AI Can Now Draft Grant Reviews — But Humans Still Decide
Your nonprofit's funding application might get a robot's first read
Every year, foundations hand out billions in grants, and the people who decide who gets the money are usually volunteers drowning in applications — forms, budgets, attachments, letters. They have to apply the same detailed scoring rules to every one, and explain their reasoning so colleagues can check it. It is slow, repetitive, and easy to do inconsistently. A research team from Russia built a system called SAGE to take on the first pass.
SAGE (short for Schema-Guided Aspect-Based Grant Evaluation) turns a grant's scoring rubric into a structured checklist. Then, instead of just spitting out a score, it links every judgment back to specific evidence in the application — the sentence about the budget, the paragraph about the team. It also sorts its own statements into three buckets: confirmed, disputed, or unaddressed. That means a reviewer can see not just what the AI concluded, but why, and where it may have gotten something wrong.
The test results are the interesting part. The team ran SAGE on 35 real nonprofit grant applications and compared its output to 105 reviews from the original competition. Its agreement with humans was only "fair" — a score called kappa of 0.29, where 0 means pure guesswork and 1 means perfect agreement. Then the foundation's reviewers did a second round, with SAGE's draft in front of them. In that assisted round, agreement jumped to 0.58, more than doubling a simpler AI approach that scored just 0.33. The pattern is clear: the AI's drafts were not great alone, but genuinely useful as something humans reacted to and corrected.
The catch is that 0.58 is still only "moderate" agreement, not reliable enough to trust unsupervised. The study covered one foundation and 35 applications — a small slice of a huge, varied world. And there is a real risk that busy reviewers start rubber-stamping AI drafts instead of thinking hard about them. Since grant decisions shape which charities survive, the human signature still has to mean something.
- SAGE turns a grant's scoring rules into a checklist and shows which parts of the application support each judgment — an auditable paper trail, not a black box
- Human agreement with the AI rose from a 'fair' 0.29 to a 'moderate' 0.58 once reviewers worked alongside its drafts, beating a simpler AI method at 0.33
- The AI never decides anything — it produces a rough draft that experts correct, meaning grant reviewers could save hours while staying accountable
Why It Matters
Faster grant reviews could mean smaller charities hear back sooner and spend less on paperwork.