A weekly agent pulls new federal opportunities and researches foundation fit, and a drafting assistant assembles first drafts from your own approved program language, which your staff then rewrite and own.
For most small nonprofits, grants are the budget, and the grants pipeline is one overworked person. That person spends hours finding opportunities that fit, reading eligibility rules, and then writing applications that are largely the same organizational history, program description, outcomes, and budget narrative rearranged to each funder's questions and word limits. Two parts of that job are mechanical. Finding federal opportunities is a query against public data: the Grants.gov search2 and fetchOpportunity APIs require no authentication, so a weekly job can pull new postings that match your keywords and eligibility and drop them in a pipeline. Researching a private foundation's fit is reading public records: a foundation's Form 990-PF lists, in Part XIV, the grants and contributions it paid during the year with recipient, purpose, and amount, which shows who it actually funds and at what size. The third part — assembling a first draft from language you have already written and had approved — is what a language model does well, as long as the draft is built from your material and not invented. The constraint is the funder. Some now say how they treat it: NIH's notice NOT-OD-25-132 (released July 17, 2025, effective for the September 25, 2025 receipt date) says it "will not consider applications that are either substantially developed by AI, or contain sections substantially developed by AI, to be original ideas of applicants." It allows that AI tools "may be appropriate to assist in application preparation for limited aspects or in specific circumstances," warns that AI use detected after award may be referred to the Office of Research Integrity alongside enforcement action up to termination, and separately caps applications at six per principal investigator per calendar year (T activity codes and R13 conference grants excepted). Most foundations publish nothing on AI. So this workflow treats every application as the organization's own writing: the model retrieves and assembles approved language and flags gaps; staff write the program-specific substance and own every word submitted.
Collect your best submitted applications, your current program descriptions, outcome data, logic model, board list, budget narrative, and organizational history into one Notion workspace, one page per reusable block, each marked with the date it was last approved by your executive director and the source of every number in it. This library is what the drafting assistant draws from. If a fact is not in the library, the assistant may not use it.
An n8n schedule calls the Grants.gov search2 API weekly with your keywords, eligibility category, and funding categories, and fetchOpportunity for each new result. It writes title, agency, opportunity number, close date, award ceiling and floor, and eligibility to an Airtable pipeline. A person marks each one pursue, watch, or skip in under a minute; the reasons for skips teach you which keywords to drop.
For each foundation prospect, use Perplexity to gather the foundation's stated priorities, geographic focus, and application process from its own site, with citations you can click. Then read Part XIV of its most recent Form 990-PF. Line 2 has a checkbox the foundation ticks if it only funds preselected organizations and does not accept unsolicited requests, and lines 2a to 2d give the application contact, form, deadlines, and restrictions. Line 3a lists grants paid during the year — recipient, purpose, and amount — and 3b grants approved for future payment. Record typical grant size from 3a, whether it funds organizations like yours, and whether it accepts unsolicited requests. A foundation that checks the preselected-organizations box on line 2 goes to watch, not pursue.
Before drafting for any funder, check its guidelines for AI-use, originality, or certification language and record it in the pipeline. Where a funder restricts AI-developed content — NIH is the clearest example — skip step 5 for that application, use the library only as your own reference, and have staff write the draft in step 6. Where the funder is silent, the rules in step 5 still apply.
Give the Claude API the whole library as a fixed prefix, then the funder's questions and word limits and the drafting prompt in the assets. It picks the relevant blocks, maps each question to them, adapts length and order, and inserts a GAP marker wherever the question needs something the library does not contain — this year's numbers, a new partner, the specific project budget. It never invents an outcome, a statistic, a partner, or a quote. Cache the library prefix so drafts and revisions done in one working session read it at the cache-hit price; the default cache lasts 5 minutes, and a 1-hour cache costs more to write.
The program lead fills every GAP and rewrites the need statement and project design in their own words for this funder; the development lead edits for voice; the executive director approves. Check every number against its source in the library. The assembled draft is a scaffold that saves the reshuffling time — the program thinking that wins grants is the part a person writes.
The Airtable pipeline carries close dates with reminders at 30, 14, and 3 days, award decisions, reporting due dates for awarded grants, and funder feedback. Once a quarter, update library blocks with newly approved language from submitted applications and retire stale numbers. A library that is not maintained produces confident drafts with last year's data.
Use these templates as-is or customize for your business.
You are assembling a first draft of a grant application for [Organization] to [Funder]. You may ONLY use facts, numbers, names, and claims that appear in the LIBRARY blocks provided. You are reorganizing and condensing approved language, not creating new content. For each funder question, in order: 1. Quote the question and its limit (words or characters). 2. List the library block IDs you drew from. 3. Write the draft answer within the limit. 4. Insert [GAP: what is needed] wherever the question asks for something the library does not contain — current-year data, project-specific budget lines, new partners, specific timelines, anything about this particular project. Do not fill a GAP with a plausible guess. RULES: - Never invent a statistic, outcome, partner, testimonial, or quotation. - Keep every number exactly as written in the library, with its block ID in brackets after it so staff can check the source. - Do not use language the funder's guidelines prohibit: [paste any] - If two library blocks conflict, use neither and insert [GAP: conflicting figures in blocks X and Y]. End with: a list of all GAPs, and a word/character count per answer against its limit.
