My AI Donor Advisor's Report Card
Often as good as a human, but not a replacement.
My AI donor advisor is a folder of 1,000 files, linked together like Wikipedia, storing everything that crosses my desk as a funder: pitch decks, meeting notes, update reports, email threads, research studies, and notes about key ideas. Claude Code does the work of organizing and linking things together.1
Sounds cool, but does it work? Mostly it’s amazing, but like everything AI it’s “spiky” — great at many things, but sometimes humorously bad.
The Report Card
A: better than the best person. B: on par with a capable human. C: some useful elements. D: just adding steps.
Funding inputs: B+
The advisor researches with the depth of a full human team. It reads FEC filings and Forms 990, and does internet searches on the project and its key people. The corpus holds 350+ research studies, and every pitch gets matched against them.2 But unlike a human advisor, it hasn’t had years of conversations learning which people and organizations have done well or poorly.
Funding evaluation: B
My evaluation process is more complex than I’d realized; it covers 10 topic areas spanning 86 questions, with 51 red and green flags at the end. The advisor runs the full playbook on each project, which lets me focus mental energy on key items rather than parsing.3
Many of the advisor’s observations and questions are ones I would have generated myself — as if written by someone deeply familiar with my work. But it also generates lower-quality output, which is visibly related to its limited data about the world.4 Plus it’s regularly too cryptic or verbose.
Funding judgment: C-
On actual decisionmaking, it’s not there yet. It recommends funding almost everything tied to my interests, tweaked to focus on some particularly aligned area.5
Remembering: A
Vetting a pitch through the advisor yields a useful byproduct: a tracker of asks, key questions, and next steps. Eighteen months ago I passed on a pitch as basically “good story, no numbers.” Last week that project came back through a mutual contact. Instead of half-remembering my objection, the advisor pieced together the full history: what was asked, what I gave instead, and the exact question I’d left on. My prep for the new meeting: have the numbers arrived?
Analyzing tactics: B+
I had a half-formed idea for a new persuasion tactic, and asked the advisor what it would cost per net Democratic vote. The advisor interviewed me one question at a time, then produced a memo and spreadsheet on par with what I’ve seen from sophisticated analysts.6
Collaborating: B+
Smart donor advisors circulate slates of their highest funding priorities. These lists encode where they see the best leverage in the moment; my playbook encodes where my marginal dollar goes furthest. I ran one slate through the AI advisor and got back the intersection: the handful worth a further look.
Drafting emails: D
I’m on draft seven of a three-paragraph email. A friend sent me a pitch that was vague on one key detail, and the system could not get the reply right. Finally I wrote it myself. The re-drafting process helped me identify principles I follow but had never written down such as “understand before judging,” which the advisor can now use for next time.
Things it still doesn’t do
The advisor doesn’t connect me to people I’d enjoy meeting. Lots of key information surfaces only in conversation, not in written materials that the advisor can find. The advisor can’t do exploratory meetings where it unpacks an idea to find promising seeds. It can’t put the right questions to the right people. Some of this is bottlenecked on me, some on the technology.
Conclusion
Everything above is a huge improvement from out-of-the-box AI, which was “not remotely what I wanted.” I expect these grades to rise because every miss teaches the advisor something for future runs (the “I know Kung Fu” moment in an earlier essay), and the technology is advancing fast.
Human advisors may worry about their future, but I think they end up equally valuable in an AI world, just focusing on different tasks than today.
What’s the report card for your own AI usage so far?
The broader term for this setup is a “second brain,” and I use it for much more than political work. I made a shareable starter kit for those interested; email or message me for a copy.
The studies’ mundane use is to guard against overoptimism, e.g. flagging if a proposal uses unreasonably large effect sizes. The more exciting use is figuring out what new things might work; for any new idea, I can quickly start building from the existing evidence. To facilitate that, all 350 studies are cataloged on multiple axes: tactic, election salience, target population, etc. See the section “Analyzing tactics” for an example.
In retrospect, it’s no wonder that evaluating projects (a) took a lot of energy, and (b) always came with a sense I was missing something. Instantly running a funding opportunity through this playbook is just so cool.
The advisor’s output is noticeably richer and more accurate with more information about a given organization. This is why I wrote that organizations should start packaging their information for funders’ AI agents.
I expect that this is fundamentally a data issue. I haven’t yet given it full information on what I’ve historically funded or passed on, nor asked it to holistically consider a full portfolio.
The only reason I don’t rate this more highly is that I can’t be sure that all of its numbers are as accurate as a human expert in the domain, but to my eye the numbers are absolutely good enough for a back-of-the-envelope analysis.