How to Use AI to Write Grant Applications, Step by Step

Use AI for the mechanical work and keep the evidence yours. Have it extract the funder's criteria and limits, check your fit, draft each answer from an evidence pack you've assembled, cut to the character limit, and then score the draft as an assessor would. You check every figure and tailor the story. The evidence pack is the step that makes the rest work.

The mistake that sinks AI-assisted bids is asking the model to write the application from a project description. It fills gaps with plausible statistics, generic need statements and outcomes you can't evidence, and assessors who read dozens of bids a round spot that immediately. When every claim traces back to a document you supplied, AI becomes a fast, patient drafting partner instead of a liability. Whether funders accept AI help at all is covered in using AI for grant writing without funders rejecting it.

Follow me on Instagram@sagnikteaches

Step 1: Turn the funder's guidance into a criteria table (15 minutes)

Paste the guidance notes, the application questions and the scoring criteria into a chat on a business or non-profit plan. First, check whether the funder says anything about AI use; some require a declaration, and you need to know before you start.

Connect on LinkedInSagnik Bhattacharya

From the funder guidance below, produce a table with:
question number | exact question wording | word or character
limit (and whether spaces count) | scoring criteria that apply |
what evidence the funder explicitly asks for.
Then list: eligibility rules, deadline, required attachments,
anything said about AI use. Quote; do not summarise rules.

Check the table against the original once; models occasionally drop a sub-question. Two rows from an illustrative table for a community fund:

Subscribe on YouTube@codingliquids

The guidance actually asked how you'd involve people in designing and evaluating the project. The table dropped the second half, and an answer written from it would have missed marks on a question the funder had spelled out. One read against the original caught it. What you get is a one-page map you'll use for every later step.

Step 2: Build an evidence pack you reuse for every bid (2-3 hours)

This is the foundation. Collect into one document, with a date on each item:

• Mission and a two-paragraph description of what you do, in your own words
• Numbers: people or animals served per year, sessions delivered, waiting lists, volunteer hours
• Outcomes you've measured, with how you measured them
• Three to five short beneficiary quotes you have permission to use, names removed
• Your most recent accounts summary and the project budget
• Previous monitoring reports to funders
• Staff and volunteer roles relevant to the project
• Policies funders ask for: safeguarding, equality, environmental, data protection

Give every item a date and a source, so a figure can be traced in seconds when an assessor or auditor asks. A few illustrative lines from the animal rescue's pack:

• Animals rehomed, 2025: 212. Source: rehoming log export, January 2026.
• Foster carers active: 11, placing 26 animals last year with one part-time coordinator. Source: foster rota, March 2026.
• Vet support: two practices give reduced rates for rescue animals. Source: rate letters on file. No formal partnership agreement, so don't call it one.
• Quote (permission given, name removed): "I didn't think I could foster with a full-time job. The coordinator made it work around my shifts."

Store it in a Project (ChatGPT Business, Claude Team) or a Gem (becoming a skill from November 2026), or as a source in Gemini Notebook (formerly NotebookLM), so every drafting chat can see it. Remove anything that identifies an individual. Update the numbers every quarter and the pack stays ready.

Step 3: Check the fit before you write a word (10 minutes)

Using the criteria table and our evidence pack, assess our fit.
For each eligibility rule and funder priority: Met / Partly / Not met,
with the evidence from our pack. Then list the weakest area and
what evidence would strengthen it. Be blunt; don't reassure me.

A small community music charity running this against an arts fund found (illustrative) that it met every priority except "reaching people who don't usually take part in the arts", where its pack had no data on who attended. That was fixable: a short attendance survey from the last term gave them the numbers. Without the check, they'd have discovered the gap in the rejection letter.

Step 4: Plan each answer against the scoring (20 minutes)

For every question, ask for an outline: the points that will score, in order, and which evidence supports each. Answers that follow the funder's own criteria order are easier to score, and assessors notice. Edit the outline before drafting. This is where you add what AI can't know: the story behind the project, the local relationship that makes it work, the reason now.

For the rescue's "why are you best placed to deliver this?" question, the AI outline listed years operating, volunteer numbers and animals rehomed. All true, and all things any rescue could say. The coordinator's edit moved one point to the top: the two vet practices that already give reduced rates, which is why the budget per animal is lower than a new service could manage. Then a second: the waiting list of people who have asked to foster but can't be trained with one coordinator. Neither was obvious from the pack's headings, and both answer the question better than a longer history.

Step 5: Draft section by section, to the limit (2-4 hours)

Draft the answer to Q3 using the approved outline and ONLY facts
from our evidence pack. Limit: 1,500 characters including spaces.
- Every number must appear in the evidence pack. If a point needs
  a figure we don't have, write [NEED DATA: ...].
- Plain English. Active voice. No superlatives ("vital",
  "unique", "life-changing") unless quoted from a beneficiary.
- Answer the question as asked; follow the outline order.

Models are poor at counting characters, so paste each draft into your word processor's counter or the funder's portal preview. If it's over, ask for "the same answer in 1,350 characters, keeping all evidence", which leaves room for your edits.

