Case studies of AI adoption in developing countries.

 

Agriculture is the primary source of sustenance for billions of people around the world and, in many developing countries, employs more than half the working population.Agriculture is well suited for AI-powered productivity improvements because of its high volumes of unstructured data, reliance on labour and complex supply chain logistics, as well as the significant number of farmers who would value customized services that are not locally available. Rural agricultural areas are typically short of the prerequisites for AI adoption (e.g. electricity, Internet access and digital literacy). Despite these challenges, the following case studies demonstrate how AI can be used in three main agricultural applications in developingcountries, with significant impacts on the yield and quality of crops, as well as the livelihoods of farmers (table II.2).

                                                                                            

Globally each year, pests and diseases decimate up to 40 per cent of the world’s crops, causing substantial detriment to farmers. Effectively addressing such problems requires specialist knowledge; it can take years of experience to diagnose infestations in a timely fashion and apply appropriate treatments. Such expertise is generally in short supply, particularly in areas in which smallholding farmers do not receive agricultural extension services. With the use of AI in agriculture, however, expert information can be made instantly available to any farmer who has a mobile telephone. In Colombia, the International Centre for Tropical Agriculture, for example, has developed a mobile application that helps farmers diagnose infestations of banana plants using photos of crops, called Tumaini, which means “hope” in Swahili.

Tumaini uses a deep-learning–based computer vision system that has been trained on thousands of images of banana plants, both healthy and infected, and labelled by agricultural experts, providing the algorithm with comprehensive visual references in order to identify unique patterns indicative of crop diseases, which are often too subtle for untrained eyes to detect. A farmer uploads a photo of the plant and the application provides an instant diagnosis and suggests dedicated countermeasures. Tumaini can detect five diseases and one pest with an accuracy of above 90 per cent, giving farmers a diagnostic capacity comparable to that of highly trained experts. The application is also available in offline mode, although there may be some loss of accuracy, and can therefore be widely used even in rural areas that lack reliable Internet access. To date, Tumaini has been downloaded over 10,000 times in 15 countries across Africa, Latin America and South-East Asia (Tumaini, 2024). Crop diseases in developing countries can also be addressed with the use of GenAI-powered chatbots. MkulimaGPT, for example, created for farmers in the United Republic of Tanzania, is a large language model that has an elaborate sensor-based disease detection system for maize. The chatbot is delivered through a commonly used mobile messaging app, to facilitate diffusion among local farmers. A farmer uploads a photo of the crop, which is cross-referenced with an internal database and, if the application detects an abnormality, it initiates a chat session, offers a diagnosis and guides the user through the appropriate action, thereby significantly lowering the skill barrier for the average maize farmer. One limitation of deploying large language models in developing countries is a lack of training data in local languages. To address this with regard to MkulimaGPT, the developers have obtained funding from a private charitable foundation, to collect high-quality local data and build a chatbot that speaks Swahili, to ensure that the chatbot is tailored to local needs.

Another common application of AI inagriculture is in predicting local crop yields in order to allow farmers to make informed financial and management decisions about their crops. Such use also offers Governments accurate data needed in monitoring and ensuring food security. Conventional data collection methods for crop yields, such as field surveys and aerial imagery, are costly and difficult to scale. In addition, traditional statistical methods struggle to capture the many complex factors that contribute to yields, such as climate and soil conditions and crop genotypes. The ability of AI technology to jointly analyse different data from unconventional data sources can help unlock new opportunities. Drawing upon and analysing free opensource data, AI can generate reliable cropyield predictions. Researchers at Beijing Normal University, for example, have used AI techniques with three open-source datasets to estimate the yield of rice crops. Their model relies on climate and soil data from Google Earth Engine, historical crop yield data from official publications and open-access satelliteimagery, all of which are readily accessible on the Internet; open-source data can thus help fill gaps when local data are sparse.

 Once models have been calibrated and key information pre-processed, AI can offer an accessible and effective solution. Compared with traditional regression models, the deep neural network proved more efficient in extracting crop-yield–related features from the data, with up to 88 per cent accuracy compared with only 42 per cent when using traditional regression models. When used with data from China, the new model enabled accurate predictions of rice yields at the county level, covering 94 per cent of the rice cultivation area. This case study shows that the use of AI can open new ways to use data for accurate crop-yield prediction in low-resource conditions. In addition, in China, researchers from the South China Agriculture University have applied machine-learning techniques to images from unmanned aerial vehicles, to predict yields of cotton. Compared with satellite imagery, imagery from such vehicles offers higher resolutions and can thus facilitate yield predictions at a much more granular level, even individual fields. As in the previous study, the deeplearning model achieved significantly higher accuracy than one based on linear regression, namely, 80 per cent compared with 66 per cent. Such a model may be particularly helpful for smallholding farming communities that need to plan harvests and choose which crops or activities to invest in.  

