Can you use AI for grant writing?

Yes. AI excels at funder research, drafting needs statements, building logic models, writing budget narratives, and running compliance checks. The key is matching the right model to each stage — Gemini 2.5 Pro for research, Claude Sonnet for narrative prose, GPT-5 for logic models. AI should amplify the PI's original ideas, not replace them.

Does NIH allow AI in grant proposals?

Partially. NIH Notice NOT-OD-25-132 (effective September 25, 2025) states that proposals where the scientific ideas are substantially developed by AI will be rejected. AI is acceptable for editing, formatting, grammar review, and literature summarization — as long as the PI's original scientific concepts drive the content. Disclose any AI use per NIH guidance.

Does NSF require AI disclosure in grant proposals?

NSF's notice to the research community on AI advises transparency about AI use in proposal preparation. While NSF has not yet issued a blanket prohibition, reviewers expect disclosure of significant AI-assisted content generation. Use the NSF template provided on this page as a starting point, and always check the current RFP guidance.

Is using AI for grant writing ethical?

Yes, when used appropriately. AI is a drafting and research tool, not a substitute for the applicant's expertise, mission knowledge, or original ideas. The ethical line is transparency — non-disclosure of substantial AI use while claiming sole authorship creates research-misconduct risk. This guide treats disclosure as mandatory, not optional.

Will AI-generated grants be detected by funders?

AI detection tools (GPTZero, Originality.ai) are imperfect — they produce false positives on well-edited human text and miss well-edited AI text. Detection capability is not the right frame. The risk is non-disclosure, which is a research-misconduct issue if discovered. Disclose appropriately and ensure the PI's authentic expertise drives the proposal.

How much time does AI save on a federal grant application?

Federal grant applications average 80–200 hours according to Instrumentl's 2025 grant statistics report. One case study (Pets for Patriots) reported an 80–90% reduction in proposal completion time with AI assistance — though this represents a single reported case, not a sector-wide average. Foundation grants, which average 15–20 hours, also show meaningful time savings.

What is the best AI model for writing a needs statement?

Claude Sonnet 4.6 is the leading choice for needs statement narrative. It excels at empathetic, story-driven prose that speaks to a funder's mission alignment. The key is providing Claude with your organization's program data, community demographics, and theory of change before asking it to draft — context drives quality.

Can I use ChatGPT for SBIR grant proposals?

You can use GPT-5 (via ZeroTwo) for SBIR, but the same NIH and NSF disclosure rules apply. GPT-5 is particularly strong for logic models and structured reasoning sections in SBIR Phase I/II proposals. Pair it with a compliance check pass using Claude to verify SBIR-specific requirements (commercialization plan, budget caps, page limits).

How does ZeroTwo compare to Grantable or Grantboost?

Grantable and Grantboost offer funder databases and proposal templates specialized for grant writing. ZeroTwo is a multi-model orchestration platform — Claude, GPT-5, and Gemini in one chat — giving you the best model for each stage of the grant workflow. The recommended approach: use Instrumentl or Candid for funder discovery, then ZeroTwo for the drafting and compliance workflow.

What does a compliant AI disclosure statement look like?

A compliant disclosure identifies the specific tasks where AI was used (e.g., literature summarization, grammar editing) and explicitly states AI was not used to generate the PI's scientific ideas. This page provides four funder-specific templates — NIH, NSF, DoL, and private foundations — as starting points. Always verify against the current RFP guidance before submission.

Grant Writing Workflow

AI for Grant Writing: A Model-by-Stage Workflow for Nonprofits

TL;DR

AI for grant writing works best when you match the model to the stage — Gemini 2.5 Pro for funder research, Claude Sonnet 4.6 for narrative prose, GPT-5 for logic models, and a compliance pass before submission. This guide covers the full five-stage workflow for NIH, NSF, DoL, and foundation grants, with copy-paste prompts and funder-specific AI disclosure templates. The #1 risk: NIH's September 2025 policy rejects substantially-AI-developed proposals.

Start your grant workflow free — Claude + GPT-5 + Gemini in one chat

No credit card required. Free tier available.

