Can you use AI to write a grant application? Yes — but almost certainly not for the part you think.
The prose is the worst place to point it. A program officer reads applications in stacks, and generic, fluent, faintly hollow writing is exactly what they have been trained by volume to skim past. Where AI genuinely earns its place in a nonprofit's grant cycle is upstream and downstream of the writing: finding the right funders, reading their guidelines properly, structuring the logic model, sanity-checking the budget, and turning one funded proposal into the next four. Do that, write the narrative yourself, and you will be ahead of the organisations doing the opposite.
Why AI-written narratives fail review
Reviewers are not running detection software, and mostly they do not need to. Three things give the game away.
It says nothing specific. AI prose defaults to the general case, because it produces the most probable sentence. "We serve a diverse population facing significant barriers" is a sentence with no information in it. What a reviewer wants: the exact household count, the two clinics that make most of the referrals, and the caseworker's account of why the referral pathway keeps breaking. A model cannot produce that. It does not know it. Only you do.
It has no voice. Grant panels often include people who have run programmes like yours. Prose that reads like a polished stranger writing about your community, rather than someone who has been inside it, registers as a distance problem before it registers as an AI problem.
It answers a different question. Models pattern-match to what a grant application usually looks like, not to what this funder actually asked. The single most common cause of rejection is not writing quality. It is failing to answer the prompt as written, and a fluent draft makes that failure harder to notice, not easier.
The hard part of a grant application was never the sentences. It was knowing what you actually do and being able to prove it works. AI does not help with either. It helps with everything around them.
Where AI genuinely helps
Funder research and prospecting
This is the highest-value use, and it is not glamorous. Perplexity is well-suited to funder discovery because it cites its sources — you can click through and verify the foundation exists, is still funding, and funds your area. That verification step is not optional. Models hallucinate plausible-sounding foundations, and a fictional funder is a wasted week.
Dedicated grant-prospecting platforms such as Instrumentl and Candid's databases remain the serious tools here; a chat model is a research accelerant, not a replacement for a funder database. See our guide to AI tools that cite real sources for why this distinction matters more than any other.
- Who should use it: every organisation without a dedicated development officer.
- Who should not: nobody — but nobody should skip the verification step either.
Reading guidelines and RFPs properly
A federal RFP can run to 80 pages, and the requirement that disqualifies you is usually on page 61. Upload it to Claude or ChatGPT, or a document-specific tool such as ChatPDF or Humata, and interrogate it: what are the eligibility criteria, what is the page limit, what attachments are required, what must the budget narrative include, what are the reporting obligations if we win.
Then ask the question most applicants never ask: what would disqualify us? Models are good at this — it is a comprehension task with the answer in the document — and it is worth an hour of a program director's time every single cycle.
Structuring the logic model
Theory of change, logic model, outcomes framework: whatever your funder calls it, it is a structure, and structuring is what these tools do best. Describe your programme in plain language, ask for inputs, activities, outputs, outcomes, and the assumptions connecting them, then argue with the result. The argument is where the value is — the model will propose a causal chain, you will see immediately which link is fantasy, and that is the link the reviewer was going to find anyway.
The first draft you are going to rewrite
If a blank page is the blocker, get words down. Feed the model your notes, your last funded proposal, your programme data. Then rewrite every paragraph. The output is scaffolding, not a draft — the version you submit should retain the structure and almost none of the sentences. Our guide to making AI writing not sound like AI is directly relevant, because a grant panel is the least forgiving audience for that failure mode.
Budget and data work
Julius and similar analysis tools will take a spreadsheet of programme data and produce the summary tables and charts a proposal needs, without a staff member spending a Saturday in Excel. The numbers must be yours and you must check them — but the formatting, the pivoting, and the chart-making are genuinely mechanical work.
The funder presentation
If you get to a site visit or a pitch, Gamma turns an outline into a presentable deck in minutes. For a small organisation with no designer, this is a real saving. The presentation tools category has the alternatives.
The reporting burden
Nobody talks about this and it may be the biggest win. Grant reporting — quarterly narratives, outcome summaries, the same information reformatted for six different funders — is repetitive, deadline-driven, and low-stakes in terms of prose quality. It is nearly ideal AI work, and it consumes a startling amount of a small nonprofit's capacity.
