← Cases
KenyaHealthcare and life sciences2019 – ongoing

Jacaranda Health

PROMPTS AI-assisted maternal health helpdesk

Jacaranda Health's PROMPTS sends about 7% of messages directly to nurses, with 18 nurses monitoring the system around the clock

WorkingPress reporting

Problem

Kenyan mothers in public clinics had questions about pregnancy and newborn care but little access to staff between visits. Danger signs can be missed when messages pile up.

Approach

The nonprofit set up PROMPTS in 2017 as an SMS service and added AI in 2019 to sort messages by urgency. A Kiswahili health language model, UlizaLlama, was finished in 2024 and answers routine questions. Messages that show danger signs go to staff, and 18 nurses monitor, audit and answer questions around the clock. The service works in English, Swahili and Sheng and has spread to Ghana, Nigeria and Tanzania.

Results

18nurses who monitor, audit and answer PROMPTS messages around the clockPress reporting[1]
~7%share of messages flagged as potentially urgent and sent straight to a nurse, bypassing the AIPress reporting[1]
~30%share of cases where the program still makes errors, Jacaranda's own estimateCompany's own claim[2]

Takeaway

Jacaranda estimates the program still errs in about 30% of cases, so human review stays and adds cost. Experts quoted by NPR said such tools work best where health systems already function.

Categories

Evidence
Press reporting

Sources

  1. Technology in support of maternity
    IDRC (Québec Science)20 May 2025News article

Source summaries

[1]NPR17 Sep 2026

NPR reports PROMPTS runs in 24 Kenyan counties, with daily questions growing from under 100 to about 15,000. Eighteen nurses monitor the system, about 7% of messages go straight to a nurse, and upkeep costs about $2 per user, paid by the nonprofit. Experts warned that the system could misclassify an urgent message.

[2]IDRC (Québec Science)20 May 2025

IDRC, a funder of PROMPTS, reports that the service started in 2017 and added the UlizaLlama model in 2024. It reports 650,000 mothers registered in 2024 and 4,000 to 15,000 messages a day. Jacaranda estimates errors in about 30% of cases, and a human takes over when a mother is unhappy with an automated reply.

Full details are in the original sources.

Added 25 Oct 2026 · Updated 25 Oct 2026
Open case page ↗