Zanzibar: Jamii ni Afya AI bot

The problem

Jamii ni Afya (JnA) has transformed how community health is delivered in Zanzibar. Community Health Workers (CHWs) now register households, record visits, and log key health data on a digital platform instead of relying on pen and paper. That means supervisors have more information than ever about what is happening in their catchment areas. But more data has not automatically translated into better supervision.

Supervisors are expected to meet with their CHWs every month to review local health issues, discuss difficult cases, and plan outreach activities. In practice, supervisors often log into their app and see only high-level counts, numbers of visits or registrations, without a clear picture of what is changing, where, and why. For many, preparing for a single supervision meeting can take hours. A supervisor may jump between dashboards, paper notes, and phone calls to CHWs to piece together what is really happening. Patterns like a gradual rise in malaria cases, or a drop in postnatal visits in one area, can easily be missed until they become more serious. Downstream, CHWs walk into meetings and receive feedback that feels generic rather than tailored to the realities of their community that month.

The problem is not a lack of effort or commitment. It is the lack of visibility and timely, synthesized insights at the moments when it matters most. This is where AI can play a productive role, not by making clinical decisions, but by doing the heavy lifting of summarizing patterns, surfacing trends, and suggesting focus areas, while leaving judgment and action firmly in human hands. To explore this possibility, the Ministry of Health, D-tree, and partners chose to start from the ground up, working alongside supervisors and CHWs to understand their lived experience of this supervision gap, and to co-design an assistant that would actually help.

The solution

The Zanzibar Ministry of Health set the vision and ensured any new tool would strengthen, not disrupt, the national community health programme. D-tree, as the long-standing custodian of JnA, brought deep relationships with frontline teams. Dalberg Data Insights joined as the technical partner, leading the solution development, while Dalberg Design led the human-centred design, field immersion, and workflow integration that ensured the assistant aligned with frontline supervision practices. Together, they agreed on a simple principle: any AI solution should make supervision easier, more meaningful, and more humane for the people doing the work.

To live up to that principle, we grounded every major design decision in supervisors’ and CHWs’ own experience. Field visits and interviews across two districts confirmed the core supervision gap: supervisors were spending hours manually pulling insights from JnA data, while CHWs still received guidance that felt too broad or too late. We also saw how diverse supervisors' realities were, different comfort levels with digital tools, different geographies, different connectivity, which pushed them to design for a range of users, not an “average” one. Five supervisors were brought in early as testers, providing feedback on real message flows that shaped specific design decisions: shortening outputs into a few scannable points, and presenting trends in a way that supported collaborative problem-solving rather than individual performance review.

From the many AI ideas we explored, we prioritised a single, practical job: turning raw data into clear visibility. Working with supervisors from the start, we co-designed the JnA AI Bot, a WhatsApp-based assistant that turns routine JnA data into short “trend briefs” and ready-to-use supervision agendas. By surfacing the trends that matter most, the assistant ensures supervisors can focus on the right issues while also reducing the administrative burden of preparing for meetings.

Near the end of each month, the assistant scans recent data in a supervisor's catchment area. It sends a WhatsApp message highlighting a small number of priority topics, for example “Rising child fever cases,” with a one-line explanation of why each was flagged. The interaction is conversational, not technical. Supervisors can ask follow-up questions in plain language, such as “Is this rise across all villages?”, and receive immediate answers grounded in Ministry of Health guidance. Once a topic is selected, the assistant generates a ready-to-use coaching agenda: the learning objective, talking points, anticipated CHW questions, and space for action items. The assistant’s design reflects choices made directly with supervisors during the co-design process. Supervisors were clear that they did not want a new app. WhatsApp was already part of their daily routine, and delivering insights through it meant the tool could fit into existing habits rather than creating new ones. The assistant is designed around community-level data rather than individual records, does not provide patient diagnoses, and operates within the boundaries of national protocols.

The co-design process also helped build understanding and trust around AI. The team explained what the assistant would and would not do, how it used JnA data, and why it would never replace supervisors' judgment. Their questions helped refine guardrails and communication, which later became central to training materials.

