HOW MANY laws and regulations are needed to govern an archipelago, home to 287m people, spread across some 13,500 islands, and stretching more than 5,000km? For Indonesia, the answer is 251,232. Yet all too often this bureaucratic morass only worsens persistent problems, from rampant graft to sluggish growth.
Indonesia is a vivid example of a challenge common to Asia’s populous poorer countries. Governments need to manage complex economies and provide services to diverse societies with limited resources. Officials across the
HOW MANY laws and regulations are needed to govern an archipelago, home to 287m people, spread across some 13,500 islands, and stretching more than 5,000km? For Indonesia, the answer is 251,232. Yet all too often this bureaucratic morass only worsens persistent problems, from rampant graft to sluggish growth.
Indonesia is a vivid example of a challenge common to Asia’s populous poorer countries. Governments need to manage complex economies and provide services to diverse societies with limited resources. Officials across the region have long turned to digital technology for help making their operations more transparent and effective. Yet Indonesia has launched more than 27,000 digital public-sector platforms and apps, to little avail. Now they hope artificial intelligence can do better.
AI models could help boost the very capabilities that poor states often lack. They are adept at analysing lots of data, from tax filings to farm yields, at a relatively low cost. Properly adapted to local conditions, AI could allow developing countries to do “in a decade what might otherwise take a century”, the World Bank argues.
Public-sector experimentation with AI in poorer parts of Asia has so far been promising, albeit limited. Consider the courts in Pakistan, which face a backlog of more than 2m cases. In 2024 trial-court judges were given access to JudgeGPT, a generative-AI assistant grounded in Pakistani statutes and court judgments. The tool helped judges with legal research, summarising and drafting. Those who used it cleared around 6.3% more cases a year. Greater speed did not mean lower quality: appeal rates did not increase. Similarly, in India, where there is an even bigger legal logjam, Adalat AI, a startup, claims that its transcription tools and document-processing software have reduced case times by 30-50% across 11 Indian states.
The biggest gains may come from helping governments make better predictions, says Indermit Gill, the World Bank’s chief economist. Rampant theft and poor infrastructure make it hard for electrical utilities in poor countries to estimate demand. AI can help them take giant leaps, says Mohak Mangal, the co-founder of Pravah, a startup that is helping utilities in six Indian states deliver power more efficiently. In Telangana, a southern Indian state, a recent study found that AI-enhanced monsoon forecasts helped farmers reap savings of as much as 42,000 rupees ($560) in a cropping season, the equivalent of three months’ minimum wage in the region.
Overloaded officials hope AI can help meet a perennial challenge, directing welfare benefits more accurately. For example, a big food-assistance programme the Indonesian government runs reaches less than 50% of its intended recipients, according to SMERU Research Institute, a local think-tank. One pilot project in Banyuwangi, a port town in eastern Java, used AI to analyse data from different government databases to help verify recipients’ eligibility; a government-commissioned study found that the new system reduced the time taken to register recipients from 75-200 days to less than one. The government hopes to roll out the initiative to the rest of the country by the end of the year.
But wider adoption poses risks. Poverty can be dynamic, specific to local conditions and difficult to quantify, point out researchers at SMERU. More than 9,000 households challenged the AI’s initial categorisation in Banyuwangi. Researchers in Bangladesh tested whether a machine-learning model trained on mobile-phone use could examine data on phone top-ups, call patterns and mobility to improve the targeting of a cash-transfer programme. Only 32% of the households it identified were eligible, so more than two-thirds of the programme’s funds went to households that should not have received them.
Such findings point to the bigger challenge poor countries face when deploying AI: the places that could most use the technology to supplement state capacity are the same places that lack capacity to deploy the technology safely and effectively, and to develop models of their own. In a World Bank survey of officials from nearly 60 countries, around 80% of those from lower-income countries said they had no tools, or only basic ones, for evaluating the performance of their AI experiments.
Governments may not like the idea of running essential public services on technology controlled by foreign powers, yet most poor countries must rely on frontier models developed in America or China. The World Bank recommends using interoperable systems drawing on suppliers from different countries and starting by adapting simpler models to local conditions. Mr Gill thinks parts of Asia, such as India’s southern states and countries in South-East Asia, are well placed to do so.
To succeed, poorer states will need to focus on less glamorous back-end processes. This includes basics such as improving data quality, training officials and procuring technology sensibly, says Shahrukh Wani, who heads AI partnerships at the International Growth Centre, a British research institute. JudgeGPT is a case in point: the striking gains appeared only when access to the model was accompanied by targeted training on how to use it; when judges were handed the tool alongside a generic legal course, they were no more productive than before. Unless states lay the groundwork, AI could end up only deepening the bureaucratic morass.
For exclusive coverage of Asian politics, economics and security, sign up to Asia Bulletin, our weekly subscriber-only newsletter.
One Subscription.
Get 360° coverage—from daily headlines
to 100 year archives.
Archives
HT App & Website