How to Use AI to Write a Grant Application

Grant applications are graded against a published rubric by an assessor working through a stack of them. That single fact should shape how you use AI here. The parts of your application that are scored on how clearly you explain something are exactly where a language model earns its place. The parts scored on whether a number is true are exactly where it will lose you the grant.

Split the document that way before you write a word, and everything else follows.

Start with the rubric, not the form

Almost every funder publishes assessment criteria, weightings, or at minimum a set of priorities. Large public portals such as Grants.gov attach the full announcement to each opportunity, and smaller foundations usually state their criteria on the page. Find that document first.

Then do this, in order:

1. Paste the criteria into your AI tool and ask it to restate them as a scoring table: criterion, weight, and what a top-scoring answer would demonstrate. You are checking your own reading, not outsourcing it.

2. Ask it to map each criterion to the section of the application form where it must be evidenced. Funders frequently score something in a section that does not carry its name.

3. Ask what evidence a strong application would include for each criterion. Note which of that evidence you actually have. That gap list is the real output of this step.

That third list is worth more than any draft. Most applications fail because they never address a criterion, not because the prose was weak.

Where AI genuinely helps

Turning your evidence into the funder's language

You know your organisation does outreach in three neighbourhoods. The funder asks about geographic equity of service delivery. Same fact, different vocabulary. Restating what you already do in the terms an assessor is scanning for is the single highest-value use here, and it does not invent anything.

Budget justification prose

The numbers come from your finance person. The paragraph explaining why 0.4 of a coordinator post is the right level of staffing for this activity is prose, and it is prose most people write badly under time pressure. Give the model the numbers and the reasoning and ask for the justification narrative.

Compressing to a word limit

Cutting 900 words to 500 without losing a scored point is tedious and AI is good at it. Give it the criteria alongside the text and instruct it that anything mapped to a criterion must survive.

Answering the same question in four applications

If you apply for several grants a year, your organisational background and track record sections repeat with different framings. Keep one accurate master version, then adapt it per funder. This is close enough to how writing an SOP with AI works that the same discipline applies: one source of truth, many renderings.

A critical read before submission

Ask it to assess your draft against the rubric you built in step one and identify the lowest-scoring criterion. Then ask what evidence would raise it. Assessors are looking for reasons to differentiate applications, and this simulates that reading.

Where it will cost you the grant

Some content in an application is a factual representation to a funder, and in public funding it can carry legal weight.

• Outcome numbers, beneficiary counts, and past performance figures. These come from your records or they do not go in. A plausible-sounding number is the worst possible failure mode here.

• Eligibility statements. Whether you meet a size threshold, a registration requirement, or a geographic restriction is a matter of fact, checked against the guidance, by a person.

• Partner commitments. Never let a draft describe what a partner will contribute until the partner has said so in writing.

• Citations and evidence for need. Models produce convincing references that do not exist, which is a general problem covered in how to stop AI making up citations. In a grant application an invented statistic is not embarrassing, it is a misrepresentation.

• Anything the funder asks you to attest to personally.

A useful rule: if a sentence contains a number, a date, or a name, a human verifies it against a source before submission. No exceptions, however tired you are on deadline day.

A workflow that holds up

The pattern is consistent: the model handles shape and language, you supply and verify substance. That division is the same one that makes preparing a business loan application with AI work, and it fails in the same place when people let it slide.

A worked example of the translation step

The translation step is the one people skip, so here it is concretely. Say you run a small community food project and the criterion reads: demonstrates sustained engagement with underserved populations.

What you have, in your own words: we deliver on Tuesdays and Fridays to four estates, we have done it for three years, about 90 households, most referrals come from two GP surgeries and the local school.

What a good prompt does with that is not embellish it. It asks which parts of that answer the criterion is actually testing, and what is missing. Sustained maps to the three years and the twice-weekly cadence. Underserved maps to the referral routes, because referrals from a GP surgery and a school are evidence of need identified by other services rather than self-selection. Engagement is the weakest of the three, because delivery is not engagement, and an assessor may read it that way.

That gap is the useful output. It tells you the section needs one more fact you already have and did not think to include, perhaps that 30 of the 90 households have been with the project for over a year, or that six volunteers came from recipient families. Neither is invented. Both were sitting in your head, unmentioned, because you did not know the criterion was asking for them.

Run that exercise per criterion and the application improves in a way no amount of prose polishing achieves.

Common rejection reasons AI will not save you from

Worth knowing before you invest a weekend in a draft:

• Ineligibility. Wrong organisation type, wrong region, wrong income threshold. Check first, always, against the guidance rather than the summary page.

• Missing attachments. Accounts, governance documents, safeguarding policies. These are usually pass or fail and no narrative compensates.

• Asking for the wrong amount. Funders publish ranges and typical awards. A bid at three times the typical award needs a reason stated in the bid.

• Costs the fund does not cover. Core costs, capital, existing staff. Many funders exclude these explicitly and applicants still ask.

• Late submission. Portals close on the hour, not at the end of the day.

None of these are writing problems, which is precisely why a strong-looking draft can lull you into skipping the checks that decide the outcome.

On disclosure

A growing number of funders ask whether AI was used in preparing an application, and some prohibit it for specific sections. Read the guidance. Where a declaration is requested, answer it accurately; the risk of an undeclared yes discovered later is far larger than any competitive cost of declaring.

For organisations doing this regularly, the wider toolkit in ai tools for nonprofit organizations covers the reporting side that comes after a successful bid, and using AI across a small business puts grant work in context with the rest of the admin load.

FAQ

Can AI write my whole grant application?

It can produce a complete-looking draft, and that draft will contain unverifiable claims where your real evidence should be. Use it for structure and language, supply every fact yourself, and check the guidance for restrictions before you rely on it.

Will assessors be able to tell AI wrote it?

Detection tools are unreliable, but assessors reliably notice generic applications with no specific local evidence, which is what happens when a model fills gaps you should have filled. Specificity is what scores, and specificity has to come from you.

What is the most useful single prompt?

Ask it to score your finished draft against the funder's published criteria and name the weakest section with the evidence that would improve it. That reproduces the assessor's job more closely than any drafting prompt.

Is it safe to paste our data into an AI tool?

Depends on the tool and the data. Beneficiary information and anything personal needs the same care as any other confidential record, including checking retention and training terms before you paste anything.

Should we disclose AI use to the funder?

If they ask, yes, accurately. If they prohibit it for certain sections, follow that. Undeclared use discovered after an award is a far bigger problem than a declaration.