Data infrastructure and institutional readiness as the foundation for AI impact

On the sidelines of the 2026 Skoll World Forum at Oxford Town Hall, IDinsight convened AI-for-Good pioneers to examine what it truly takes for AI-based social interventions to deliver on their impact promise, starting with the upstream data infrastructure that makes them work.

Meg Battle, Senior Director of External Affairs and Global Strategic Partnerships, delivered the opening address. 

Meg acknowledged that the AI-for-good sector has grown rapidly, with real applications in healthcare, agriculture, and humanitarian response, but has had deeply uneven results, with many deployments failing to reach the people they were designed to serve. Two patterns consistently separate impactful AI projects from those that fall short. First, data fundamentals matter: AI tools are only as good as the data infrastructure beneath them, and fragmented, incomplete, or siloed data limits even the most sophisticated applications. Second, shared public infrastructure accelerates progress; tools like knowledge graphs could dramatically improve AI relevance in sectors like education, but are not being built fast enough to keep pace with deployment. With ODA budgets shrinking and foundational data programs being cut, she urged actors in the sector to urgently ask how to sustain these unglamorous but critical investments.

“The AI tools that get attention are only as good as the data quality and infrastructure underneath them. If that infrastructure is fragmented, incomplete, or inaccessible, the AI’s usefulness is limited regardless of how sophisticated the technology is.” 

 

— Meg Battle, IDinsight

The panel discussion was moderated by IDinsight’s Mallika Sobti. The panellists were four practitioners who have built impactful AI systems: Rikin Gandhi; CEO and co-founder of Digital Green who’ve developed an AI system supporting over a million farmers across Africa, Asia, and Latin America; Becky Faith, a Senior Research Fellow at the Institute of Development Studies and the UK Foreign, Commonwealth and Development Office, a leading voice on how AI and digital systems shape development outcomes; Sid Ravinutala, IDinsight’s Chief Data Scientist who has built a wide range of AI-for-good solutions for education, health and agriculture; and Anubhav Arora, Co-Executive Director, Programs & Platforms, Noora Health, a remote caregiving tools that keep patient families connected to care across India, Bangladesh, Indonesia, and Nepal.

Key Insights

To begin, Becky Faith acknowledged the funding pressures that are pushing organisations to adopt AI in contexts where it doesn’t fit or where privacy isn’t adequately protected, but invited leaders to interrogate whether AI adds real value to their program before they make a decision about investing in it.

Start with product fundamentals, not infrastructure

Anubhav Arora’s opening provocation reframed the whole conversation. Noora Health didn’t start with data pipelines. Building data infrastructure before defining what you’re actually measuring is a costly mistake. You need to know your users’ journey, the funnel stages, and which metrics leadership cares about before a single data warehouse is set up. The plumbing only makes sense once you know what you’re trying to measure.

“Even before building the data infrastructure, I’d begin with product fundamentals. What are your participants’ real pain points and how will your tool serve them? What format should your AI tool take?”

 

— Anubhav Arora, Noora Health

Instead of building a chatbot, Digital Green designed entirely around the farmer’s specific decision moment, grounding their AI tool’s responses in local agronomy, crop stage, and recent weather. Rikin emphasized that defining the exact intervention, before engaging the technology, was the decisive early move.

“For us, the key early decision was being very clear we were not building a generic chatbot. A farmer might send a voice note saying, ‘my maize leaves are turning yellow this week,’ and unless we could respond in that exact moment with something locally relevant, nothing else mattered.” 

 

— Rikin Gandhi, CEO Digital Green

Data infrastructure is the silent bottleneck

Sid Ravinutala described a telling moment when a senior government official watched a live LLM demo of a data analysis that took minutes. He then turned to his staff and asked why the same task takes them weeks. What the senior bureaucrat didn’t realise was that AI doesn’t replace the work of chasing, cleaning, and stitching together siloed datasets. The demo’s magic was contingent on clean, structured, queryable data existing in the first place. Without it, even the most capable model is bottlenecked by the same organisational dysfunction it was meant to solve.

“The ‘AI does your analysis instantly‘ promise only holds if you’ve already done the hard organisational work of getting your data into shape.”

 

— Sid Ravinutala, IDinsight

Evaluation and learning are fundamentals

Rikin emphasized the value of evaluating the impact of AI tools and what approach they found to be most valuable for this purpose. Digital Green uses Agency Fund’s four-level evaluation framework, assessing model quality, product experience, user adoption, and real-world impact, because each layer surfaces different failures. A simple counting error once inflated their reach numbers in Nigeria and no model benchmark caught this, underscoring how blind single-layer measurement can be. Organizations need to explore not just whether their model works, but what the return is on investing in the infrastructure around it.

Funders need to back more than just the innovation

All panellists agreed that more investment is needed in generalizable tools. Both funders and implementers need to think beyond their own use cases and build tools that can be utilised by others in the community working on similar problems.

Funders were challenged to go beyond tools and models and fund data infrastructure as the foundation for AI’s impact. Anubhav credited an early Snowflake partnership with giving Noora Health the push to consolidate data infrastructure they never would have prioritised otherwise. Rikin put it plainly: the highest-return investments for Digital Green weren’t in the model. They were in curating agronomy knowledge, building feedback loops from millions of queries, and ensuring voice-based accessibility. 

Becky mentioned FCDO’s partnership with the Canadian Government to build safe, development-focused AI ecosystems in Africa and Asia, bringing together funders with different priorities. This collaborative’s ability to avoid duplication, embrace plurality and complement each other’s work is both rare and consequential.

“Funding innovation alone is not enough. You also have to invest in talent, data, and responsible AI practices. Our approach is fundamentally about partnership and ecosystem-building to ensure all these pieces develop together.”

 

— Becky Faith, IDS/FCDO

Lessons for practice

• Interrogate AI’s fit before investing. Funding pressures are pushing organisations to adopt AI where it doesn’t belong. Leaders should first ask whether AI genuinely adds value to their program and whether they can deploy it responsibly.
• Start with product fundamentals, not data pipelines. Before building any infrastructure, map your users’ journey, define funnel stages, and align leadership on the right metrics. The data plumbing only makes sense once you know what you’re measuring.
• Clean data is a precondition, not a given. AI’s promise of instant analysis only holds if the underlying data is already structured, accessible, and queryable. Organisations that haven’t done the hard work of data consolidation will hit the same bottlenecks, regardless of the model.
• Measure impact at every layer. Model accuracy alone is insufficient. A technically sound model can still produce a confusing product that fails to change behaviour. Evaluation frameworks need to assess model quality, product experience, user adoption, and real-world impact separately.
• Funders need to back the infrastructure, not just the innovation. The highest-return investments are in knowledge bases, feedback loops, accessible interfaces, and shared tools, not just the model itself. Ecosystem-building approaches that pool funder priorities and avoid duplication are both rare and necessary.