Implementing AI in education is both very similar to earlier challenges of integrating edtech — and at the same time very different. It is similar in that costly infrastructure constraints limit access, and skills upgrades and training are required for effective uptake at scale. What is notably different about AI is the speed of its development and roll-out. FOMO (fear of missing out) is an unfortunate but real driver of many efforts to quickly deploy AI in education before knowing exactly how to do it, who benefits and who does not, what the enabling conditions for success look like, and what it all costs. Move fast, and you might well make mistakes. But not doing anything risks that countries and communities already on the wrong side of the digital divide fall further behind their peers.

A new working paper from the World Bank, Costing AI Use in Education in Low- and Middle-Income Countries, identifies approaches to budgeting and financing that ministries of education will need to consider as they support teachers and learners in their use of AI. Given resource constraints and infrastructure gaps, cost-effective strategies should prioritize approaches that minimize upfront investment, manage recurrent costs, and generate policy-relevant evidence.

Rather than starting with large-scale, systemwide deployment, countries can begin with targeted applications that address binding constraints to teaching and learning while building institutional capacity.

1. Start with teachers, not students. Interventions that support teachers through lesson planning, feedback, coaching, or administrative assistance are often more cost-effective than direct student-facing tools, particularly in early stages. They require fewer devices, are feasible even where student‑level access is limited, and improve instructional quality at a lower fiscal cost per student — while generating evidence about what additional investments may be needed to support wider AI use by students.
    
2. Weigh AI costs against the high cost of the status quo. Estimates of global learning poverty — the share of 10-year-olds unable to read and understand a short text — stand at 70 percent. Expanding traditional teacher‑based service delivery is fiscally demanding. In Sub‑Saharan Africa alone, reaching SDG4 would require additional annual spending of US$6.4 billion for primary and US$38.7 billion for secondary teachers by 2030. AI-related costs should be assessed against this baseline, not in isolation.
    
3. Build on infrastructure already in place. Leveraging existing hardware, like shared school devices and current mobile networks, removes the need for massive capital investments. This shifts the focus from expensive hardware procurement to affordable, scalable software deployment, freeing up resources for teacher training and locally relevant digital content.
    
4. Rethink financing and procurement. Traditional models may need to be reimagined. Ministries of education can adapt existing funding streams like national Universal Service Funds (USFs) to subsidize cloud computing costs. Outcome-based contracting and flexible, pay-as-you-go usage models reduce the financial risks associated with fast-evolving AI tools and avoid locking systems into outdated technologies. Public-private partnerships can combine public oversight with private capital and expertise. Separating the underlying AI layer from the user-facing application layer also helps education ministries swap out AI models as newer, cheaper, or more powerful ones emerge.
    
5. Design for low connectivity and limited budgets. AI applications that function without connectivity and synchronize when bandwidth becomes available can be particularly valuable in rural or underserved areas. For example, tools that record and analyze classroom practice and deliver feedback to teachers offline. Lightweight small language models, which require less computing power and cost less to run, may offer more affordable, cost-effective approaches where infrastructure and budgets are most constrained.
    
6. Treat evidence generation as a core investment, not an afterthought. The evidence on cost‑effectiveness, scalability, and implementation of AI in education in low- and middle-income countries remains very limited. Governments and development partners can hepfully fund well-designed pilots, collect comparable data on total cost of ownership and outcomes, and create feedback loops that allow systems to adapt as technologies and prices evolve. Ministries would collectively benefit from sharing what they learn with peers. Without this, they risk making multi-million-dollar commitments to technologies that may not work, cannot scale, and could harm learners and the wider learning ecosystem.

Decisions to ‘scale up’ AI use in education should be informed by proven pedagogical approaches rather than speculative hype — at costs that are understood, affordable, and sustainable, balancing the potential for impact against what achieving that impact will cost. A follow-up post will explore how the emergence of AI changes the way ministries of education can plan for the costs of educational technology.