One of the most important resources for agriculture is water, which is often scarce. According to the Food and Agriculture Organization (FAO), 1.2 billion people live in agricultural areas with very high levels of water stress, mostly in developing countries. In recent years, the problem of water stress has been exacerbated by climate change, with the increasing intensity and frequency of droughts.  

These impacts can be alleviated by a combination of AI and other technologies. In Tunisia, for example, there have been regular severe droughts, the impacts of which are aggravated by intense agricultural production. Agriculture accounts for over 70 per cent of the country’s freshwater withdrawal; it is therefore both the main cause and a casualty of water shortages. The issue is being addressed, for example, by ifarming, a startup founded in Tunisia in 2017 to reduce water consumption through more accurate farming. The main service of the startup is Phyt’Eau, an AI-based programme that can danalyse data on water use collected in real time through an array of IoT sensors on farms (Agritech, 2024). The sensors collect information thatmeasures water stress on crops, including on temperature, soil humidity and wind. Based on the data, Phyt’Eau prescribes an optimal irrigation management plan for the plot that, when connected to an irrigation system, can be administered automatically. In initial trials, the prototype reduced water use by 20 per cent and increased crop production by 20 per cent. IBM offered access to advanced AI and IoT platforms, and this collaboration boosted the water-saving capability of Phyt’Eau to 40 per cent and productivity by up to 30 per cent. AI is also used in precision agriculture in Malaysia, for example, where drones equipped with AI vision systems are being deployed in palm-oil plantations to spray nutrients and pesticides with speed and precision. In Fiji and Samoa, an AI-based system developed in Australia is being used for automatic weeding and pesticide spraying. These and other projects are leveraging AI with other automation technologies to achieve more sustainable and productive farming.

Manufacturing plays a key role in economic development, stimulating growth in different upstream and downstream sectors and generating significant employment opportunities. Examples from developing countries such as Brazil, China and India show how industrialization can reduce poverty and accelerate economic growth.

 Manufacturing has been subject to successive waves of technological innovation, the latest of which is Industry 4.0 technologies. Developing countries that have applied these technologies have boosted productivity and growth rates in manufacturing value added and GDP. The following case studies show how developing countries can use AI to cut costs, create better working environments and increase efficiency (table II.3).

 A major domain for AI applications in manufacturing is robotics. Over recent decades, industrial robots have automated many repetitive processes and replaced human workers in hazardous and physically demanding environments. One disadvantage is that they can be fairly rigid, generally built and programmed for particular tasks, and it is costly to adapt them to new tasks. The use of AI enables robots to be moreversatile and adaptive. In China, for instance, a technology company has developed a fully automated AI-driven robot for welding. Its deep-learning algorithm uses three-dimensional laser sensors to recognize objects in real time and distinguish between various metal parts and weld joints and it can guide the robotic arm to perform accurate welding operations. An advantage of the technology is that it can weld on shiny metal surfaces, whereas previous robots could not make the necessary distinctions due to reflections. In Indonesia, for example, More importantly, while traditional welding robots need to be reprogrammed for each new product, an AI-powered welding robot can quickly adapt to different functions and the new dimensions of incoming parts while requiring minimal human intervention. This can significantly reduce retraining costs and shorten downtimes. Within the field of AI-guided industrial robots, an emerging trend is the use of collaborative robots, or cobots. These are unlike ordinary robots in that they are designed to work in close interaction with humans. Typically, they are smaller and less expensive and have built-in mechanisms that reduce the need for additional safety fencing. Due to these features, cobots can be more readily integrated into small-scale production lines or labour-intensive manufacturing settings. AI enhances the collaborative qualities ofcobots by improving safety and by enablingthem to work in more dynamic environments.  

Addressing equipment breakdowns can be costly. Breakdowns cause delays in production and require expensive replacements of parts. They are particularly burdensome for manufacturers in developing countries where skilled technicians and stocks of specialized spare parts may be in short supply. Many of these problems can be prevented by using AI for predictive maintenance. In traditional machine maintenance, technicians carry out inspections and repairs manually, either when scheduled, or when a machine breaks down. In predictive maintenance, machinery is constantly monitored for signs of failure using IoT sensors, with data analysed by AI processors. By cross-referencing with past data, an AI processor detects patterns indicative of a future malfunction and alerts plant operators ahead of time. In Türkiye, for example, Vestel Electronics, a home appliances manufacturers, has collaborated with a university to apply machine learning to predict the remaining useful life – the expected amount of time until a machine’s next breakdown – of plastic injection moulding machines. The algorithm is trained on historical sensor data, including the clamping force of a machine, oil temperature and injection time, then analyses real-time sensor data in the factory. According to a study by the company, the algorithm correctly predicted the remaining useful life of the machines 98 per cent of the time. Equipped with this information, managers can schedule maintenance and purchase spare parts in advance, thereby lowering costs and downtimes. Predictive maintenance only requires AI data processors and a set of  IoT sensors attached to machines. It is thus versatile and adaptable to different industrial environments. In Chile, for example, large mining companies such as Codelco are using the technology to monitor the fleet of autonomous mining trucks. Smaller manufacturers can also use the technology given the increasing availability of less expensive, standardized packages.

