AI for Public Good: Where Should We Build First?

In late 2027, the Minister of Planning of a coastal lower‑middle‑income country stared at a map he had seen too many times that year.

Clusters of red dots marked schools and clinics damaged or destroyed by the latest monsoon floods and cyclones. Blue dots marked facilities that had narrowly escaped serious damage but had been forced to close for weeks. Grey patches on the map—rapidly growing peri‑urban belts—had almost no dots at all.

At the cabinet table, he had promised three things:

• No child would be forced out of school for months every time it rained heavily.
• No pregnant woman would need to cross flooded roads to reach a functioning clinic.
• The country would “build back better”—not just rebuild what was there, where it was.

Parliament had approved a US$3.5 billion Social Infrastructure Resilience Program over eight years, with financing from a development bank and a coalition of climate funds. The program would:

• Retrofit or reconstruct 1,000+ existing schools and clinics in high‑risk zones.
• Build 500+ new facilities in underserved growth areas.
• Co‑invest in supporting infrastructure: access roads, drainage, small bridges, solar and backup power.

But the money was not nearly enough to do everything at once. Choices had to be made about where to invest first.

The Data Problem

For the first time, the Ministry had access to a centralized infrastructure data platform built with support from an international consortium. It pulled together:

• Project and asset records from multiple ministries (education, health, public works) and donor portfolios.
• Historical disaster loss data (flood depths, wind speeds, damage reports) from the national disaster agency.
• Population and poverty layers from the statistics office and international datasets.
• Basic school and clinic performance indicators (enrollment, utilization, staffing) where they existed.

On paper, this “national Canvas” of projects and assets was exactly what the Minister had always wanted. In practice, it exposed uncomfortable truths:

• Many facilities had no formal coordinates or consistent IDs.
• Loss and damage reporting was patchy and politically influenced.
• There were dozens of half‑completed or stalled projects—foundations poured, no roof; clinics never staffed.

The consortium proposed going further: deploying a domain‑specific AI assistant, powered by a platform called ConstructChat, trained over the country’s infrastructure records and global project benchmarks. Instead of manually merging spreadsheets and shapefiles, planners could ask questions in natural language and get:

• Ranked shortlists of candidate facilities by risk and need.
• Maps overlaying hazard, population, and service gaps.
• Benchmarks from similar countries: how they prioritized school retrofits or rural clinics.

The Minister liked the idea—but also knew that whichever districts came first and last would shape political narratives for years.