Testing the limits of AI deep research

The experiment

I used Abacus AI's Deep Agent to conduct research on what I thought was an unsolvable problem: finding small family foundations willing to support grassroots voter turnout efforts. And you know what? I was right. It is an unsolvable problem.

My starting assumption was that small family foundations would be a good fit for us as a small, all-volunteer organization: smaller grants, fewer reporting requirements, more flexibility. Although the organization supports voter turnout across the country, we thought family foundations might support efforts in their states / communities.

The mistake

I hoped that AI could wade through mountains of 990s and annual reports to find a donor providing just what we needed. But like any experiment — you can't think your way out of a faulty hypothesis. I was too focused on what we needed, and not focused enough on what these kinds of donors are willing to fund.

---

What I should have known

• Philanthropy is based on relationships — prospecting donors is just not strictly an information problem, and not well suited for AI research.
• Funding for voting issues is almost entirely funded through large national organizations and networks.
• Family foundations are locally focused, and not interested in geographically diverse organizations.