Foundations often sit on a wealth of qualitative evidence. Across grant proposals, progress reports, evaluations and grantee stories, the lessons, outcomes and innovations that matter most for strategy are documented in detail.
However, this rich evidence base is often scattered, narrative-heavy and difficult to read across at portfolio level and thus goes unused.
Traditional approaches to qualitative data analysis and use rely on foundation teams and evaluators reading and coding reports manually. These approaches can be slow, narrow in scope, and place a heavy burden on already stretched teams.
Read: the evaluator’s guide to generative AI
Over the past two years, Itad has been learning how to build AI-powered systems that change this. These systems extract, classify and synthesise qualitative insights from grant documentation so that foundations can learn from what they already hold.
How AI is changing the way we learn from data
Recent advances in large language models make it feasible to analyse text at scale while preserving nuance. What used to take months of reading can now happen in hours across hundreds or thousands of documents, expanding the volume and range of evidence that teams can review and compare and making space for greater sensemaking and interpretation.
Two recent Itad projects demonstrate the impact of this well:
- We are supporting a foundation to extract results and lessons from its extensive grants database; our system integrates content from the database to generate storyline summaries and thematic syntheses, as well as providing a chatbot style interface that allows users to query the extracted data.
- We have developed an AI-driven grants assistant for a major global health fund that lets staff ask plain-language questions across their portfolio and surface cross-portfolio trends in minutes rather than weeks.
In both cases, our tailored AI solutions are leading to faster, deeper learning. They helped surface patterns (for example, common success factors across different grants) that were previously hidden in plain sight. They are also freeing up staff time from laborious manual reviews, allowing evaluators and program managers to focus on interpreting findings and making decisions.
Importantly, these projects are building confidence that AI-driven analysis can be done in a secure, ethical, and user-friendly way in a philanthropic context.
Lessons for building useful AI-powered systems
Design your learning architecture before the AI architecture
Clear analytical thinking remains essential, if not even more important, when using AI. The systems that work best are the ones where we have invested in designing the analytical framework up front. Without clarifying learning questions and taxonomies aligned to a foundation’s strategy, AI extracts a lot of information that is not very useful.
Instead, investing in a strong analytical framework ensures AI outputs are recognisable to the people who need to act on them.
Although the AI integration is new, the underlying discipline for evidence-based decision-making is not. We have spent over four decades helping organisations design frameworks and approaches that enable this – integrating AI now allows us to do it at speed and scale.
Trust, transparency and humans-in-the-loop are essential
By this point, we are all aware of the fallacies of AI’s credibility, and organisations are right to ask hard questions about how far AI outputs can be trusted. Hallucinated summaries, opaque reasoning and one-shot outputs that no one has checked are not good enough for strategic use.
To mitigate this, we design our systems so every insight links back to its source document and paragraph. AI-generated outputs must be reviewed and verified by researchers through a human-in-the-loop model, and uncertainty is flagged rather than hidden.
It’s also important to be clear with partners about what these systems are not: they do not substitute for humans making sense of what findings mean in context. The technology produces material for collaborative sensemaking — it does not replace it. Itad is adept at helping partners reflect on evidence and understand what works, why, and how to adapt; using AI-supported analysis helps broaden the evidence base to do so.
Build in partnership, iterate in sprints
Useful AI-powered systems are not built in one pass. They need to be developed through a partnership-based, iterative process, starting with pilot analyses and refining prompts, rubrics, and workflows through short sprints.
This means evaluators, subject-matter experts and foundation teams work alongside developers throughout. Their feedback helps tune the analysis, correct errors and false positives, and ensure the system reflects the language, evidence needs and decision-making context of the people who will use it.
This collaborative approach makes the final solution better calibrated, easier for staff to use and interpret, and more likely to become part of everyday learning practice rather than a separate technical product.
Integrate AI into existing learning workflows
AI-powered systems are most useful when they are designed around the places where learning and decision-making already happen. Rather than delivering a tool and expecting teams to change their behaviour around it, the system needs to feed into existing cycles such as strategy reviews, board reporting, learning events and portfolio reflection sessions.
This means designing outputs in partnership with programme, impact and learning teams so that the analysis is timely, recognisable and actionable. For example, AI-generated thematic syntheses can be used to support learning briefs, strategy conversations or board papers, making insights easier to use at the moments when decisions are being made.
Embedding AI in this way also requires new capabilities. Involving foundation staff in developing taxonomies, testing prototypes and interpreting outputs helps build confidence and ownership. With the right support and training, teams are better able to maintain, question and evolve the system after the initial engagement ends.
What’s next for AI analysis?
AI-powered qualitative analysis is moving from an interesting experiment to a dependable resource. Foundations now have a real opportunity to turn dispersed grant reports, proposals, evaluations and partner narratives into a structured, searchable and traceable evidence base.
This allows Foundations to see patterns that would otherwise remain hidden: which approaches are gaining traction, where bottlenecks are recurring, which partners or geographies are showing promising signals, and where strategies may need to adapt. This creates a stronger basis for portfolio reviews, board conversations, grantee dialogue and strategy refreshes, helping teams move from anecdotal learning to more systematic, evidence-informed decision-making.
Ultimately, the value of an AI-powered system is not simply that it makes analysis faster, but that it helps foundations act with greater confidence on the issues that matter most to them. By making evidence easier to access, compare and use, these systems can support sharper strategic choices, earlier course correction and more focused conversations with partners. Used well, AI can help foundations direct attention and resources towards the approaches, relationships and opportunities most likely to advance their mission and achieve greater impact.
How can Itad help design AI-powered systems?
If you are considering an AI-powered qualitative analysis project we’d love to discuss how Itad can support you, including: what is feasible, what good looks like and our experiences and learning.
Get in touch with us:
