Building an AI-powered clinical advisor to strengthen frontline healthcare in Ethiopia

Decision-maker’s challenge

Community health workers operating in remote areas are frequently required to make critical, life-saving decisions autonomously. In Ethiopia, community health workers—known nationally as health extension workers, or HEWs—serve as a critical interface of the primary healthcare system; however, they often lack immediate access to guidance from senior clinicians when managing intricate cases.

In practice, health extension workers rely on direct phone calls to their supervisors at nearby health centers for clinical guidance. While this provides some support, the process is inconsistent, lacks standardization, and has no system for documenting consultations or tracking outcomes. Supervisors themselves can often be unavailable due to competing responsibilities, leaving health extension workers without timely guidance at the point of care. An additional barrier is language: with health extension workers and communities primarily communicating in local languages such as Amharic and Afan Oromo, any technology solution would need to work across linguistic boundaries to be effective.

When a health extension worker is uncertain regarding a diagnosis or a suitable treatment plan, the resultant delay in receiving care can significantly impact the patient’s trajectory. Conversely, an unnecessary referral could burden an already-stretched health centre with a case the health extension worker could treat within the community. Last Mile Health, working closely with the Ethiopian Ministry of Health, identified the need for real-time, expert-level clinical support for health extension workers that can operate in resource-constrained environments, and subsequently engaged IDinsight as a technical implementation partner to contribute specialized AI expertise to the development of HEP Assist.

Impact opportunity

Last Mile Health leads this initiative in partnership with Ethiopia’s Ministry of Health to strengthen the Health Extension Program through AI-enabled clinical decision support. It provides direct support to improve health outcomes for millions across Ethiopia by strengthening the primary healthcare system.

In its first phase, the Health Extension Program (HEP) Assist tool, also known as HEP Assist, provides health extension workers with instant access to a call center staffed by trained health professionals who use the AI assistant to respond to queries live. When a health extension worker encounters a complex case during a household visit or at a health post, they call a toll-free number and receive evidence-based clinical guidance from a human who uses the assistant to obtain answers from over 100 Ministry of Health training manuals and service delivery guidelines. This means faster, more accurate diagnoses and treatment decisions at the point of care, reducing unnecessary referrals, cutting delays, and ensuring patients in remote communities receive the right care closer to home.

The deployment of HEP Assist across five major regions—Oromia, Afar, Sidama, South Ethiopia, and Central Ethiopia—holds the potential to engage thousands of health extension workers. Given an average catchment area of approximately 2,500, the potential reach of the tool across 408 health extension workers is an estimated 1,020,000 community members. By enhancing the precision of frontline clinical consultations, the project ensures families in rural areas receive quality care, minimizing unnecessary referrals and preserving lives through prompt intervention.

The model underpinning HEP Assist is designed to be scalable and replicable. Its architecture (an open-source language model drawing on locally vetted clinical guidelines, delivered through a call center staffed by health professionals) can be adapted for community health systems in other countries facing similar challenges. Any context where community health workers operate in remote areas with limited supervision and need real-time clinical decision support could benefit from this approach, making it a blueprint for strengthening AI-enabled primary healthcare across sub-Saharan Africa and beyond.

Our approach

Last Mile Health led the design, implementation, deployment, and operationalization of HEP Assist in partnership with Ethiopia’s Ministry of Health. IDinsight served as AI Advisor and Technical Implementation Partner, embedding senior data scientists within the Last Mile Health technical team to co-develop an AI solution tailored to Ethiopia’s Health Extension Program.

This methodology, adopted jointly by the partnership, emphasized responsiveness and handoff, ensuring that the AI-powered decision-support tool was rigorously validated for accuracy and designed for local management, thereby promoting long-term sustainability.

Key milestones

Milestone 1: Development of the AI solution, HEP Assist

Last Mile Health led the development of HEP Assist, working with Ethiopia’s Ministry of Health and IDinsight to build an AI-powered clinical advisor grounded in Ministry-approved clinical guidance. The tool draws its content from over 100 Ministry of Health training manuals, service delivery guidelines, and chart booklets. IDinsight designed and implemented the LLM pipeline and technical architecture, while Last Mile Health aligned the tool with frontline workflows, implementation requirements, and long-term sustainability. Under Last Mile Health’s technical leadership, IDinsight supported the assessment and implementation of an open-source AI model, initially Llama 3.3 70B, with the architecture designed to remain flexible, customizable, and free from vendor lock-in. This approach enables Last Mile Health and the Ministry of Health to retain long-term ownership of the platform and continue adapting and managing it as the technology evolves.

Milestone 2: Regional rollout and implementation

Last Mile Health led the deployment and implementation of the tool across five regions: Oromia, Afar, Sidama, South Ethiopia, and Central Ethiopia. The rollout involved structured training for 20 call agents and hands-on orientation for more than 400 health extension workers, ensuring they understood how to access the tool, what clinical support it could offer, and how to use it reliably and independently. Feedback from continuous surveys and qualitative interviews was used to refine the AI model to ensure it adequately served frontline needs and delivered on its promise of safe, accurate guidance.

Milestone 3: AI call center integration

Last Mile Health established and operationalized the AI-supported clinical call center. IDinsight provided technical support to integrate the AI assistant into the call agent workflow. This integration enabled health extension workers to receive immediate, automated guidance via a dedicated call agent, bridging the gap between remote communities and clinical expertise.

Milestone 4: Continuous evidence generation

The Last Mile Health team conducted continuous surveys with call agents, internal staff, and external Ministry of Health experts to track user experience and technical performance. These feedback loops ensured the tool’s advice remained safe, accurate, and truly helpful for health extension workers in the field. Last Mile Health used these findings to guide continuous improvement of HEP Assist, drawing on technical support from IDinsight and clinical input from Ministry of Health experts.

The results

• Successful deployment: Last Mile Health launched HEP Assist as a core clinical decision support resource in five regions of Ethiopia.
• High accuracy: The partners validated the AI model to ensure clinical advice provided to health extension workers is accurate, context-appropriate, and aligned with national health guidelines.
• Real-time support: Last Mile Health established and now operates a functional AI call agent system that provides health extension workers with immediate clinical case consultations, reducing the isolation of frontline work.
• Evidence-based iteration: Last Mile Health, in collaboration with Ministry of Health domain experts, completed comprehensive evaluation cycles (surveys and qualitative analysis) to improve the tool continuously based on actual user feedback from the field.
• Improved diagnostic accuracy: 19 Ministry of Health domain experts assessed the AI tool’s responses across 580 real-life clinical questions and case studies spanning 23 service areas. 91.2% of responses were rated above average, with 51.9% rated “Excellent” and 25.7% “Very Good,” exceeding the project’s original 85% accuracy target. As of 2025, endline survey data confirmed 83.3% medical accuracy in real-world use, with clinical supervisors reporting reduced unnecessary referrals and greater independence in case management over time.

Key features

HEP Assist currently:

• Supports two local languages (Amharic and Afan Oromo). Voice interaction in both languages is currently in development.
• Automatically extracts images from guidance documents and surfaces them within the chat conversation.
• Ingests and responds with training videos at the relevant timestamp.
• Includes LLM-based auto-checking of responses as a guardrail against hallucinations.

This project is ongoing. We will share further results as they become available.