The Bot Stops Here

✳The argument

Off-the-shelf agentic AI tools are now capable of meaningful financial and qualitative grantmaking due diligence, from extracting and reconciling data to surfacing risks, summarizing evidence, and drafting follow-up questions. As agents take on deeper analytical work, foundations need ways to preserve human judgment, skepticism, and skill development. My concept of cognitive checkpoints is a practical design approach for skill plugins and other AI workflows that keep foundation staff actively engaged during AI-assisted due diligence.

Cognitive checkpoints are deliberate pause points in an AI workflow where a human must read, evaluate, and confirm before anything continues. They keep teams in the driver’s seat instead of just watching AI drive, and they make the AI’s work clear enough to trust.

✳Code repo

An AI-assisted workflow for reviewing nonprofit financial packets and drafting a due diligence summary, gated so the agent stops for human review at each step instead of turning extraction into judgment.

• →Built around the Hollyhock Foundation's grantee review process, with assumptions you'd want to adapt before reusing it.
• →Not a finished product, and no substitute for your own security, privacy, or legal review.
• →The plugin runs Python scripts, so it only works in a desktop coding tool like Claude Code or Codex. A lighter browser-only version lives in a separate repo.