NODE 1 — Schedule: weekly, Monday 7:00
NODE 2 — HTTP Request
Method: POST
URL: https://api.grants.gov/v1/api/search2
Body (JSON):
{
"keyword": "[your program keywords, e.g. youth workforce]",
"oppStatuses": "forecasted|posted",
"eligibilities": "[your eligibility code, e.g. 12 for 501(c)(3) nonprofits other than institutions of higher education]",
"rows": 50
}
(No API key required for search2 and fetchOpportunity per the Grants.gov API guide. Eligibility code 12 = Nonprofits having a 501(c)(3) status with the IRS, other than institutions of higher education; the full code list comes back in every search2 response. search2 returns IDs, titles, agencies and dates; award ceiling and floor come from fetchOpportunity.)
NODE 3 — Filter: keep only opportunity IDs not already in Airtable
NODE 4 — HTTP Request per new ID
POST https://api.grants.gov/v1/api/fetchOpportunity
Body: { "opportunityId": <id> }
NODE 5 — Airtable: create record
title, agency, opportunity_number, close_date, award_ceiling, award_floor, eligibility, link, status = NEW
NODE 6 — Slack or email digest to the development lead: count of new opportunities with linksfoundation | name ein | from the 990-PF source_990pf_year | tax year read stated_priorities | from the foundation's own site (link) geography | where it funds accepts_unsolicited | yes | no (Part XIV line 2 preselected box checked) | LOI only | unclear application_process | from Part XIV lines 2a-2d: contact, form, deadlines, restrictions grants_reviewed | count of Part XIV line 3a (paid) entries read typical_grant_range | from Part XIV line 3a amounts (low / median / high you observed) funds_orgs_like_us | yes | partly | no — with two example grantees and purposes ai_or_originality_rules | quote or 'none published' with date checked deadline / cycle | dates ask_amount_planned | within typical range relationship | none | contact made | prior grantee decision | pursue | watch | skip, with one-line reason RULE: a foundation whose Part XIV shows no grants to organizations like yours goes to watch, whatever its website says.
Get a new AI workflow every week. Prompts, tool stacks, and ROI math included.
Retrieval-augmented generation: the agent answers strictly from a curated corpus of your documents and history. Cheaper, more controllable, and fewer hallucinations than open-ended generation.
Learn the agentic glossary →Where this workflow tends to break in production — and what to put in place before you ship it.
Draft invents an outcome, statistic, or partner
Mitigation: Prompt restricts facts to library blocks with block IDs after every number; GAP markers for anything missing; staff check every number against its source.
AI-assisted draft submitted to a funder that restricts AI-developed content
Mitigation: Funder AI/originality rules recorded in the pipeline before drafting; restricted funders skip the drafting step.
Library holds stale numbers that flow into new applications
Mitigation: Each block carries its approval date and source; quarterly library update; conflicting blocks produce a GAP, not a choice.
Pipeline fills with poor-fit opportunities
Mitigation: Eligibility filter in the Grants.gov query; one-minute pursue/watch/skip triage; 990-PF grantee review before any foundation is pursued.
Generic drafts lower award rates
Mitigation: Program lead rewrites need statement and project design; track award rate per quarter, not drafts produced.
Do not use the drafting step for any funder that restricts AI-developed content — NIH does not treat substantially AI-developed applications or sections as the applicant's original ideas, and AI use detected after award can lead to a research-integrity referral and enforcement action — and check every funder's guidelines before drafting rather than assuming silence means permission. Skip the drafting assistant entirely if you do not yet have a body of approved, sourced language; it will either produce thin drafts or, worse, tempt someone to let it fill the gaps. Never submit an assembled draft without the program lead rewriting the project-specific sections: funders read hundreds of applications, and a well-formatted response that says nothing particular about this project loses to a plain one that does. And remember that a federal application also needs current registrations and certifications that no workflow here handles.
A phased approach to get this workflow running and delivering ROI.
Days 1–30
Foundation
Days 31–60
Optimization
Days 61–90
Scale
Most enterprises do not have an AI adoption problem. They have an AI inventory problem — and the first honest count is usually three to five times what anyone predicted.
None of the three big legal AI platforms publishes a per-seat price. Here is what each one does publish (research content, security commitments, plan terms) and how a 25-lawyer firm should structure the quote.
The managed tool layers for AI agents price the same action four different ways. Here is what Composio, Arcade.dev and Pipedream Connect publish, what running your own MCP servers costs you instead, and a worked monthly bill.
One practical AI workflow per week. No fluff.
Get the full guide with step-by-step setup, workflow templates, and copy-paste assets.