Industries / Government, education and community / Nonprofits / Foundations
Measuring results
Foundations try to find out whether their grants worked, by asking grantees, commissioning evaluations and, in some cases, paying only for results.
4
Milestones achieved, from 2003 to 2018.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2003
Grantees grade their funders
The Center for Effective Philanthropy built the Grantee Perception Report, a survey of a foundation's grantees whose results are compared with those of peer funders. The Hewlett Foundation has commissioned one regularly since 2003, and CEP says more than 300 foundations have used it.
Credit
Center for Effective Philanthropy; William and Flora Hewlett Foundation
2016
Evaluations kept in-house
In a 2016 benchmarking survey of 127 senior evaluation or program staff, run by the Center for Evaluation Innovation and CEP, only 57% said their foundations share evaluation findings outside their own institutions, and 71% said their foundations invest too little in sharing evaluation data.
LessonFoundations' own evaluation staff said they share too little of what they learn.
Credit
Center for Evaluation Innovation; Center for Effective Philanthropy
2018
Paid only for results
In what Brookings calls the world's first development impact bond for education, the UBS Optimus Foundation put up $270,000 for the charity Educate Girls, and the Children's Investment Fund Foundation (CIFF) agreed to repay it, with a 15% internal rate of return (a yearly return), if results were met. The evaluator IDinsight, using a randomized trial to measure learning, found that by year three 768 girls, 92% of those identified, had enrolled, and learning gains reached 160% of the target.
LessonBrookings noted that the results show the program worked, not that paying by results beats other ways of paying for it.
Credit
UBS Optimus Foundation; Children's Investment Fund Foundation; Educate Girls; Instiglio; IDinsight
2018
Better measures, same results
From 2009 through 2016, three school districts and four charter school groups spent $597 million on the Gates Foundation's Intensive Partnerships for Effective Teaching, $237 million of it from Gates grants. RAND and the American Institutes for Research found that the sites improved how teachers were evaluated but that the initiative "failed to achieve its goals for improved student achievement and graduation".
LessonThe evaluators pointed to incomplete implementation without proven models to copy, conflicting goals for teacher evaluation, and changing state and local conditions.
Credit
Bill & Melinda Gates Foundation; RAND Corporation and American Institutes for Research (evaluation)
Watch for
An AI reader for reports
The GitLab Foundation says a staff member built an internal AI tool to help review grantee reports and research literature, and that it "has reduced manual analysis time by up to five hours per grantee application". No check of the tool's accuracy was published.
Achieved whenThe foundation or an outside party publishes how often the tool's summaries match a person's reading of the same reports.
Claimed by
GitLab Foundation, June 2026
Open
01
What the evaluations found
An open index of evaluations that foundations have commissioned and published, coded by program, funder, country, method (randomized trial, comparison group, interviews only) and main result (improved, no clear effect, worse), starting from Candid's Issue Lab collection and adding reports from foundation websites.
The briefCredit · unclaimed02
Reports sized to grants
A dataset of what 150 foundations ask grantees to report, from their published reporting guidance and grant terms: how often, how many questions, what data and attachments, set against each foundation's typical grant size.
The briefCredit · unclaimed
Open 01 · the brief
What the evaluations found
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open index of evaluations that foundations have commissioned and published, coded by program, funder, country, method (randomized trial, comparison group, interviews only) and main result (improved, no clear effect, worse), starting from Candid's Issue Lab collection and adding reports from foundation websites.
Why it is still open
Issue Lab holds more than 36,000 research reports, case studies and toolkits and is moving into the permanent collection of the Rockefeller Archive Center, but Candid's description of it mentions no coding of what the reports found. In 2016 only 57% of foundation evaluation staff surveyed said their foundations share findings outside. This research found no dataset that codes the results of published foundation evaluations.
Built with what exists
An AI model that reads long PDF reports, such as Claude, Gemini or Muse Ember (see the Note "The jobs agents repeat"), to draft the codes; a coding guide modeled on those used in systematic reviews; two human coders; Python.
First step
Pull 50 evaluations on one topic, such as education, from Issue Lab and code them by hand.
Done looks like
1,000 evaluations coded; two coders agree on the main-result code for at least 85% of a 200-report sample; the share of null results reported by method and field; data published under an open license.
Open 02 · the brief
Reports sized to grants
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A dataset of what 150 foundations ask grantees to report, from their published reporting guidance and grant terms: how often, how many questions, what data and attachments, set against each foundation's typical grant size.
Why it is still open
In 2020, 64% of foundations CEP surveyed said they had reduced what they ask of grantees, a self-report made during the pandemic. Giving Evidence found that few foundations track the costs their processes create for charities. This research found no public comparison of reporting demands with grant size.
Built with what exists
A web agent such as Browser Use to collect public reporting guidance and templates (see the Note "Deleting 80% to start over"); an AI model to code the demands; 360Giving's GrantNav and US tax return data for grant sizes; a spreadsheet.
First step
Gather the reporting guidance of 20 funders and code it by hand.
Done looks like
150 foundations coded by two people who agree on at least 90% of codes; a published chart of reporting demands against median grant size; the count repeated after a year to see whether demands fall.