How Should an AI Agent Score a Foundation's Fit With My Program?
Build an AI agent in Gumloop that reads a funder's own pages and hands you a one-page brief, scored against your rubric, with questions to ask.
From a funder's own priorities page and recent grants, an AI agent you build in Gumloop scores the fit against a rubric you wrote. It hands back a one-page brief with a score for each criterion, a quoted line of evidence for each score, and questions for the program officer. You decide whether to apply.
You've got a foundation's website open in a pile of tabs. The priorities page uses words your program fits.
The signal is in what the foundation has funded lately, on the grants page and in the latest annual report. An AI agent can read both in one pass against the rubric you write, and it quotes what the funder published.
What does the one-page brief hold?
You give the AI agent a funder's name, two to four links from the funder's own site (priorities, recent grants, the latest annual report), and a short summary of your program. It hands back a score for each criterion in your rubric and a total, a short quote with its link as evidence for every score, the recent grants it could see, and four questions for the program officer that the pages didn't answer.
If the funder doesn't list its grantees, the brief says "Not found" and the question list picks it up.
The score is a read. In a podcast interview I said "You need a human in the loop with your AI tools," and here the human is you.
How do you build the funder fit scorer in Gumloop?
- 01
Write the rubric on paper first.
Pick four to six criteria, scored 0 to 3, with a sentence on what a 3 looks like in your words. Geography, population served, type of support, grant size and timing are the usual candidates. What Should You Do on Paper Before You Build an AI Agent? walks through it.
- 02
Create the AI agent and paste in the instructions.
Start a new AI agent in Gumloop and answer guided setup in one sentence: score a foundation's fit with my program from its own pages. Then open the tab named Agent, replace the instructions box with the block below, and press Save (Ctrl/Cmd + S).
- 03
Keep Web Fetch and Web Search on.
Fetch reads the links you hand it, and search finds a page on the funder's own site you didn't have a link for. Which Abilities Should I Turn On for My AI Agent? covers the rest.
- 04
Connect Google Drive only if your program summary lives there, and change its approval.
Find the connector in the Connectors list, click the approval icon next to it and choose Ask for writes/deletes, because the default is Always allow. Leave Gmail and every other connector off. Or paste the summary into each task instead.
- 05
Copy the instructions out, then turn off Edit Own Instructions.
It's on by default and has no version history, so paste your instructions into a document you keep. Reflections can also propose instruction changes on a schedule, so if it's on (Performance in your AI agent's sidebar), set Apply Suggestions to Review Queue and a suggested change to your rubric waits for you.
- 06
Run it on a funder you already know.
You're the trigger: open a new task and type the funder's name, the links and your program summary. Don't set up scheduled triggers for now. Read the brief beside your own opinion of that funder.
I am an ambassador for Gumloop. Code NONPROFITAGENTS takes 20% off your first month.
Checked on 2026-11-09 against Human in the loop, which lists Always allow as the default approval mode and Ask for writes/deletes as an option.
Checked on 2026-11-09 against Gumloop's AI agent builder page, which describes guided setup for a new AI agent, lists Web Search, Web Fetch and Edit Own Instructions as on by default, the last with no version history, and says nothing saves until you press Save.
Checked on 2026-11-09 against Gumloop's triggers page, which describes scheduled triggers that run an AI agent with a prompt you configure.
Checked on 2026-11-11 against Gumloop's Reflections page, which says new AI agents on the Pro plan and above start with Reflections on and set to Auto-Apply Suggestions, and that with Review Queue every suggestion waits for you.
What do you paste into the instructions for the funder fit scorer?
You are my grants research assistant AI agent. I'm the [YOUR ROLE] at [ORGANIZATION], and I decide whether we apply. My goal is to apply only where a funder's published priorities match [PROGRAM NAME]: [TWO SENTENCES ON WHO YOU SERVE AND WHAT YOU DO]. When I give you a funder and links, read those pages and others on that funder's own website. Don't use tax filings, government databases or third-party grant databases. If a page won't open, tell me. Score the funder from 0 to 3 on each criterion: [CRITERION ONE]. A 3 means [WHAT A 3 LOOKS LIKE]. [CRITERION TWO]. A 3 means [WHAT A 3 LOOKS LIKE]. [CRITERION THREE]. A 3 means [WHAT A 3 LOOKS LIKE]. [CRITERION FOUR]. A 3 means [WHAT A 3 LOOKS LIKE]. Give me one page: each score and the total, a short quote and link as evidence for each, the recent grants you could see, and four questions for the program officer the pages didn't answer. If you can't find evidence for a criterion, write "Not found on the pages provided" and leave the score blank. Never guess a grant size or a deadline. Never contact anyone. Be proactive about gaps. The decision to apply is mine.
How do you tell whether the scores can be trusted?
Open every link in the first few briefs. Each quote should sit on the page word for word. If one doesn't, the AI agent paraphrased, so tell it to quote or leave the evidence blank.
Score a funder you know well and compare the total with your own read. If the AI agent runs high, you'll find a criterion where a broad word earned a 3. Tighten what a 3 means.
Count the "Not found" lines. Each one should've turned into a question for the program officer.
When a funder says yes, that's a different job. Next, read How Do I Build a Donor Stewardship Tracker With an AI Agent?.
Frequently asked questions
Does the AI agent decide whether we apply?
No. Against your rubric, the AI agent scores the funder and lists the questions the pages left open.
Why do the instructions keep it to the funder's own pages?
The brief quotes a line from a page behind each score, and the program officer can read the same pages. The instructions rule out tax filings, government databases and third-party grant databases.
Can it check my funders on a schedule?
Yes, later. Scheduled triggers run an AI agent with a prompt you set. Run it by hand for three weeks first.