Where AI helps, and where it must not touch
| Grant cycle stage | What AI is good for | What to never hand over |
|---|---|---|
| Prospecting | Shortlisting funders, summarising priorities | Trusting a funder exists without verifying |
| Reading the RFP | Extracting requirements, finding disqualifiers | Assuming it caught everything |
| Logic model | Structuring inputs, outputs, outcomes | The evidence that your programme works |
| Narrative | Outline, gap-check, tightening your own prose | The story, the specifics, the voice |
| Budget | Formatting, tables, charts, arithmetic checks | The figures themselves |
| Reporting | Reformatting outcomes for multiple funders | Outcome data you have not measured |
Swipe the table sideways to see every column →
Disclosure: check before you submit
A growing number of funders now ask whether AI was used in preparing an application, and some restrict it. Policies vary widely — from open permission, to disclosure requirements, to outright prohibition on AI-generated narrative text.
Three rules:
- Read the funder's AI policy before you draft, not after. If it is not in the guidelines, ask the program officer. Asking is not a red flag; it reads as diligence.
- If disclosure is requested, disclose accurately. "We used AI tools for research summarisation and drafting support; all narrative content was written and verified by staff" is an honest, unremarkable sentence, and it is far better than being asked afterwards.
- Never let AI generate an outcome number, a beneficiary count, or a citation. This is the failure mode that ends relationships rather than just applications. Everything factual must come from your own records.
The free and discounted tier nonprofits routinely miss
Before buying anything, check what you can already get.
- Google for Nonprofits — eligible organisations get Google Workspace at no cost, along with Ad Grants for search advertising. Gemini features are bundled into parts of that surface.
- Microsoft for Nonprofits — donated and discounted licences, including Microsoft 365, with Copilot available on eligible plans.
- Canva for Nonprofits — the pro tier free for eligible organisations, which covers most of what a small charity needs from a design tool.
- TechSoup — the clearing house for donated and heavily discounted software; if you are eligible for anything, this is where it will be listed.
- Free tiers of the major models — ChatGPT, Claude, and Gemini all have usable free tiers. For an organisation doing a few applications a year, that may be the entire AI budget.
Eligibility criteria differ by country and organisation type, and terms change; verify with each programme directly. But a great many nonprofits pay for tools they qualify to receive free, which is a quiet and entirely avoidable waste of restricted funds.
A workflow for the next application
- Read the guidelines with a model. Twice. Once for requirements, once for disqualifiers.
- Verify every funder detail by hand. Click the link. Confirm the deadline on the funder's own site.
- Build the logic model in conversation. Argue with it until the causal chain holds.
- Write the narrative yourself. Use AI to find the gaps, never to fill them.
- Insert three specifics per section. A number from your records, a name, a moment. This is what the model cannot fabricate and the reviewer cannot forget.
- Check the AI-use policy and disclose if asked.
- Reuse the structure, not the sentences. Build a library of your funded proposals — that is your best training material, and it is yours.
FAQ
Can AI write a grant proposal?
It can produce a competent, generic draft, which is precisely the problem. Reviewers read many applications and generic fluency is what they skim. Use AI for funder research, requirement extraction, logic-model structuring, and reporting, then write the narrative yourself with specifics only your organisation has. The tools in the AI writing tools category are drafting aids, not applicants.
Do funders allow AI in grant applications?
It depends entirely on the funder, and the landscape is moving quickly. Some are relaxed, some require disclosure, and some prohibit AI-generated narrative outright. Check each funder's guidelines before you draft, ask the program officer if the guidelines are silent, and disclose accurately if asked. Getting caught not disclosing is a far worse outcome than disclosing.
What is the best AI tool for nonprofits?
For most small organisations, a general-purpose model on its free tier plus Perplexity for sourced research covers the bulk of it. Add a document tool such as ChatPDF or Humata if you regularly wade through long RFPs, and check whether you qualify for donated licences through TechSoup, Google for Nonprofits, or Microsoft before spending restricted funds. The research tools category is a good place to compare options.
Can AI find grants for my nonprofit?
It can shortlist and summarise, but it must not be trusted alone. Models will confidently describe foundations that do not exist or that stopped funding years ago. Use a citing tool like Perplexity so you can click through and verify, and treat a real grant database as the source of truth. A fabricated funder costs a week you do not have.
How do we stop our applications sounding AI-generated?
Add specifics. Numbers from your own records, the name of the referral clinic, the reason the programme changed shape in its second year, a sentence a participant actually said. Generic prose is the tell, not word choice — and a model cannot invent the details it does not have. Also read the draft aloud; the AI cadence is easier to hear than to see.
Put the saved hours somewhere useful
The point of any of this is capacity. A small organisation that stops losing Saturdays to reformatting outcome data for six funders gets those Saturdays back for programme work — which is, after all, the thing the grant is for.
Start with the research tools for the funder work and the writing tools for drafting support. Lean organisations should also see our small business use-case guide and the free tools list.