When people talk about AI in education, they usually mean AI in classrooms: devices, chatbots, adaptive platforms. The conversation quickly stalls because many schools in low- and middle-income countries lack the infrastructure. So, AI becomes something for well-resourced systems, and everyone else waits.

But what if the problem is not lack of infrastructure, but lack of organization?

World Bank President Ajay Banga draws a useful distinction between “big AI” (massive compute, specialized talent) and “small AI”: practical, task-specific tools that run on everyday devices. Small AI is already transforming agriculture and healthcare in developing countries. In education, it could do the same, not necessarily by putting devices in classrooms, but by changing how ministries produce and deliver what shapes learning in schools.

Ethan Mollick tracks this across industries: AI is a productivity breakthrough for individuals, but organizations haven’t captured the gains yet. His challenge: stop waiting. Reorganize around what exists.

Something We Saw Firsthand

Working in a country in Latin America, we faced a universal problem: teachers need to know not just which students are behind, but what and why. That requires diagnostic questions where each wrong answer reveals a specific misconception. Most systems don’t have them, because producing thousands of aligned items is slow, expensive, and rarely prioritized.

Using Claude Code, a tool that lets non-developers give AI instructions in plain language and have it read documents and produce structured outputs, we tried a small AI approach. Four teachers wrote seed questions from real classroom errors. Claude Code read the 225-page national textbook, learned its terminology, and scaled those seeds to 3,950 diagnostic questions. Every step had a human decision maker: experts chose which errors mattered, reviewed what AI produced, and rejected what didn’t hold up. Six weeks. Four people.

The counterfactual is not “handcrafted questions by a large team.” It was that teachers would continue starting the year without knowing where each student stands, relying on intuition. Small AI did not replace an expensive process. It replaced the absence of one.

The Imagination Gap

What struck us was that most education institutions don’t know this is possible. They might think Claude Code is for developers. IT and curriculum sit in different meetings. Nobody has put small AI in front of the people who run education processes. Mollick calls this the gap between individual capability and institutional adoption. In education, we’d call it an imagination gap:

• Regulatory incoherence. Countries accumulate thousands of overlapping normatives across government levels. Feed those 10,000 regulations to Claude Code: it flags contradictions, identifies duplicates, produces a simplified compliance guide in days.
• Ministry workflows. Give Claude Code the procedure manuals: it maps workflows, identifies bottlenecks, suggests streamlined alternatives. Not replacing judgment but showing where time is wasted.
• Teacher training that fits the school. Most ministries deliver one training to all teachers because designing differentiated versions takes more time than any team has. Claude Code can adapt core training content for each school profile, adjusting examples and focus areas. Training stops depending on what the ministry has time to develop and starts depending on what each school needs.
• School reporting reimagined. Principals spend 76% of their time on paperwork. What if they sent a WhatsApp voice note about how the week went, and AI transcribed it into the compliance report the ministry requires? Better data, because principals report honestly instead of copying last month’s template. A team with Claude Code could prototype this in weeks and test it before any procurement begins.

What Would It Take?

Mollick argues organizations capturing AI gains need three things: leadership that decides to experiment, a lab that turns ideas into solutions fast, and a crowd using these tools daily. Most ministries have none of the three.

Nobody has a proven playbook. But a starting point is modest: one person who is AI proficient (not a software engineer, not just AI literate, but someone comfortable enough with Claude Code to turn a real problem into a working prototype) paired with people who know the problem deeply, a curriculum specialist, a training coordinator, a supervision lead. Take one task: turning 200 school visit reports into a regional summary that identifies which schools need support and why. Run it with Claude Code alongside the staff who normally do it. Compare. AI suggests, humans evaluate. Any pilot must follow proper data privacy protocols, and starting small is how you manage that risk. If results hold up, invest further. If not, the cost was weeks, not years.

Small AI, Big Change

The grand vision of AI-powered classrooms may—and should—arrive gradually, grounded in evidence on what actually improves learning. That will require full connectivity, devices, cheaper data processing, and deep AI literacy among teachers and students. Small AI requires none of that from schools. It does require AI-proficient people inside the institutions that support them: people comfortable enough with the tools to turn a real education problem into a working prototype and working with the cross-disciplinary team that know well the subject and the students and teachers involved. The binding constraint is not technology. It is whether institutions can create the conditions for those people to work together with the right support, so that solutions do not only perform well but have a real impact on learning.

Mollick’s question for CEOs applies here: will education systems build the organizational capacity to capture these gains, or keep debating the grand vision while everyday work stays stuck in the last century?

“Credit: World Bank Group. All rights reserved”