5-Stage Grant Workflow

Funder Research

Gemini 2.5 Pro

2M-token context ingests RFP corpora + past awards

Needs Statement

Claude Sonnet 4.6

Story-driven, empathetic narrative prose

Logic Model

GPT-5

Structured outcomes, theory of change tables

Budget Narrative

Claude + data tools

Reconciles line items to narrative, catches math errors

Compliance Check

Claude

Word counts, page limits, required sections

80–200 hrs

Avg federal grant writing time

Instrumentl 2025

~25%

Federal grant approval rate

Instrumentl 2025

15%

Nonprofits with responsible-AI policy

Candid 2025, n=850

Sept 25, 2025

NIH enforcement date for AI proposals

NIH NOT-OD-25-132

What is AI for grant writing, and what does it actually do well?

AI for grant writing means using large language models to research funders, draft proposal sections, build logic models, and check compliance — not to generate the applicant's original ideas.

Federal and foundation grants require structured storytelling: a compelling needs statement, a defensible logic model, a detailed budget narrative, and precise formatting. AI accelerates each of these tasks without replacing the expertise, mission knowledge, and lived experience that funders actually fund.

According to Instrumentl's 2025 grant statistics report, federal grant applications take an average of 80 to 200 hours to complete. Foundation grants average 15–20 hours. AI-assisted workflows can compress the research and drafting phases significantly — one reported case (Pets for Patriots) showed an 80–90% reduction in proposal completion time, though this represents a single use case, not a sector average.

The Candid 2025 AI Equity Project (n=850 nonprofits) found that 96% of nonprofits report a basic understanding of AI capabilities, yet only 15% have a responsible-AI policy. Grant writing is the highest-leverage use case — and the one requiring the most care.

For specialized AI grant writing workflows, see also our guide to AI workflows for consultants and grant writers.

The five-stage AI grant writing workflow

One model per stage — each chosen for the task, not the trend.

1Funder Research + Lit Review

How do you use AI for funder research?

Use Gemini 2.5 Pro. Its 2M-token context window can ingest an entire corpus of past RFPs, prior award abstracts, 990s from peer organizations, and the funder's strategic plan in a single session. No other model currently matches this for large-document synthesis.

Provide Gemini with: the funder's published priorities, 3–5 past awards in your domain, and your organization's program summary. Ask it to identify alignment gaps, priority language matches, and recommended emphasis for your specific application.

Contextual CTA

Switch to Gemini 2.5 Pro in ZeroTwo to load your full RFP corpus into a single thread.

2Needs Statement Narrative

How do you write a needs statement with AI?

Use Claude Sonnet 4.6. Needs statements require empathetic, story-driven prose that connects community data to human reality. Claude's writing is characteristically nuanced — it can hold the tension between statistical evidence and narrative voice better than more technically-oriented models.

Required inputs: your community needs data, geographic scope, target population demographics, and the theory of change. Give Claude the funder's stated priorities (from Stage 1) so it can mirror their language.

3Logic Model / Theory of Change

How do you build a logic model with AI?

Use GPT-5. Logic models require structured reasoning: inputs → activities → outputs → outcomes → impact. GPT-5's reasoning capabilities handle tabular output chains and multi-column logic model formats reliably, with fewer hallucinations in structured data contexts.

Provide your program activities, intended beneficiaries, short- and long-term outcomes, and any existing program evaluation data. Ask GPT-5 to output a structured Markdown table you can paste directly into your proposal document.

4Budget Narrative + Justification

How do you draft a budget narrative with AI?

Use Claude with data tool use. Budget narratives must reconcile line items to prose justification — a task riddled with human error. Claude with file upload can ingest your budget CSV and cross-check every line item against the narrative description, flagging discrepancies before reviewers do.

Contextual CTA

Attach your budget CSV in ZeroTwo and have Claude cross-check line items against the narrative in one thread.

5Compliance Check + Formatting

How do you run a compliance check?

