What automation hits first · AI, Ethics, and Society

There is a way of talking about AI and jobs that is so abstract it stops meaning anything. Some jobs will go. Some will change. Some new ones will appear. Everything depends, the analyst says, on what kind of jobs we mean — and then never says which kind. This section refuses to do that. In 2026, the line between work that current generative models eat for breakfast and work they cannot touch is actually fairly clear, and it runs through Nepal in a specific, identifiable place. Pretending the line is fuzzy is a way of avoiding the conversation; the line is not fuzzy, and the conversation is overdue.

The shape of exposure: text-in, text-out, repetitive, remote

If you want a single sentence that predicts which jobs current AI is most likely to absorb, it is this: work that is text-in and text-out, repetitive in structure, and done remotely from the person it serves. Each of those three properties matters, and the combination matters more than any one of them.

Text-in, text-out means the inputs and outputs are language — emails, tickets, transcripts, code, copy, translations, summaries, scripts. Current models are essentially very large pattern engines for language. They are extraordinarily good at producing more language that fits the pattern of the language they were shown. If your job is mostly converting one piece of language into another piece of language, your job is on the exposed surface.

Repetitive in structure does not mean boring. It means the underlying task has a stable shape. A customer-support reply has a stable shape — acknowledge, identify, instruct, close — even if the words change. A basic translation has a stable shape. A weekly status update, a property listing, a templated legal notice, a routine bug-fix request — all stable. The model has seen ten million instances of the shape; it can produce the ten-million-and-first cheaply.

Remote from the person it serves matters because if the work happens through a screen, the AI can sit on the same side of the screen as the worker, and the customer cannot tell the difference. The moment work has to happen in the room — fitting a pipe, calming a frightened parent, negotiating in a tea shop — the model is on the wrong side of physics.

What this looks like, concretely, in Nepal

Apply that test to the actual Nepali labour market and the pattern is not subtle. The most exposed categories include entry-level English-Nepali and Nepali-English translation, transcription work for international clients, templated copywriting for digital-marketing agencies, first-line customer-support chat for global SaaS companies routed through Kathmandu and Lalitpur, basic software-testing scripts, junior content-moderation triage, and the bulk of what the BPO sector in Bhainsepati, Pulchowk, and Sanepa actually sells. These are exactly the jobs that have absorbed thousands of educated young people who finished a +2 or a bachelor’s and did not want to leave for the Gulf.

The least exposed categories, in the same labour market, are the inverse. A cement mason in Banepa is not exposed; a trekking guide on the Annapurna circuit is not exposed; a tea-shop owner in Asan is not exposed. A counsellor at a mental-health NGO in Jhapa is not exposed in the substance of the work, even if some paperwork is. A municipal ward secretary in Kavre, whose job is half-paperwork and half being-the-person-villagers-trust-to-show-up, is exposed in the paperwork half and shielded in the trust half. A last-mile delivery rider for Foodmandu is shielded by the last mile, even though the dispatch is fully algorithmic above them.

The BPO question is a labour-market question, not a tech question

It is tempting to talk about the BPO sector as if it were one industry among many. In Nepal, that understates it. The BPO and IT-enabled services sector — depending on how you count and which year you trust — employs somewhere in the tens of thousands of people directly, and several times that indirectly through training institutes, recruitment agencies, real estate, and the small-business ecosystem around the office parks. The wages are not high by Bay Area standards. They are very high by Nepali standards, and they go disproportionately to women, to first-generation college graduates, and to people who would otherwise be on a labour-permit queue at the Department of Foreign Employment.

If those jobs contract by even a quarter over the next several years — through automation of the easiest tasks, through foreign clients re-shoring with AI-assisted teams, through per-seat fees that erode margins — the second-order effects show up in remittance dependence, in female labour-force participation, and in the credibility of the entire “study hard, get a desk job, stay in Nepal” social contract. This is not a technology story. It is a political and labour-market story that happens to be triggered by a technology.

What is not coming, at least not on this timeline

The forecasts that talk about AI replacing “60% of jobs” are doing a particular trick: they are counting tasks, not jobs, and then assuming that every task automated leads to one fewer worker. That is not how labour markets work in any country and especially not in Nepal. Tasks get re-bundled. A worker who used to spend forty per cent of their week on the easy emails now spends sixty per cent on the hard cases, where the customer is angry and the situation does not fit a template. Whether that is a better job or a worse one depends on whether the pay goes up to match the increased difficulty — which is the topic of the next section.

What is genuinely not coming, on the 2026 horizon, is full automation of work that has to be physically present, work that requires high-trust relational judgment, and work that depends on being-the-person-vouched-for in a community. The lines could move. They will not move evenly. And the policy mistake to avoid is treating “the lines will move eventually” as a reason to do nothing about where they sit today.

Check your understanding

Quick check

—

Which of the following jobs is most exposed to current (2026) generative-AI automation in the Nepali context?

A trekking guide leading a fourteen-day Annapurna circuit.
A cement mason on a residential construction site in Banepa.
A junior English-to-Nepali translator at a Lalitpur BPO handling templated marketing copy for foreign clients.
A municipal ward secretary in Kavre whose work depends on knowing residents personally.

Quick check

—

Why does the exposure of Nepal’s BPO and IT-enabled services sector matter disproportionately for the country’s labour-market policy conversation?

Because BPO workers are the highest-paid workers in the country.
Because the sector employs tens of thousands directly — disproportionately women and first-generation graduates — and is one of the main domestic alternatives to foreign labour migration; contraction there ripples through remittances, female participation, and the credibility of the "stay in Nepal" path.
Because BPO companies pay the largest share of corporate tax in Nepal.
Because the BPO sector controls Nepal’s internet infrastructure.

What comes next

Knowing which jobs are exposed is only half the question. The other half is what happens to the productivity gains when AI does absorb a task — do they flow to the worker as a raise, to the owner as profit, or to a foreign platform as a per-seat fee? The next section, Who captures the gains, argues that this is the question that determines whether AI is good or bad news for Nepali labour, and that the answer is decided by bargaining power and policy, not by the technology itself.