This series picks up where the four-part series on AI sovereignty ended: Who Holds the Switch?, Three Powers, Three Philosophies, The Five Sovereignties and the Hidden Bill and Pool or Be Colonised, Again!. That series asked who controls AI, the question of AI sovereignty: a country’s ability to decide how AI is built, used and governed within its borders. This one asks what a country that mostly takes AI should write into its national AI policy. Cross-cutting lenses (gender, rural people, the poorest, minorities, disability, children, faith communities and institutional capacity) are noted in each part and will be treated in full in one of the articles in the series. Bangladesh is the worked example, not the limit of the argument. The facts, figures and recommendations apply, with local adjustment, to AI takers (countries that mostly use AI built elsewhere rather than build the leading systems themselves) across the Global South (the developing and emerging countries of Asia, Africa, Latin America and the Pacific): Nepal, Pakistan, Sri Lanka, Kenya, Nigeria, Ethiopia, the Philippines, Indonesia, Bolivia and many more. Each article closes with a short “Beyond Bangladesh” section to make that link explicit, and explains key terms in plain words where they first appear.
From Switch to Statute: What an AI Taker’s Policy Must Actually Do
Sovereignty is a word. A policy is a set of decisions. Bangladesh is on its third attempt to write one.
Picture a district hospital in 2030. A young doctor uploads an eye scan to an AI tool that flags diabetic damage. The tool was bought by a ministry, runs on a foreign company’s cloud, and learned mostly from other people’s eyes. If it misses a case, who is answerable? Who tested it on Bangladeshi patients? What happens if the supplier raises its price, changes its terms, or switches it off?
No one in that room can answer. That is not a technology problem. It is a policy gap.
Bangladesh is the case study here, but the hospital could be in Kathmandu, Nairobi, Lagos, Karachi or La Paz. The question is the same for every AI taker: once a country accepts that it will depend on others for AI, what does it actually write down? A national AI policy is where sovereignty stops being a slogan and becomes rules, budgets, institutions and rights.
Third time lucky?
Bangladesh has tried before. A national AI strategy was drafted in 2019–2020. A National AI Policy draft in 2024 proposed a National AI Centre of Excellence, an advisory council, a public register of AI systems and a review every three years. It was never finalised.
In January 2026, for public comment, the interim government published a strong draft National AI Policy 2026–2030. It sorts AI systems by risk, bans mass surveillance and social scoring (rating people’s trustworthiness from their behaviour in order to reward or punish them), requires algorithmic impact assessments (checks, before an AI system is used, of whom it could harm and how), and commits to the Council of Europe’s AI convention. One analysis praised its accountability cycle, which included annual reporting, a mid-term review in 2028, and a sunset clause, a rule that ends the policy in 2030 unless it is renewed. The policy was meant to be replaced by a permanent AI Act by 2030.
Then came an election. The new government’s ICT minister told a global meeting in July 2026 that Bangladesh is drafting a “human-centric” policy. So the country now has a good draft, a new government, and an open question: will this one survive?
History is not encouraging. Political disruption explains only part of past failures. The bigger reason was the absence of binding institutional commitments. Documents were written; but, nobody was obliged to carry them out. That pattern is familiar far beyond Bangladesh. Many developing countries have national AI strategies that were launched with ceremony and never funded.
Where we start
Honesty about the starting point matters. The IMF’s AI Preparedness Index gave Bangladesh a score of 0.38, placing it 113th of 174 economies. India scored 0.49 and Sri Lanka 0.43. Bangladesh ranks 75th in the Oxford Insights Government AI Readiness Index.
The most detailed picture comes from the AI Readiness Assessment. It was led by the ICT Division with UNESCO and UNDP in late 2025. It listed 15 priority actions. Some of the key findings included data scattered across ministries; there are too few computer chips for AI; university courses are outdated; and, almost no teaching of AI ethics, with a severe gender gap in the AI workforce. One commentary on the report put it bluntly – drafts stuck in bureaucratic limbo, data locked in silos, and widening digital divides.
None of this means Bangladesh should wait. It means the policy must be designed for the country that exists, not the one in a brochure.
The AI taker’s doctrine
Countries like Bangladesh will not build the world’s leading AI models. That is not defeatism. Even rich countries are making the same calculation. The World Bank’s World Development Report 2026 urges developing countries to adapt existing AI tools to local needs rather than spend billions building from scratch.
An AI taker’s policy can rest on five plain verbs.
- Adopt. Use proven AI where it clearly helps people, starting with problems that matter most, such as health, farming and disaster warnings.
- Adapt. Make imported tools work in Bangla, on local data, under local law. A tool that does not understand the people it serves is not a public service.
- Govern. Set rules, rights and institutions that apply to everyone who sells or uses AI here, including the government itself.
- Prepare. Plan for the day a supplier raises prices, changes terms or cuts access. Previous articles discussed who holds the switch. A policy must say what happens when it is flipped.
- Pool. Work with other Global South countries on language data, testing, computing power and bargaining, because no single small market has enough leverage alone.
Five tests for any draft
How will the steering committee know whether its draft is good? A useful policy passes five tests.
• It decides rather than lists. “Promote AI in health” is a wish. “The health ministry will test every AI diagnostic tool on Bangladeshi patients before purchase” is a decision.
• It names owners and money. Every commitment needs an institution responsible for it, a budget line and a date.
• It binds the state. Governments are the biggest buyers of AI and the most dangerous misusers. Rules that exempt the state protect no one.
• It protects the weakest first. If a policy works for a Dhaka software firm but not for a farmer on a river island, it has failed most of the country.
• It survives the next government. That means grounding key parts in law, building independent institutions and creating a broad public constituency for the policy.
Honest objections
“A policy is just paper.” Often true. That is why the tests above focus on law, money and institutions, not aspirations.
“Regulating now will scare investors.” Most of what an AI taker regulates is how AI is bought and used, not how it is invented. Clear rules can attract investors who want certainty.
“Global powers will set the rules anyway.” Partly. But rules on how AI is used inside Bangladesh, in its hospitals, courts, schools and elections, are still Bangladesh’s to make. Giving that up would be a choice, not a necessity.
A lens to keep in view
Readiness scores are national averages. It underscores that a university graduate in Dhaka and a garment worker in Gazipur do not live in the same “AI readiness.” The same is true for a woman without her own phone, a Chakma speaker in the hills, or a person with a disability. While the series’ discussion will allude to each of the lenses that are pertitent, the gamut of these issues will be discussed in one of the upcoming articles.
Beyond Bangladesh
Bangladesh is the example, not the exception. On the IMF’s index, Pakistan scored 0.37 and Nepal 0.35, close to Bangladesh’s 0.38, while advanced economies scored at least 0.7. Most low- and lower-middle-income countries sit in the same band, face the same shortages of data, chips and skills, and buy their AI from the same handful of companies. The five verbs and five tests above are not a Bangladeshi recipe. They are a starting point for any AI taker, from Dhaka to Nairobi to Lima.
Back to the hospital we began with. The doctor deserves a tool that works, a regulator who checked it, a contract that protects the country, and a patient who can ask why. None of that is out of reach. It just has to be written down, paid for and enforced.
Political and International Affairs Analyst; Research Fellow, Bangladesh Neeti Gobeshona Kendra (Policy Research Center, NGK)





