I Build AI Tools for Grantmaking – Here’s What They Revealed About Equity
By Michael Abdullahi
There is a conversation happening right now across philanthropy, EdTech, and education policy that is moving faster than it should.
The conversation goes like this: AI is transforming how we work. It is making grant review faster, learning more personalised, and resource allocation more efficient. The sector that adopts it first will have an advantage. The communities that benefit from these systems will receive better, faster, more targeted support.
That conversation is not wrong. But it is dangerously incomplete.
And the part it is missing is the part that matters most.
The efficiency trap
I manage the Daily Grant Programme at The Pollination Project Foundation, where we distribute $365,000 in annual grants to grassroots changemakers across more than 100 countries. I also research generative AI in education – with three peer-reviewed publications examining how AI-powered tools reshape learning outcomes, teacher practice, and access to quality instruction.
In both roles, I have built AI tools. I have seen what they can do. The system I trained at TPPF reduced our evaluation reporting time by 35%. That is real. That matters. A faster pipeline means grassroots leaders wait less time for funding decisions that can determine whether their work continues or stops.
But here is what that same tool also showed me.
When I trained it on our historical evaluation data, it learned our patterns. It learned what a strong grant application looks like – according to us, based on who we had previously funded, according to the criteria our rubric had historically rewarded.
And our rubric, like every rubric in philanthropy, was not written in a vacuum. It was written in a sector that has systematically favoured applicants who write in fluent, formal English, who lead registered organisations with audited financials, who have prior grant histories, polished proposals, and professional networks that connect them to funders before they ever submit an application.
The AI did not create that bias. But trained on our data, it was ready to perpetuate it, at scale, at speed, and with the false authority of algorithmic objectivity.
The same pattern, faster
I review 280+ grant applications every week. The applications that perform best in our system share characteristics that have almost nothing to do with the quality of the proposed work. They are written in sector-standard language. They use phrases like “theory of change” and “MEL framework” and “stakeholder engagement”, language that signals fluency with philanthropic norms rather than depth of community impact.
The applications that struggle, even when the underlying work is extraordinary, often come from community leaders writing in their second or third language. From first-time applicants who have never had to package their work for a Western donor audience. From grassroots organisers whose impact is documented in relationships, trust and years of sustained presence, none of which fits neatly into a dropdown menu or a 500-word narrative box.
When we layer AI onto a system that already disadvantages these applicants, we do not solve the problem. We automate it. We make exclusion faster, more consistent, and harder to challenge because it now carries the legitimacy of a system rather than the accountability of a human decision.
This is the part of the AI conversation that philanthropy is not having loudly enough.
What I learned in a classroom
Before I managed grantmaking pipelines, I stood in front of a classroom of 5th-8th graders in Boston.
My research on AI in education found the same pattern at the classroom level that I now see in grantmaking. When we surveyed 250 teachers about generative AI adoption, the finding that surprised us most was not about performance metrics. It was about trust.
Teachers did not trust AI tools to account for what they already knew about their students, the child who always looked down when she was confused but never raised her hand, the student whose engagement dropped every October for reasons unrelated to the curriculum. That knowledge- relational, embodied, accumulated is not in the dataset.
And the students most likely to be misread by an AI system were the same students most likely to be misread by a standardised test, a deficit-framing intervention, or a grant rubric that rewards polish over substance.
The communities most in need of support are often the least legible to the systems we build to support them.
That is not a coincidence. It is a design failure. And it is one we keep reproducing.
What needs to change
I am not arguing against AI in education or philanthropy. I use it every day. It has genuine potential to reduce administrative burden, surface patterns human reviewers miss, and extend the reach of under-resourced teams doing important work. But potential is not the same as outcome. And right now, we are adopting AI faster than we are auditing it.
Here is what actually needs to change:
First, bias audits before deployment. Every AI tool adopted in a grantmaking or educational context should be tested for differential impact across applicant demographics, languages, and organisational profiles before it goes live, not after it has already filtered out the communities it was supposed to serve.
Second, practitioners at the design table. The people who understand what exclusion looks like at ground level – programme officers, teachers, community organisers, grantees, etc. need to be involved in designing the systems meant to serve their communities. Not consulted once, but present throughout.
Third, human override by design. AI should flag, surface and accelerate, but the decisions that determine who gets funded and who gets taught need to remain accountable to a human being who can be asked to justify them. Algorithmic neutrality is a myth. Accountability is not optional.
Fourth, community-defined success metrics. We cannot audit AI systems against equity goals we have not clearly defined. And we cannot define equity goals without asking the communities most affected by our decisions what equitable looks like to them.
None of this is technically difficult. All of it is institutionally inconvenient.
Which is exactly why it rarely happens.
The question underneath the question
At its core, the debate about AI in education and philanthropy is not really about technology. It is about who gets to decide what counts as a good student, a strong proposal, a fundable idea, a worthy community. Those decisions were never neutral. AI does not make them neutral. It makes them faster and harder to see.
I work at this intersection because I have been on both sides of it as a grantee who rarely applied, as a teacher whose students were routinely underestimated, as a researcher who has seen what the data does and does not capture, and as a practitioner who builds the tools and lives with what they reveal.
The question is not whether AI will transform education and philanthropy. It will. It already is.
The question is whether we will be honest and brave enough to build it differently from everything else we built.
I believe we can. But only if we stop treating equity as a feature to add at the end and start treating it as the architecture we build from the beginning.
(Michael Eneye Abdullahi is Senior Manager of the Daily Grant Programme at The Pollination Project Foundation and a researcher in AI, education equity, and philanthropic grantmaking.)
