MUSYG · AI ADOPTION

Evidence and references

Evidence note: AI in grant administration

Check what a source measured and whether its result applies to your task.

Estimate my scenario

Last checked: 19 August 2026; screening and fundraising evidence added on 18 November 2026. Version française.

This note bounds the synthetic A2 grant-dossier case. It separates administrative handling from a decision that affects access to funding.

Useful anchors

  • In Candid's 2025 survey, 1% of 529 responding foundations reported using generative AI to screen applications or assist decisions. Summaries and administrative tasks were described as more acceptable uses. This measures reported practice, not productivity.
  • Degrees of Change uses a central system to handle more than 1,000 applications and support 150 volunteer reviewers. Staff review extracted information and matching recommendations. The supplier case gives a useful workflow analogue, not a causal time estimate.
  • The National Zakat Foundation case reports an 80% reduction in disbursement delay after several data, automation, and agent changes. Treat it as a high scenario, not a forecast.
  • Swiss FDPIC guidance explains the information and human-review rights that can apply to fully automated individual decisions.

The worked case therefore lets the agent check a published checklist, prepare a packet, and perform writes after approval. It does not score merit, vulnerability, or success probability. A high productivity scenario can still be explored for a standard digital sub-flow, while offline, ambiguous, sensitive, and new-program cases stay visible in the whole-workload denominator.

For a real pilot, measure active human time, corrections, escalation, critical errors, complaints, appeals, reversed decisions, access by channel and group, and the downstream outcome. A nominal human review is not enough unless that person has the evidence, time, competence, and authority to disagree.

Screening and fundraising evidence (18 November 2026)

No study measures active human time for a typical nonprofit task (a grant application, an impact report, a donor appeal) with and without AI. What exists measures decisions, agreement or donations.

Study Design and sample Finding Use here
Lane et al., MIT Solve screening Randomized, 48 solutions, 228 screeners, 3,002 decisions (2024) AI advice lowered pass rates from 68% to 62% (score only) and 57% (with a written rationale). Decision time capped at 2.5 minutes, so no time result Risk of anchoring on AI; no time saving
la Caixa prescreening Hidden parallel trial, then operational use, about 500 proposals per call Reviewers rescued 30 of 216 (2022) and 35 of 98 (2023) flagged proposals; two funded projects would have been excluded Every exclusion needs human review
Thelwall et al., UKRI proposals 2,267 past proposals scored by language models Correlation with individual reviewers 0.186, below reviewer-to-reviewer agreement of 0.242 Retrospective test before use
Rospocher, controlled proposals Synthetic proposals with identical content Lower scores for weaker publication record and lower-tier institution Test prestige bias
Caffier et al., donation appeals Online experiments, about 650 participants each, small real donations Undisclosed AI appeals did as well or better than human ones Context; disclosure changes the result in other studies

Funder practice is split. MacArthur does not use AI to decide or assess grants. Foundation Scotland uses AI to shortlist, as guidance only, with mandatory human review and public disclosure. In the UK, 13% of funders ask applicants to declare AI use (Charity Digital Skills Report 2026, 123 grantmakers). A declaration field in the application form costs little and gives the funder its own data.

To print or save as PDF: Ctrl+P (⌘P on Mac).

Source and history · GitHub