow-up continued through 24 August 2022. [1]

The platform was operating in Kenya when the trial began. [1]

SHA-256: aa41a37d6114…

AI Incident Database: 545.

Not assessed

Not assessed

8 revisions. Latest: 25 October 2026.