Start With the Problem, Not the Model

Institutions often begin AI adoption with a product demonstration. A model answers questions fluently, produces a dashboard or promises to automate a difficult workflow. Leaders then search for somewhere to deploy it. This reverses responsible design. The institution becomes organised around the tool before it has defined the human problem.

A better sequence is problem → people → context → baseline → smallest sufficient intervention → evidence → governed scale → revision or exit. The model comes after the mission. Sometimes the appropriate answer will be a large generative system. Sometimes it will be a narrow classifier, a rules-based service, a searchable knowledge base, an SMS workflow, better teacher preparation or no AI at all.

Starting small does not mean thinking timidly. It means keeping the problem visible, reducing unnecessary data and cost, learning under controlled conditions and preserving the institution's ability to change direction. A limited intervention that people can understand, maintain and improve may create more lasting value than a powerful system that imports dependency without solving the underlying need.

On 14 October, the World Bank Group's Annual Meetings programme foregrounded Small AI as a route to more and better jobs in developing countries. The event describes practical and affordable applications designed for local languages, infrastructure and budgets, and asks how countries can move from promising pilots to impact at scale. World Bank…