Measure in the portal itself before you call an answer finished. In one illustrative case, an answer measured 1,480 characters in the word processor and 1,532 in the funder's portal, which counted each paragraph break as characters. The portal refused to save it the night before the deadline. Paste each answer in as soon as it's drafted, not on submission day.

One more rule belongs in every drafting prompt: beneficiary quotes stay word for word. Given the quote "I didn't used to go out. Now I've got somewhere to be on Tuesdays," an illustrative draft "tidied" it into "The programme has given me a renewed sense of purpose and routine." The rewrite is grammatical, generic and no longer something the person said. Add "reproduce quotes exactly, including informal grammar" to the prompt, and check each quote against the pack yourself.

A before and after on a need statement, from an illustrative adult literacy charity. The generic first draft, from a prompt without an evidence pack:

Literacy is a vital life skill, yet millions of adults struggle with reading, which impacts every aspect of their lives. Our unique programme transforms lives by helping every learner reach their full potential.

The draft from the evidence pack:

Last year 64 adults joined our reading groups; 41 more are on the waiting list, 18 of them for over six months. Most are referred by an employment adviser or a housing officer after struggling with forms. In our end-of-course survey, 52 of 58 respondents said they could now complete a basic form without help. This grant would fund a second weekly group and cut the wait to under two months.

The second version has no adjectives doing the work. The numbers do it, and an assessor can score them.

Step 6: The budget and the budget narrative (45 minutes)

Never let AI invent a cost. Build the budget yourself in a spreadsheet, then use AI to check it against the narrative:

Here is our budget table and our project description.
1. Check every cost in the narrative matches the table.
2. Check totals add up.
3. List any activity in the narrative with no budget line,
   and any budget line not explained in the narrative.

When an animal rescue (illustrative) ran this check on a $15,000 bid for a foster-carer programme, it found that the narrative promised "training for 20 new foster carers" but the budget only costed venue hire for two sessions of eight, and that the $1,200 for vet checks wasn't mentioned in the narrative at all. Both were fixed before submission. Mismatches like these are among the easiest ways to lose marks, and among the easiest for AI to spot.

Give the check the funder's budget rules as well, because that's where a quick sum saves a rejection. Suppose the guidance caps overheads at 10% of the grant. On the $15,000 bid, the rescue had put in $2,100 for a share of rent, insurance and bookkeeping: 14%, or $600 over the cap. Pasting the rule into the prompt ("flag any line that breaks these funder rules") caught it. The rescue cut the overhead line to $1,500 and said in the narrative that it would cover the remaining $600 from its own unrestricted funds.

Step 7: Trace every figure and claim (30 minutes)

List every number, statistic, date, name and factual claim in
this application. For each, say where in our evidence pack it
comes from, quoting the source line. Mark any you cannot find
as UNSUPPORTED.

An illustrative extract of what comes back for the rescue's draft:

• "212 animals rehomed last year": evidence pack, animals rehomed 2025. Supported.
• "working in partnership with two local vet practices": pack says reduced rates, no partnership agreement. Partly supported; reword as "two vet practices give us reduced rates".
• "one in four pet owners has considered giving up a pet": UNSUPPORTED.
• "rescuing animals since 2011": pack says the charity registered in 2012. Mismatch; check which date you mean and use it everywhere.

Then look at each UNSUPPORTED item yourself. Typical finds: a national statistic the model "remembered" (remove it or find the real source), a rounded number that's now wrong ("nearly 100" for 64), and a partner organisation mentioned without their agreement. Every claim in a grant application is something you may be asked to report against, so if you can't evidence it, don't submit it. Catching made-up figures in AI-drafted proposals goes deeper on this step.

Step 8: Let AI play the assessor (20 minutes)

Open a fresh chat so the model has no stake in the draft, paste the scoring criteria and the full application, and ask:

You are an experienced assessor for this funder. Score each
answer against the criteria using the funder's scale. For each:
the score, the strongest sentence, the weakest point, and what
evidence would raise the score. Do not rewrite the answers.

For the animal rescue, an illustrative response scored the "difference your project will make" answer lowest: "States activity (foster placements) but not outcome. What changes for the animals or the community? No target." They added a target (40 more animals placed in foster homes within the year, based on last year's 26 with one coordinator) and the measure they'd use. Treat the scores as a prompt for thinking, not a prediction; the model doesn't know how this funder's panel actually scores.

If every answer comes back scored 4 or 5 out of 5, the model is being polite, not useful. Ask instead: "Rank the six answers from weakest to strongest, and for the weakest two, say which sentence an assessor would underline in red." A forced ranking always produces a bottom answer, and that's the one to spend your remaining hour on.

Step 9: Voice, portal and submission checks (30 minutes)

• Read it aloud. Any sentence you wouldn't say to the funder on the phone gets rewritten.
• Paste into the portal early. Portals strip formatting and count characters differently; bullet points can vanish.
• Check attachments against the criteria table from step 1.
• Declare AI use if the funder asks.
• Save the final answers into a library of approved paragraphs, tagged by topic, for the next bid.

