Nepal’s National AI Policy (2025) promises higher productivity, better public services and a workforce ready for the Artificial Intelligence (AI) economy. These are reasonable, even necessary goals for a developing country like ours. But the policy is largely silent on socio-economic questions such as what happens to employment, the tax base and the state’s ability to fund itself when machines begin to replace workers at scale.

In Nepal, unemployment already stands at around 10 percent. Tax revenue accounts for roughly 90 percent of government revenue, while remittances have recently been equivalent to 26 to 33 percent of national GDP. In the near future, if automation significantly shrinks formal wage employment, it could weaken these very pillars of our economy, leading to serious fiscal consequences. For a country that neither builds frontier AI systems nor maintains a strong social safety net, the first policy question should not be how much productivity AI can unlock, but whether it will deepen the fractures already visible in our economy.

For most of modern economic history, labour, growth and state revenue moved together. Workers produced value, consumers spent it, and governments taxed the results. AI disrupts this relationship because capital in the form of AI models, agents and robots can now perform tasks that once required skilled human time, thereby removing labour from the equation. As these systems become more capable and autonomous, this substitution may extend beyond low-skill work into high-skill sectors, precisely the areas Nepal aspires to build its future workforce around. 

Nobel laureate economist Daron Acemoglu used the term “productivity bandwagon” to explain how technological progress can improve society broadly. But for that to happen, two conditions must be met: (i) Technology must increase workers’ productivity, and (ii) Workers must retain enough bargaining power to claim a share of the resulting gains.

Nepal’s AI policy focuses almost entirely on the first condition and says almost nothing about the second. When labour becomes easier to replace, employers gain greater leverage over wages and working conditions, while a larger share of the gains goes to the owners of the technology. Something similar happened during the Industrial Revolution, when Britain went through a long stretch known as “Engels’ Pause”, during which output and profits rose while worker wages stagnated and their living conditions worsened. A similar imbalance is beginning to appear today: record numbers of applicants per opening, shrinking entry-level roles and widespread layoffs.

Nepal’s exposure to such vulnerabilities is immediate. For instance, the country’s IT and software service exports are estimated at nearly $1 billion annually, which is an important and growing source of foreign earnings, yet the sector is one of the first exposed to AI automation. IT is only one part of a much broader shift, as pressure will eventually move towards other white-collar jobs. And as AI advances alongside robotics and physical automation, it will reach a wider range of jobs that still employ large numbers of Nepalis at home and abroad. 

Meanwhile, the infrastructure that powers most advanced AI sits in a handful of wealthy economies, referred to as the “Compute north”. Nepal, like most developing economies, sits in the “Compute south”: a consumer of AI capability rather than a producer. Payments for such AI services flow outward even as the efficiency they enable reduces local labour demand. As a result, Nepal risks bearing the domestic cost of AI-driven displacement while much of the corresponding value is captured abroad.

The policy response should therefore begin with socio-economic and fiscal risks, not only with productivity targets. Three areas of action could help the government improve Nepal’s position.

First, build a real data baseline. The government should publish an annual assessment of how automation is affecting employment and output by sector. Without reliable numbers, every policy-level decision is just guesswork.

Second, protect workers during the transition. Upskilling programmes are necessary but insufficient on their own. Temporary income support, training stipends and active job-matching for those displaced by algorithmic change could reduce the human and political cost of the shift.

Finally, Nepal should explore new revenue sources that track the technology itself. As the traditional income-tax base comes under pressure, it should study mechanisms, already being discussed in several advanced economies, to capture value from AI usage at the point of consumption. Major AI platforms already meter and bill AI tokens (which is the basic unit of data processed by AI models), so the technical capacity to track domestic use exists. A tax on AI tokens consumed in Nepal could retain a share of the economic activity that occurs in Nepali territory.

Thus, Nepal cannot simply import AI strategies designed for countries with deep capital markets and extensive welfare systems. Our vulnerabilities are different and more immediate. A national AI strategy that treats employment, tax and remittance dependence as central rather than secondary concerns is the only realistic route to capturing the technology’s benefits without hollowing out the state’s ability to function. The window for that adjustment is open now; it will not stay open indefinitely.