In large-scale manufacturing, multiple AI-enabled systems can be integrated within a single plant, to provide significant gains in production, savings in energy and greater profits. The synergistic effects of AI and other frontier technologies may also enable manufacturers in developing countries to catch up with counterparts in developed countries. In India, Tata Steel, one of the country’s largest steel manufacturers, has implemented more than 250 machine learning systems across various production processes. One such application assesses the quality of welds on steel tubes. A machine-learning algorithm can automatically detect a faulty weld with more than 80 per cent accuracy and thereby significantly lower the number of defective products. The use of AI can also help optimize the chemical mix in steelfurnaces and speed up the transportation of materials within and between plants. Such improvements, combined with other digital technology upgrades, have increased the corporation’s pre-tax profits. Another example is Unilever, who has built the world’s largest laundry detergent powder factory in Indaiatuba, a municipality in the state of São Paulo, Brazil. The company has made the factory more agile and cost efficient while minimizing environmental footprint by using technologies suchas AI and digital twins, that is, virtual representations of physical objects.  

A digital twin is used with machine learning to establish the optimal process parameters for new formulations of laundry powder. Reducing the need for physical trials has accelerated the launch of innovations while cutting energy consumption. Between 2018 and 2023, the company also used AI-driven predictive maintenance to halve the cost of life cycle management for pneumatic devices. Other key use cases include a biomass-powered machine-learning spray-drying tower that has achieved a 96 per cent reduction in carbon dioxide emissions and a digitally enabled sealing solution that has eliminated chronic defects, reducing customer complaints about leakage by 94 per cent. As a result, the technologies have reduced innovation lead times by 33 per cent and production costs per ton by 23 per cent, while also reducing carbon dioxide emissions. In 2022, in recognition of its achievements in the field of advancedmanufacturing, the Indaiatuba site was listed by the World Economic Forum as one of the 29 new “lighthouse” factories worldwide. 

The use of AI offers significant opportunities for improving access to and the quality ofhealthcare services in both developed and developing countries. Many developing regions lack medical services and infrastructure, which challenges citizen well-being and poverty reduction goals. With regard to healthcare services, the use of AI can improve both access and quality. The following case studies illustrate how AI has been implemented in developing countries to provide expert diagnoses of diseases, extend the coverage of healthcare services and manage pandemic outbreaks (table II.4).

The timely and accurate treatment of diseases requires high-quality diagnostics, which are often unavailable to patients in developing countries, particularly in rural areas, due to a lack of skilled medical professionals, laboratory facilities and infrastructure. AI offers the prospect of new and cost-effective diagnostic methods and equipment in low-resource settings.  

AI can, for example, be used to diagnose perinatal asphyxia, a birth complication that leaves infants unable to breathe properly and, in developing countries, is one of the top three causes of newborn deaths. Most cases can be treated if quickly diagnosed; in developed countries, this is commonly done by sending a sample of an infant’s blood to a laboratory, for analysis of signs of low blood oxygen, a service that may be out of reach in rural areas. 

In Nigeria, a team of AI researchers has offered a novel, simple and inexpensive alternative involving analyzing an infant’s cries. Crying and breathing rely on the same set of muscles, and irregular vocal sounds in an infant’s cry are a reliable indicator of asphyxia. Such minute differences may not be heard by human ears, but can be readily detected by a machine-learning algorithm trained on a data set of infant cries. The researchers developed Ubenwa – meaning “cry of a baby” in Igbo – an AI-driven mobile application that analyses a short audio clip of a newborn’s cry and can detect perinatal asphyxia with an accuracy of 86 per cent, securing valuable time for treatment. Another example of an AI system that can enhance traditional diagnostics is a battery-powered X-ray machine with an embedded AI-driven tuberculosis screener. In countries with few expert radiologists, this can serve as a valuable tool for doctors. Unlike traditional X-ray machines, the battery-powered machines are portable and can be deployed in remote areas that may have limited electricity connections. For example, such machines are being used by health authorities in South Sudan and Tajikistan, with support from the United Nations Development Programme. In Tajikistan, 15 machines have already been used to screen 120,000 people in 2023, covering 15 per cent of the country’s total diagnosed cases of tuberculosis.