A library entry is more useful with a little context attached. An illustrative example:

TOPIC: Need / waiting list (adult literacy)
Approved text (612 characters): "Last year 64 adults joined our
reading groups; 41 more are on the waiting list..."
Figures as of: March 2026 (update before reuse)
Used in: community foundation bid (funded), arts and learning
fund (not funded; feedback: strong need, weak outcomes)
Evidence source: evidence pack s.2, waiting list export

The "figures as of" line is the one that saves you: reusing a paragraph with last year's numbers is an easy way to submit something untrue without meaning to. Before each new bid, ask AI to list every library paragraph you plan to reuse whose figures are more than six months old, then update those numbers from the current evidence pack first. The "used in" line also tells you which paragraphs have been in successful bids, which is more reliable than any score a model gives your draft.

How the hours compare on a real-sized bid

For the illustrative animal rescue's application (six questions, a budget and three attachments), the time looked like this:

The second application from the same pack took under six hours. The saving compounds, which is why the pack is worth building properly.

What a language café for newcomers changed in its second bid

A different illustration: a volunteer-run language café that helps adults new to the area practise conversation had been turned down twice with feedback that its outcomes were "unclear". The fit check in step 3 flagged the same weakness: its evidence pack described sessions, not results. The volunteers added a simple before-and-after self-rating question at registration and after eight weeks, collected a term's worth, and rebuilt the outcomes answer around it: "Of 37 learners who completed eight weeks, 29 rated their confidence speaking to a stranger at 4 or 5 out of 5, up from 6 at the start." AI didn't win that bid. Better evidence did, and AI made it quick to spot the gap and rewrite around the fix.

Which AI tool to use for this

Any of the main assistants will do the steps above; what matters more is a plan that keeps your documents private and lets you store the evidence pack. Several vendors discount for non-profits: OpenAI for Nonprofits offers ChatGPT Business Standard seats at $8 a user a month billed annually ($10 monthly), Claude for Nonprofits prices the Team plan at $8 a user a month, and Google Workspace for Nonprofits now includes the Gemini app and Gemini Notebook at no extra cost. Eligibility rules and verification vary, so check each programme's terms. Specialist grant tools add funder databases and application tracking; AI grant writing tools compared for small non-profits weighs them against a general assistant, and turning charity data into impact reports covers the reporting you'll owe once the grant lands.

Grant writers' follow-up questions

Should I tell the funder I used AI?

Follow the funder's guidance first; some ask applicants to declare AI use and a few restrict it. Where there is no rule, a short honest line is rarely a problem if asked, because funders assess the project and your evidence, not the typing. What they do object to is generic text and claims that don't hold up, both of which this method is designed to prevent.

Can I reuse the same AI draft for several funders?

Reuse your evidence pack and approved paragraphs, not whole applications. Each funder's priorities and scoring differ, and assessors who read many bids recognise boilerplate. Start each application at the criteria step, so the draft answers that funder's questions in that funder's order, and adapt your library paragraphs to fit.

Is it safe to put beneficiary information into AI?

Only anonymised, aggregated information: numbers served, outcomes, and quotes you have permission to use with names removed. Use a business or non-profit plan that doesn't train on your content, and never paste case files, safeguarding details or anything that could identify a person. Check your data-protection policy if in doubt.

How long does an application take with AI?

Once the evidence pack exists, a mid-sized application with five or six questions typically takes six to ten hours instead of fifteen to twenty. The first application takes longer, because building the pack takes two to three hours. The time saved is mostly in drafting and cutting to length; checking figures and tailoring to the funder still take as long as they should.

Further reads

• Donor Thank-You Letters With AI That Still Feel Personal — Thank individual donors in a way that still feels personal.
• AI Tool Discounts for Charities and Non-Profits: What to Claim — Non-profit pricing on the AI tools used in this method.
• A Five-Minute Fact-Check Routine for AI Output Before It Goes Out — A five-minute routine for checking any AI draft before it goes out.
• How Small Charities Can Use AI With Almost No Budget — Doing all of this on free and discounted tools.
• Claude vs ChatGPT for Business Writing, Proposals and Reports — Which assistant drafts long-form applications better.
• How to Write an AI Policy for a Small Charity — Write down how your team uses AI in fundraising.
• Do Small Charities Need Outside Help to Adopt AI? — What a small charity can do with AI on its own, the five signs it needs outside help, and how to brief that help so the work stays yours.
• Charity Appeals and Social Posts With AI That Keep Your Voice — How a small charity turns past appeals into a voice card, briefs AI with a real consented story, and edits drafts until they sound like itself.
• Can Farmers Use AI for Farm Records and Paperwork? — How farmers can hand the typing of spray logs, receipts and forms to AI, and the records that must still be checked against the source.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: vendor plan pages; OpenAI help page on OpenAI for Nonprofits, Claude for Nonprofits guide and Google for Nonprofits help page (checked September 2026). The application example, figures and outputs are illustrative.