發展中國家能否應付數以百萬計的 AI 智能體？

Big Tech’s newest personal agents — Meta’s Muse and OpenAI’s Dots — promise to book tables at restaurants, buy groceries, and carry out other types of digital grunt work for users.

Consumers are opting in. Muse has been downloaded 9 million times in the U.S. and Canada since its launch on September 8, market intelligence company AppMagic, told Rest of World. As many as 1.8 million people use the app daily. 

Meta aspires to a global launch. “[We] are quite excited to bring personal agents to billions of people around the world,” Alexandr Wang, Meta’s chief AI officer, told the Financial Times.

Some startups are already building agents in markets that Big Tech hopes to reach. For instance, South African startup ManyHands offers Manny, a personal assistant designed to order groceries, book appointments and make payments through WhatsApp. But adapting agents to local conditions has its challenges. Anton Moulder, ManyHands’ founder, told Rest of World that getting the AI assistant to converse in all of South Africa’s 11 official spoken languages has been difficult.

Rest of World spoke to AI researchers and other experts to understand how these personal agents would perform outside Western nations.

—

Would AI agents understand the local language and culture?

English speakers could benefit more from AI agents, which run on top of large language models that underperform in many non-English languages. 

“Linguistic inclusion is ultimately about power,” Payal Arora, professor of inclusive AI cultures at Utrecht University and the founder of Inclusive AI Lab, told Rest of World.

Meta has said that expanding Muse internationally would require support for local languages and integration with local services.

AI agents would also need to understand the local culture, including the unique ways people communicate and bargain. “In many parts of the Global South, meaning is deeply contextual, shaped by family, community, religion, local institutions, and informal social networks,” Arora said.

Without that understanding, agents might impose Western assumptions on their users, Usha Raman, an independent researcher on AI policy and digital culture, told Rest of World. “The danger then is that existing forms of cultural and economic hegemony continue and the rest of the world is forced to adapt,” she said.

What happens when millions of agents go online?

Agents searching and completing transactions for millions of people could stress networks and websites. “Internet connectivity may be poor [in developing nations], affecting not only the user’s interaction with AI, but what agents can do online,” Kentaro Toyama, professor of community information at the University of Michigan, told Rest of World.

A single task given to an agent can trigger repeated exchanges between the agent, its AI model, and different websites. Cisco describes agents as “network power users,” generating up to 450% more traffic per task than a human in its testing. Agent traffic is growing online: Daily requests from AI agents on Cloudflare’s network increased by more than 1,700% over the past year.

Rising traffic wouldn’t immediately affect regions that have spare network capacity, such as East Africa, Kyle Spencer, executive director of the African IXP Association, told Rest of World. But other places with poor networks, such as Niger, Burkina Faso, or eastern Congo, could face problems, he said.

The bigger challenge would be keeping connections reliable. Road construction often damages fiber cables, Spencer said. The network could also get disrupted if undersea cables bringing internet into a nation get damaged.

Though Muse and Dots could continue working as they are cloud-based, they may be unable to reach necessary services. They also wouldn’t be able to reach users who lose connectivity for instructions or approvals to complete a task.

Could personal agents widen Africa’s digital divide?

Only 36% of Africans used the internet in 2025, compared to 92% of Europeans and 88% of North and South Americans, according to the International Telecommunication Union. City dwellers in Africa were 2.6 times more likely to use the internet than people who live in the countryside.

“If agents were to work in Africa, they might unfortunately deepen inequality,” Jane Munga, a fellow at the Carnegie Endowment for International Peace, told Rest of World.

The cost of accessing agents could also limit who benefits, Munga said. Though Muse can be accessed for free, many other agents are paid. Dots is available for a minimum of $100 a month. “These are costs that most Global South users may not be in a position to absorb,” she said.

How would local businesses need to rewire for AI agents?

Local businesses would have to make their websites understandable, readable, and usable by bots. If they do not, they risk losing out on digital economic opportunities.

This would be challenging for small, offline businesses, such as a vendor at a vegetable market. They would have to go online, maintain an inventory, know how much stock they own, and create ways for bots to pay, Munga said. This requires money and skills.

Businesses may also have to invest more in customer support. “Agents might purchase more, and request more returns — when human users decide agents erred — putting stress on businesses,” said Toyama.

Could regulators hold agents accountable?

Agents perform better the more they know the user, Mercy King’ori, policy manager for Africa at the Future of Privacy Forum, told Rest of World. A user’s favorite restaurants, shops, and regularly visited places could help an agent build a picture of their tastes, habits, and daily life.

“An agent may draw conclusions about a person or a group, such as their likely income, health, creditworthiness, or behaviour from the data it handles,” King’ori said.

In cases where agents are developed and hosted outside the country, regulators may be unable to supervise the systems, said King’ori.