Use Claude. Compliance checking requires precise attention to word counts, page limits, required section headings, and funder-specific formatting rules. Paste the full RFP requirements and your draft into Claude and ask it to produce a compliance matrix — every requirement checked against your submission.

For NIH applications: R01 page limits, NIH-specific fonts (Arial 11pt minimum), and biosketches. For NSF: strict 15-page project description limits, data management plan requirements. Claude handles these mechanical checks reliably so you focus on substance.

!Critical: Funder Disclosure

NIH rejects substantially-AI-developed proposals

NIH Notice NOT-OD-25-132 (effective September 25, 2025) states that applications where scientific ideas are substantially AI-developed will be rejected. See the disclosure templates below.

Do you have to disclose AI use to funders? (NIH, NSF, DoL, foundations)

Yes — and the rules differ significantly by funder. NIH's September 2025 policy is the most consequential: proposals where AI substantially developed the scientific ideas will be rejected.

NIH NOT-OD-25-132 — Effective September 25, 2025

The NIH Office of Extramural Research states: applications that use AI to substantially develop scientific ideas, hypotheses, or specific aims will not be reviewed. AI use for grammar editing, literature summarization, and administrative tasks remains acceptable with appropriate disclosure.

ZeroTwo's intended use: pre-drafting, funder research, logic modeling, compliance checking, and editing — not generating the PI's original scientific ideas. The expertise, hypotheses, and study design must originate with the investigator.

Read NIH NOT-OD-25-132 in full →

"Applications that do not reflect the expertise, knowledge, and capabilities of the Principal Investigator and co-investigators are not acceptable."
— NIH Office of Extramural Research, Notice NOT-OD-25-132 (September 25, 2025)

Funder AI policy comparison

Funder	Policy Status	Key Requirement	Source
NIH (R01, R21, SBIR/STTR)	Rejection risk	No AI-developed scientific ideas. Disclose all AI use.	NOT-OD-25-132
NSF (Research, SBIR)	Disclosure required	Disclose AI-assisted content generation in project description.	NSF AI Notice
DoL (ETA, WIOA)	RFP-specific	Check each solicitation; WIOA grant RFPs vary.	Check current NOFO
Private Foundations	Varies	No sector standard; proactive disclosure recommended.	Candid / Grantmaker guidance

Funder-specific AI disclosure templates

Legal disclaimer: These templates are starting points only. Always verify against the current RFP, program officer guidance, and institutional policy before submission. Disclosure requirements change. These templates do not constitute legal advice.

NIH (NOT-OD-25-132 compliant)

AI assistance was used in the preparation of this application for the following limited purposes: [grammar editing / literature search summarization / administrative formatting]. The scientific hypotheses, specific aims, study design, and all substantive scientific content were developed solely by the Principal Investigator and co-investigators named herein. No AI tool was used to generate, substantially revise, or replace the scientific ideas in this application.

NSF (Merit Review guidance)

Artificial intelligence tools were used in the preparation of this proposal for the following purposes: [specify: e.g., literature review assistance, grammar review, figure formatting]. The intellectual content of this proposal, including the research objectives, methodology, and theoretical framework, was developed by the investigators. All claims, data, and citations have been verified by the PI.

DoL / ETA (WIOA programs)

This application was prepared with AI assistance for [specify tasks: e.g., editing and proofreading, section formatting]. Program design, participant eligibility criteria, performance targets, and budget projections were developed by [Organization Name] staff based on our direct program experience and community assessment data. We affirm compliance with all applicable requirements in [NOFO number].

Private Foundation (general)

We used AI writing assistance to support the drafting and editing of this proposal. Our programmatic expertise, organizational history, and the needs assessment underlying this request are based on [X] years of direct service experience. The proposal reflects the authentic work and mission of [Organization Name] and was reviewed and approved by [Title] prior to submission.

The complete prompt pack (one per stage)

Copy-paste prompts for each stage. Specify your inputs before sending. Use ZeroTwo's model selector to switch to the recommended model for each stage.

