AI policy in developing countries

You can listen to this podcast on Spotify, Apple Podcasts, or wherever else you get your podcasts. You can also watch this conversation on YouTube and we have started a Substack.

Being a technology minister right now is hard. The information space is overwhelmed with contradicting policy recommendations, timelines are uncertain, the geopolitics are messy, and there are no clean historical analogies to lean on. We are truly in uncharted waters.

For policymakers in developing countries, the challenge is compounded: every policy decision about AI must be made with even less evidence, tighter budgets, and higher opportunity costs.

Long story short, I’m glad I’m not the one deciding whether to invest in sovereign AI capacity or plug into global platforms, planning how many data centres to build, or trying to think through what retraining the workforce would actually look like.

In this episode of Ideas in Development, we are joined by Umar Saif, a computer scientist by training, former minister of Science and Technology, and IT, in Pakistan, and now an AI company founder – the perfect CV to grapple with these big questions. We discuss what he learned from his time in government, why the rush towards sovereign AI capacity may be a costly distraction, his worries for the future, and where he is optimistic.

Technology was never the hard part

Pakistan is a textbook case of what Umar calls a data desert. Despite having a biometrically verified national ID linked to every land purchase, bank account, car registration, and large expenditure, the country has one of the world’s lowest tax-to-GDP ratios.

“The problem is not AI as much as it is the availability of data, and for these data silos to finally begin to talk to each other.”

These silos are not an accident. Umar describes how data is guarded by bureaucracies that derive power from controlling it, and his experience digitising land records in Punjab makes this clear. When the team began putting ownership records into a live database, they discovered that cultivable land in the province appeared to increase by 20% at each wheat procurement cycle – local elites had been capturing government subsidies meant for small farmers by manipulating paper records.

Vaccination coverage was a similar story in Pakistan. Government officials believed they were running an adequate immunisation programme until the team began logging individual vaccinations with GPS coordinates, using cheap smartphones.

“As soon as we put it on a map, we realised that there was only 18% geographic coverage of Punjab in terms of vaccinations.”

Umar highlights that for this work, cheap, off-the-shelf systems were always sufficient. The real obstacles he faced in government were:

1. Local politicians protecting information arbitrage
2. Bureaucracies resisting accountability
3. Large international technology companies selling expensive licensed platforms when simple open-source tools would have served just as well

“The challenge wasn’t making a computer system and a live database of records. That was pretty off the shelf… the real challenge was convincing these two big stakeholders that we really need to do it.”

Crop yield prediction, tax compliance, and flood forecasting are all areas where AI can genuinely help, but in each case, the prerequisite is data that is accessible, structured, and not locked behind bureaucratic walls. Tech/IT/AI ministers who jump to building data centres or technology parks before tackling data availability are making a serious error.

What AI can do with the right foundations

Despite the constraints, Umar sees genuine potential in a handful of AI applications that are well-matched to developing country problems.

Crop yield prediction is one. When Pakistan’s cotton crop was destroyed by an unexpected pest, the agriculture secretary walked into a government meeting with the bad news after the fact.

“I kept wondering, it should be possible to be able to figure out before that happens if that was going to happen.”

Using freely available satellite imagery from the European Space Agency’s Sentinel-2A, Umar’s team began building prediction models that could provide early warning, a critical tool for a country whose textile exports, and therefore economy, depend on a single crop.

Education is another area of great potential. Pakistan’s rural schools are full of multi-grade classrooms, where children at different levels share one teacher. AI could enable personalised learning paths, sorting students by ability and guiding each group through appropriate material. But Umar is clear-eyed about the gap between the concept and delivery. Would it require localised data centres trained on local textbooks? Would smartphones actually be used as intended?

Healthcare triage is a third area. In a country with severe doctor shortages, an AI agent on a phone could serve as a frontline health worker, advising on prenatal care, flagging emergencies, telling a parent when a child with diarrhoea needs hospital treatment. The technology is nearly there. The question now is whether it can be made cheap, accessible, and trustworthy enough to work at scale.

The sovereignty trap and other mistakes tech ministers make

Umar is critical of several policy responses he sees developing country governments gravitating towards.

The first is the rush to build data centres. Some investment in local compute makes sense, particularly where there is excess power generation capacity that could be co-located with data centres. But the instinct to build large-scale infrastructure often reflects political incentives rather than a realistic assessment of what’s needed.

The second is the pursuit of digital sovereignty. Umar is sceptical that most countries can or should try to train their own frontier models or build their own semiconductor supply chains.

“I’m not too high on this conversation of digital sovereignty or AI sovereignty because I think that’s a very short-term view.”

With open-weight models improving rapidly, including from China, he argues that the smarter play is to fine-tune existing models with local data and context rather than building from scratch.

The third mistake is the belief that workforce retraining can fully offset AI-driven displacement.

Patient zero of AI job replacement

The Fiverr-style service jobs that workers in economies like Pakistan feed off, things like logo design, document translation, basic coding, remote radiology, are exactly the tasks AI is already doing well. And the labour cost arbitrage that made those jobs viable is precisely what AI is set to erode.

“all the jobs that you would have for the masses in developing countries because of the labor arbitrage that they used to benefit from will likely get automated.”

And the standard political response, retraining the workforce to be AI-ready, does not survive scrutiny.

“There’s no such thing. AI is going to do things that they were traditionally taught in school or in college. I can’t think of a skill that AI would not have.”

This is a more extreme argument than the standard displacement narrative. Umar is not saying that new jobs will eventually replace old ones. He is saying that the economic model of taxing labour to fund the state is itself under threat, and that developing countries with large, low-skill workforces face this reckoning sooner than most.

“The one person I want to talk to is the finance minister, because the unemployment is going to go through the roof shortly everywhere, but certainly in countries like Pakistan.”

What a tech minister should actually do

So, what is Umar’s main advice to a technology minister today? Here’s what he would prioritise:

1. Sort out your data. National AI policies should prioritise breaking open data silos, improving data quality, and creating the regulatory conditions for AI systems to access the information they need. This is not a shiny new project, but it is often the binding constraint.
2. Prepare for the labour market shock. The jobs at most immediate risk from AI, repetitive cognitive tasks done cheaply across borders, are disproportionately concentrated in lower-income countries. This is not some threat far off in the future (Umar’s timeline is three to five years), so tech ministers need to be sounding the alarm to other senior politicians.
3. Strategic investments in infrastructure. Umar argues for co-locating data centres near excess power capacity, and using open-weight models rather than expensive proprietary systems, which can then be fine-tuned for local languages and contexts.

Umar’s experience highlights that the biggest obstacles to AI working for development are the same as those that have held back technology adoption for decades. In government, political economy, bureaucratic resistance, and elite capture, all need navigating.