A common problem among developing countries is the inadequate coverage of medical services. The World Health Organization recommends at least 45 skilled medical professionals for every 10,000 people. In many developing countries, this figure is not reached, making it difficult to extend life-saving resources. It takes time for countries to build up their healthcare systems, but the use of AI can help allocate existing resources to those in greatest need.

Around 800 women died from preventable causes related to pregnancy and childbirth every day in 2020. These could be avoided with better health information and access to medical care during pregnancy. Armman, a non-profit organization, helps provide maternal and child health services in urban slums using mMitra, a free mobile messaging service. The service covers 3.6 million vulnerable women in India, sending curated voice messages aboutpreventative care measures during differentstages of pregnancy, to raise medical awareness and promote the health of both mothers and infants. Studies show that enrollment in the service enhances women’s maternal knowledge, enhances their voice within their families regarding their pregnancies and increases their likelihood of seeking professional medical services. However, about 40 per cent of enrolled women eventually stop listening to the messages and drop out. Due to limited resources, Armman staff cannot reach out to re-engage all of them, but are collaborating with Google India on an AI model that helps find and target the pregnant mothers atgreatest risk of dropping out. The model analyses each woman’s socioeconomic information, such as family size, income and age, as well as their call history, including call duration and missed calls, to predict those at highest risk of discontinuing and, of these, who would benefit most from the outreach service. Armman staff then allocate limited human resources more effectively and attempt to keep more women in the programme. 

After the introduction of the AI algorithm, engagement by subscribers rose by 30 per cent. This type of personalized messaging could be used in other sectors besides healthcare and help optimize the distribution of limited resources. There is also limited healthcare outreach in Kenya; for every 10,000 people, there are only 23 available medical doctors. Access Afya, a social enterprise, operates 12 small clinics using a telemedicine platform, mDaktari, that provides low- cost virtual doctor consultations. Using GenAI, the enterprise aims to reach more people. In a pilot programme, ChatGPT is integrated with mDaktari, to provide a chatbot that can be used as a preliminary screening tool. The chatbot receives patients’ inquiries, gathers information about symptoms and suggests that the patient should visit a clinic or collect medication at a pharmacy. This service saves clinics time on gathering patient information and, when appropriate, diverts individuals with mild conditions from the use of clinical services. AI chatbots are not foolproof; they cannot tell what is real and what is fake and can be prone to fabrications. Access Afya addresses the fallibility of chatbots by ensuring that human clinicians review and approve chatbot suggestions before they are sent to patients, in order to protect against mistakes. Use of the triage performed byAI allows human clinicians to focus onthose patients in greatest need. The early success of the programme shows the potential of using GenAI as an effective triage tool, to improve efficiency and extend the reach of existing medical services. With financial support from a private charitable foundation, Access Afya plans to expand the service to accommodate multiple languages and have a greater role in supporting clinician diagnoses.

As shown during the COVID-19 pandemic, managing outbreaks of infectious diseases requires providing public health administrators with accurate and up to-date information, for example, about demographic movements, transmission patterns and healthcare capacity. Equipped with such information, authorities may be better able to target interventions and bring an outbreak under control. In developing countries, structured healthcare data are often not available, particularly with regard to minority and vulnerable populations. As an alternative, the use of AI can unlock the potentialof significant amounts of unstructured data. 

In Brazil, for example, during the COVID-19 pandemic, in 2021, the Office of the United Nations High Commissioner for Refugees (UNHCR) worked with the Government on a machine-learning tool for predicting the inflow of refugees from the Bolivarian Republic of Venezuela and for coordinating resources to protect them from the coronavirus. The tool was used to predict future border crossings based on historical patterns. Since the pandemic had disrupted data collection, researchers used unconventional open-source data. These included Internet search activity on migration and border-related topics, complemented by data on COVID-19 cases and news reports on local unrest in the Bolivarian Republic of Venezuela. Sources also included bus timetables in border regions and schedules for salary payments, as an indicator of when people might have additional funds for travel.By triangulating between these sources of data, the AI model predicted the inflow of refugees one month in advance with a high degree of confidence. This helped UNHCR and local partners plan for the number of migrants that arrived when borders reopened in June 2021. By combining and analyzing significant and different data sets, AI can help inform key decisions during infectious outbreaks, using population movement models, such as in Brazil, or algorithms that forecast disease transmission or enable rapid diagnosis and contact tracing.