Industries / Government, education and community / Nonprofits / Foundations
Reviewing applications
Foundation staff and outside reviewers sort, score and choose among the applications they receive.
5
Milestones achieved, from 2008 to 2022.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2008
Counting the paperwork
Project Streamline, a joint effort of grantmakers and grantseekers, published Drowning in Paperwork, Distracted from Purpose, a study of grant application and reporting practices based on a survey, interviews, published research and data. It recommended shared data repositories and common platforms for applications and reports.
Credit
Project Streamline
2017
A lottery among the best
The Volkswagen Foundation added a partial lottery to "Experiment!", its grants of up to €120,000 for 18 months: an outside jury chose about 15 to 20 proposals in each round, and about as many again were drawn by lot from the others it had judged fundable. Over eight calls it received 5,051 applications and funded 183. Comparing the two groups, the foundation found "very similar research outputs and outcomes", and the lottery funded more early-career researchers and more women.
Credit
Volkswagen Foundation; EVACONSULT, Berlin (accompanying research)
2019
Blind review was not enough
Julian Kolev, Yuly Fuentes-Medel and Fiona Murray studied Gates Foundation research grant proposals from 2008 to 2017, reviewed with the applicants' names hidden. Women still received significantly lower scores, and the gap disappeared once the authors accounted for the words used in proposal titles and descriptions, where men and women differed in their use of broad and narrow words. The wording that reviewers rewarded did not predict later innovation, and women's scientific output rose more than men's after they were funded.
LessonHiding names does not remove a bias that travels in writing style.
Credit
Julian Kolev, Yuly Fuentes-Medel and Fiona Murray (National Bureau of Economic Research working paper)
2020
Grants in 48 hours
Fast Grants, run under the Mercatus Center at George Mason University, launched in April 2020 to fund COVID-19 research; its first grants went out within 48 hours and later rounds within two weeks. It raised over $50 million and made over 260 grants in 2020. In its founders' own survey of recipients, 64% said the work would not have happened without the grant, and 57% said they spend more than a quarter of their time on grant applications.
LessonIts founders wrote that much faster decisions "may be feasible by other grantmaking bodies".
Credit
Patrick Collison, Tyler Cowen and Patrick Hsu
2022
The price of applying
Giving Evidence, a UK consultancy, estimated that UK charities spend at least £900 million a year applying to foundations, at least 5.6% of the money they raise that way and at least 17.5% for small charities. It found that few foundations, and few charities, monitor what application processes cost; in 2021 the Esmée Fairbairn Foundation funded about 4% of the applications it received through its website.
LessonFoundations do not see the costs their processes create, so few design them to save charities work.
Credit
Caroline Fiennes, Gemma Bull and Sarah Sandford of Giving Evidence, with funding from the Law Family Commission on Civil Society
Watch for
Cheaper AI shortlisting
Stanford Impact Labs, a research funder at Stanford University, built a multi-agent AI system that ranked letters of interest through over a thousand head-to-head comparisons. In a backtest, a replay on more than 200 past applications, the AI "matched generalist reviewers at spotting the proposals we ultimately funded", and it produced rankings in two days for about $200, against about $3,000 for the usual process. People still review every proposal and make the decisions.
Achieved whenAn outside team repeats the test on another funder's past applications and publishes the agreement rate and any differences by type of applicant.
Claimed by
Michael Eddy, Andrés Felipe Rodríguez and Julia Spencer of Stanford Impact Labs, February 2026
An AI as second assessor
The Clothworkers' Foundation, a UK grantmaker, began testing Plinth AI in July 2026 for eligibility checks and for scoring applications against its impact framework, and compared its judgments with those of staff on past applications. In October 2026 it said the AI would soon replace one of the two assessors on eligible applications to its Open Grants Programme; the AI cannot reject applications, and staff decide when the two disagree. No accuracy figures were published.
Achieved whenThe foundation or an outside reviewer publishes how often the AI and staff agreed, and whether results differed by type of applicant.
Claimed by
The Clothworkers' Foundation, October 2026
Open
01
Name swaps for AI screeners
A public test of whether AI models used to screen grant applications score the same application differently when only the applicant's name, organization or location changes. Write 200 fictional applications to a fictional funder, make matched copies that differ in one detail, score every copy with several models under one published rubric, and report the gaps.
The briefCredit · unclaimed02
Measure the forms
A census of the application forms of 200 UK and US foundations: the number of questions, word limits, required attachments and stages, set against the typical grant size and the share of applicants funded, with an estimate of the hours each form takes.
The briefCredit · unclaimed
Open 01 · the brief
Name swaps for AI screeners
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of whether AI models used to screen grant applications score the same application differently when only the applicant's name, organization or location changes. Write 200 fictional applications to a fictional funder, make matched copies that differ in one detail, score every copy with several models under one published rubric, and report the gaps.
Why it is still open
In a 2025 study of AI reviews of research papers, Sai Suresh Macharla Vasu and colleagues found "a strong affiliation bias favoring authors from highly ranked institutions", smaller gender effects in several models, and more bias in the models' underlying probabilities than in their final scores. Human reviewers of blinded Gates Foundation proposals still scored women lower through differences in wording. Only 1% of foundations Candid surveyed in 2025 used generative AI to screen applicants, but 22% said they would or might, and the Clothworkers' Foundation is moving an AI into an assessor's role. This research found no such test for grant applications.
Built with what exists
AI models such as Claude, Gemini and GPT through their programming interfaces; Gemini Decision 1, which returns probabilities for fixed-answer questions and so can show small shifts that a final score hides (see the Note "A thousand questions, one pass"); Python for paired statistics.
First step
Write 20 fictional applications with four versions of each and run them through two models.
Done looks like
At least four models tested on 200 applications in five matched versions each; score gaps by attribute reported with confidence intervals; rubric, applications and code published so a funder can rerun the test on its own tool.
Open 02 · the brief
Measure the forms
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A census of the application forms of 200 UK and US foundations: the number of questions, word limits, required attachments and stages, set against the typical grant size and the share of applicants funded, with an estimate of the hours each form takes.
Why it is still open
Giving Evidence found that few foundations monitor what their application processes cost charities, and estimated that cost at £900 million a year in the UK. IVAR, a UK research charity, wrote in August 2026 that funders "are seeing far more applications, and far longer ones, than ever before", as AI helps applicants write. This research found no public dataset that compares application demands across funders.
Built with what exists
A web agent such as Browser Use to collect public guidance and forms without submitting anything (see the Note "Deleting 80% to start over"); an AI model to count questions and limits; timed trials by five volunteer grant writers who fill in copies of 20 forms offline for a fictional charity; a spreadsheet.
First step
Collect the public application guidance of 20 foundations and code it by hand.
Done looks like
200 funders' forms coded, each checked by a second person; hours estimated from the timed trials; a published ratio of expected applicant hours to expected grant value for each funder.