Stage 1 — Funder Research (Gemini 2.5 Pro)

I'm preparing a grant application to [FUNDER NAME] for [PROGRAM AREA]. I've attached:
- The full RFP / NOFO document
- 3–5 prior award abstracts from this funder
- Our organization's program summary (1–2 pages)

Please analyze the funder's stated priorities and language patterns, identify the top 3 alignment opportunities between our program and their priorities, flag any language in our summary that may conflict with their framing, and recommend the emphasis and narrative angle for our proposal. Output as a structured brief with section headers.

Stage 2 — Needs Statement (Claude Sonnet 4.6)

Write a grant needs statement for [FUNDER NAME]'s [PROGRAM NAME] grant.

Inputs:
- Target population: [demographics, geography, size]
- Community need data: [key statistics with sources]
- Our organization's theory of change: [summary]
- Funder priority language to mirror: [paste from Stage 1 brief]

Write 400–600 words in a compelling narrative style. Lead with a human story, support with data, connect to the funder's mission. Do not use bullet points. End with a direct statement of what we are asking the funder to fund.

After drafting, identify the 2 weakest claims and suggest how to strengthen them.

Stage 3 — Logic Model (GPT-5)

Create a standard logic model table for our grant proposal.

Program name: [NAME]
Inputs available: [staff, funding, partners, facilities]
Key activities: [list 4–6]
Target population and volume: [e.g., 200 youth ages 14–18 annually]
Short-term outcomes (6–12 months): [list 3–4]
Long-term outcomes (2–5 years): [list 2–3]
Ultimate impact goal: [one sentence]

Output as a Markdown table with columns: Inputs | Activities | Outputs | Short-Term Outcomes | Long-Term Outcomes | Impact.

Then provide a 1-paragraph "Theory of Change" narrative summarizing the model.

Stage 4 — Budget Narrative (Claude + data tools)

I've attached our grant budget spreadsheet (CSV). Write a budget narrative justifying each line item.

Requirements:
- Funder: [NAME], program: [NAME]
- Total request: $[AMOUNT], period: [DATES]
- Personnel: explain FTE allocations, roles, and their necessity
- Fringe: state rate and basis
- Indirect costs: [rate and negotiated agreement if applicable]
- Any cost-share or matching: [describe]

After writing the narrative, produce a compliance checklist: confirm each budget category in the RFP is addressed. Flag any line items where the narrative and budget figures do not match.

Stage 5 — Compliance Check (Claude)

I've attached two documents: (1) the complete RFP/NOFO for [FUNDER NAME]'s [PROGRAM], and (2) our draft proposal.

Produce a compliance matrix as a table with columns: Requirement | Location in RFP | Addressed in Draft (Yes/No/Partial) | Notes.

Include: page limits, font/margin requirements, required sections, attachment requirements, certification requirements, and any funder-specific prohibitions.

Flag any requirement that is not addressed or only partially addressed, and suggest specific language or sections to add.

Draft your first NIH/NSF-compliant section in ZeroTwo

Switch between Gemini, Claude, and GPT-5 mid-thread — no separate subscriptions. Free tier, no credit card required.

Start your grant workflow free

Frequently asked questions

Key takeaways

Match the model to the stage: Gemini 2.5 Pro for funder research (2M-token context), Claude Sonnet 4.6 for narrative prose, GPT-5 for logic models, Claude for compliance checks.

NIH NOT-OD-25-132 (effective Sept 25, 2025) rejects applications where AI substantially developed the scientific ideas. Disclosure of AI use in acceptable tasks is required.

AI saves meaningful time — federal grants average 80–200 hours — but the PI's expertise, hypotheses, and original ideas must drive the proposal.

Funder-specific disclosure templates are not optional risk management; they are the difference between compliant use and research-misconduct exposure.

For funder discovery, pair ZeroTwo with Instrumentl or Candid — ZeroTwo is the drafting and compliance layer, not a funder database.

AI workflow research and grant compliance analysis

Published: 2026-04-22

Updated: 2026-04-22

Free tier, no credit card — run the prompt pack above in ZeroTwo

Gemini 2.5 Pro, Claude Sonnet 4.6, and GPT-5 in one thread. Switch models at any stage without losing context.