BCG: The AI Model Is Only 10% of What Makes an AI Program Actually Work

A working AI model accounts for only about 10% of what determines whether an AI-driven program actually succeeds, according to a new BCG framework for scaling AI in social-impact work across low- and middle-income countries. Compute and data account for another 20%, BCG estimates, while people, process, and the surrounding ecosystem make up the remaining 70% — the inverse of where most organizations spend their attention. The framework, authored by Jim Larson, Abhik Chatterjee, Adham Abouzied, Vrishika Melanta, and Dan Grimm, lays out eight essential elements needed to scale a program: national strategy and orchestration, the AI models and applications themselves, accessible compute, quality local data, purpose-driven capital, builder talent, sustainable delivery channels, and an understanding of user behavior.

Two case studies carry the argument. In Rajasthan, an AI-enabled education program reached roughly 3.5 million students with personalized learning recommendations paired with teacher integration, producing a 10%-12% stabilized year-over-year reduction in children two or more grade levels behind. In Telangana, a government data exchange drew more than 200 stakeholders and 420-plus applications from over 240 startups, selecting six winners now deployed across transport, education, and health departments. BCG’s core prescription: stop asking what AI to build and start identifying the binding constraint blocking ecosystem-wide scaling.

That framing lines up with what Bain is telling go-to-market and investment teams about the region where a lot of this scaling has to happen. Bain’s Southeast Asia outlook forecasts 4.8% average annual GDP growth for the SEA-6 economies through 2035, but with sharply diverging trajectories by country, and names AI adoption — alongside institutional resilience and energy security — as one of three policy priorities now shaping which countries capture the upside. Both firms are converging on the same point from different angles: in emerging markets, AI adoption is becoming a structural growth variable, not just an enterprise productivity line item, and the constraint on realizing it is the surrounding institutional capacity, not the model itself.