The impact

1. Supervisors shifted from generic agendas to data-anchored discussions rooted in real trends
Before the pilot, supervisors told us they often needed more than two hours to prepare for a single monthly meeting. They moved between JnA dashboards, paper notes, and calls with CHWs, and still worried they might be missing something important. With the JnA AI Bot, most of that manual synthesis happened in the background. In our pilot across two districts, almost all supervisors opened the monthly trend alerts, and around nine in ten used the AI-generated agendas to structure their meetings. Many reduced their preparation time to 10-20 minutes. Supervisors described walking into meetings with a much clearer sense of “what matters this month” and why. Instead of spending the first part of the session figuring out which issues to discuss, they could go straight into problem-solving with CHWs.

2. Supervisors learned topics faster and coached with greater clarity
Before, when supervisors surfaced health topics, they often struggled to explain them deeply or respond to CHW questions on the spot. Now, they can learn and coach in the same flow. Each trend brief links directly to short explainers, Ministry of Health reference materials, and anticipated CHW questions with clear, evidence-based answers. This enables supervisors to internalize the concepts before the meeting and guide CHWs confidently during discussions; without needing to pause, consult colleagues, and return later with clarifications. After the pilot, 100% of supervisors reported feeling confident prioritizing topics, compared with 32% before. 100% felt they had timely information, up from 46%. And 90% felt confident coaching CHWs, up from 31%.

3. CHWs received clearer guidance and applied it in their day-to-day work
For CHWs, these changes showed up as better guidance rather than new tools. In endline surveys with 96 CHWs, about three-quarters said they received clearer answers to their questions, and over 80% felt the guidance from supervisors was better than before. All CHWs found the meeting topics easy to follow. 88% of CHWs reported applying supervision guidance directly in their day-to-day work with households, focusing on the small number of issues highlighted in each meeting rather than trying to cover everything. As one CHW shared, “After the meeting, we know which topics are most important in our communities and we can provide targeted health education based on what the supervisor taught us”. 97% said they wanted their supervisors to keep using the assistant, with many expressing interest in having something similar tailored to their own needs in the future.

Lessons learned

• Co-design with frontline staff shaped the tool in ways the team could not have anticipated from a distance. For us at Dalberg Data Insights, the pilot confirmed something we believe strongly: people-centred design is not a “nice-to-have” in AI projects; it is what makes the difference between a clever demo and a tool that actually gets used. Because supervisors and CHWs helped create the assistant, they recognized themselves in it. They could see that it respected their expertise, worked within their constraints, and aimed to make their work easier, not more complicated.
• The pilot points to a deliberate, district-by-district path for scale-up. Rather than a blanket rollout, the team plans to scale district-by-district to test diverse conditions (rural vs urban). District health teams should lead onboarding to ensure ownership sits within the system, not with external partners. The team also sees potential to apply the same logic, turning routine data into just-in-time insight, to other workflows such as referral tracking, outbreak monitoring, or symptom management support for CHWs.
• Sustaining this kind of AI requires commitment across the ecosystem. The pilot suggests that people-centred AI can strengthen, not replace, community health systems, but only if ecosystem partners keep users at the centre. For implementers, this means designing for real workflows rather than idealised ones and sustaining human-centred methods from concept through deployment. For policymakers, it means creating enabling environments where digital innovations are encouraged but governed with integrity, equity, and local ownership. For funders, it means backing long-term system strengthening rather than isolated pilots and supporting learning partnerships that connect evidence to scale. For technical partners, it means building for the realities of frontline work, basic devices, low bandwidth, and multilingual contexts from day one.
• The JnA AI pilot is a proof of concept for people-centred AI in primary health care. By making information timely, relevant, and usable, it strengthened the human relationships that underpin quality care. As health systems worldwide navigate mounting complexity, Zanzibar’s experience offers a clear signal: when AI is designed around the people who use it and the realities in which they work, it becomes not a disruptor but a quiet enabler of better systems and stronger communities.

Launch year: 2025

Attribution

Case study provided to the AI Navigator courtesy of Dalberg Data Insights, Apolitical's content partner. "From Ambiguity to Guided Action: AI-Powered insights to improve community health services in Zanzibar", Dalberg Data Insights, https://drive.google.com/file/d/1f1V25HOTxQ2Gl17EN7PjbuRA7PC0mJq-/view.