Remittances, gig work, and AI · AI, Ethics, and Society

Most Nepali AI conversation, when it gets around to labour at all, stays inside the office building. It talks about the BPO floor, the agency desk, the software shop. That is the visible part of the question. The much larger part — the part that shapes more lives, more household budgets, more municipal economies in places like Khotang and Pyuthan — is invisible from inside the office building, because it happens either on the other side of the Gulf or on the back of a motorbike at midnight in Kathmandu. Two channels. Both already changed by AI. Both barely discussed.

Channel one: the destination labour market

Roughly a quarter of Nepal’s GDP arrives, every year, in the form of money sent home by Nepali workers in Saudi Arabia, the UAE, Qatar, Malaysia, and South Korea. Nepal Rastra Bank tracks the number; it has hovered around twenty-five per cent of GDP for years, and in some recent years has pushed higher. That figure is not a statistic. It is what pays for the cement on half-built houses in Sindhupalchok, the school fees in Surkhet, and the medical bills in Tanahun. It is the foundation under entire district economies.

Most of that money is earned in jobs that current AI does not directly replace — kitchen work in a hotel kitchen, security shifts at a compound gate, construction on a Doha site, line work in a Malaysian factory, agricultural labour on a Korean farm. The worker has to be physically there. The model is on the wrong side of physics, again. So the naive read is that remittances are safe.

That read is wrong, because the destination labour market is not safe. A hotel in Riyadh that uses an AI scheduling and inventory system needs fewer assistant managers and fewer back-office staff, which means it can absorb the cost of running the floor and still hire one less migrant worker on the next visa cycle. A logistics warehouse in Kuala Lumpur that introduces robotic picking does not directly replace a Nepali warehouse hand on day one — but it shifts the demand curve, slowly, year over year, until the next contract is for fewer workers at lower wages. A South Korean factory that adopts AI-driven quality control does not need to fire its existing Nepali line workers to reduce its intake; it just opens fewer slots in the next Employment Permit System cycle.

The shock, when it arrives, does not arrive as “AI took my job.” It arrives as: the recruiter has fewer slots this year. The wage on offer is the same as it was three years ago, in nominal terms. The contract is for nine months instead of two years. The household in Dailekh that was planning on a son’s remittance to finish the house gets less, later, and for fewer years. Nepal does not control these decisions; they are made in cabinets in Riyadh and boardrooms in Seoul. But pretending that the absence of control is the same as the absence of exposure is itself a policy choice.

Channel two: algorithmic management on Nepali streets

The second channel sits closer. Open Pathao or inDriver in Kathmandu, request a ride during evening peak, and you are inside an AI system whether you think of it that way or not. The price the rider sees, the route the rider is offered, the order the rider is dispatched in, the score the rider accumulates from accepted and rejected jobs, the threshold at which the rider gets deactivated — all of these are decisions made by algorithms, on data the platform owns, with logic the rider cannot see. The same is true on Foodmandu for the delivery rider, and increasingly on other domestic platforms that have followed the same playbook.

This is not technically new — algorithmic pricing has existed since the first Uber surge — but the practical asymmetry has sharpened. The platform sits on years of trip data, can simulate the response of the rider population to a price change before it ships one, and can adjust dynamic pricing in real time across thousands of riders. The rider sits on a phone. The rider can see one price at a time. The rider does not know whether the surge applied uniformly, whether new-customer discounts are being subsidised by lower per-trip rates, whether the route they were offered is the most efficient or the most platform-margin-friendly.

When a platform tweaks its pricing algorithm and rider earnings drop by fifteen per cent, that is not weather. That is not an emergent property of the technology. That is a decision, made by people inside the platform, optimising for a specific outcome — typically platform margin or customer-side growth — and implemented through code. Calling it “what the algorithm did” is the same evasive move as calling a wage cut “what the market did.” Both descriptions hide the decision-maker behind the description.

The asymmetry, and what it costs

The structural feature of algorithmic management is the information asymmetry. The platform has the data; the rider does not. The platform has the model; the rider does not. The platform has lawyers who can frame the rider as a contractor — and therefore not subject to labour-law protections — while exerting more granular control over the rider’s working day than most formal employers ever did over their employees. The rider has a star rating and a deactivation threshold.

The cost of this asymmetry shows up in measurable ways. Riders work longer hours for the same take-home, because the per-trip rate has crept down while the dispatch is structured to keep them on the road. Riders absorb fuel-price shocks that the platform passes through, while platform commissions on each trip stay stable. Riders who try to organise — to compare notes about pricing changes, to coordinate a refusal of below-threshold trips — face the practical problem that the platform can deactivate the most active organisers without ever having to fire them, because nobody was ever hired in the first place.

None of this is unique to Nepal. The same pattern has played out in Jakarta, Manila, Lagos, and Mexico City. What is specific to Nepal is that the policy framework around it is essentially absent. Nepal does not yet have a meaningful gig-work regulation, does not require algorithmic transparency from platforms, and does not have a sectoral body for riders that platforms have to negotiate with. The space is open, and what fills it will be decided in the next few years.

Reading both channels as one question

The honest way to read these two channels together is that AI’s effect on Nepali workers is mostly indirect. It comes through changes in the demand curve in destination countries that Nepal does not control, and through changes in pricing and dispatch algorithms on domestic platforms that Nepal could regulate but currently does not. Neither channel will produce a single dramatic moment that the news cycle can name. Both will produce slow shifts that compound — fewer foreign labour slots, lower per-trip earnings, longer hours for the same take-home — that show up in household budgets in Khotang and Bara before they show up in any national statistic.

The unified policy frame is this: any AI-driven change in working conditions — whether implemented by a Saudi hotel chain, a Malaysian factory, or a Kathmandu ride-hailing platform — is a working-conditions change. It deserves the same scrutiny, the same negotiation, and the same regulation as any non-AI change to working conditions would. The algorithmic costume does not change what is underneath.

Check your understanding

Quick check

—

A Kathmandu ride-hailing platform updates its dynamic-pricing algorithm. Over the next quarter, average rider earnings per hour drop by fifteen per cent, while the platform's revenue per trip rises. The platform describes this as a “technical optimisation.” What is the most accurate framing?

It is a neutral efficiency gain — the algorithm simply found a better equilibrium and there is nothing to govern here.
It is a transfer of value from riders to the platform, dressed in technical language; it is a change to working conditions and should be governed the same way any other change to working conditions would be — through disclosure, negotiation, and, where appropriate, regulation.
It is bad luck for riders that they happened to be on the platform during the change.
It is a problem only if the riders are formally classified as employees.

What comes next

The work and economy chapter ends here, but the questions it raises do not. If automation is reshaping who earns what, and if platform algorithms are reshaping how earnings are determined, the next pressure point is the information environment those decisions get made inside. The next chapter — Truth, media, and trust — turns to what AI is doing to news, evidence, and the shared facts a democracy needs to argue from. The labour story and the media story turn out to share a structural feature: in both, an asymmetry of information is the lever, and the question of who controls the lever is political, not technical.