Why it Matters

The world's largest development agency closed its doors, taking with it six decades of evidence on what works—and what doesn't—in global development. Before the lights went out, AI helped salvage the most important lessons.

What Happened

USAID shut down in July 2025, erasing access to over 100,000 project evaluations representing $30 billion in learning across health, education, governance, and humanitarian response. DevelopMetrics, a social enterprise founded by former USAID economist Lindsey Moore, deployed their AI system DELLM (Development Evidence Large Learning Model) across the full archive before it disappeared.

The Big Picture

Traditional learning systems failed because of three human limits—cognitive overload, poor timing, and misaligned incentives. DELLM processed what no human team could: every evaluation, cross-referenced and clustered to reveal patterns that kept repeating across decades and continents.

How it Worked

The model broke each report into chunks, asking standardized questions about interventions, outcomes, and lessons. Expert coders had trained DELLM on thousands of manually tagged excerpts, creating a taxonomy of development work. The output: a ranked map of what USAID learned repeatedly for 60 years.

The Five Key Lessons

1.  Bring delivery closer to households. Programs succeed when decisions happen where people actually live—the farm, clinic, or school—not in distant offices. Rwanda's village agriculture committees let farmers directly question officials about fertilizer and budgets, creating feedback loops that adjusted district priorities.

2. Practice changes practice. Workshops don't stick. Guinea's agricultural program succeeded because field advisors were coached one-on-one on actual farms until new routines became muscle memory. Uganda's midwives "graduated" only after performing emergency procedures correctly multiple times under supervision.

3. Design for scale, not pilots. Most pilots die when extra budgets disappear. Angola's ProAgro Program required local cooperatives to co-finance their own service hubs from day one. Multiple hubs eventually self-financed because ownership was built into the design, not added later.

4. Co-creation beats consultation. Projects last when implementers share real power from the start. Central America's energy projects let communities choose committees, set tariffs, and train technicians. Madagascar's villages selected their own activities and monitored progress.

5. Strengthen the middle layer. The "middle layer"—teachers, nurses, cooperative leaders—is where policy meets people. Vietnam's Clean Air project equipped schools with air-quality sensors and trained teachers to interpret data. Within months, they organized clean air days and convinced 17,000 households to ditch coal stoves.

What's Next

Organizations need "AI Learning Officers"—not consultants or dashboards, but standing functions that ask "What did we learn last time?" The evidence isn't asking for novelty; it's asking for discipline to build, repeat, and refine what already works.

The Bottom Line

AI gave us tools to expand human learning beyond cognitive limits. There's no longer an excuse for not knowing what we already know. The responsibility to remember—and learn—now belongs to everyone else.

Go Deeper

• Read the peer-reviewed methodology behind DELLM's analysis in Land Use Policy