I use AI in grant work almost every day. So this is not a warning from someone who has not tried it. It is a report from someone who has, in a practice that has secured over $7 million across federal, state, county, and foundation funders.
The short answer: AI is genuinely useful for about a third of the job, actively risky for another third, and irrelevant to the third that decides whether you get funded.
Here is where each line sits.
What AI is actually good at
These are the tasks I hand to it without hesitation, because the failure mode is visible and cheap.
- Compressing a long funding notice. A federal notice of funding opportunity can run eighty pages. Pulling the eligibility criteria, the scoring rubric, the required attachments, and the deadlines into one page is exactly what it does well. I still read the original. I just read it faster, knowing what I am looking for.
- Reformatting narrative you already have. You wrote a strong program description for a county contract. A foundation wants the same program in 500 words under different headers. That is a mechanical transformation of language you have already verified, and it saves real hours.
- Cutting to a word count. Trimming 900 words to 500 without losing the argument is tedious, and AI is fast at it. You still decide what survives.
- Checking a draft against the instructions. Paste in the requirements and the draft, then ask what is missing. It catches the forgotten sub-question buried in part C. This one has saved me more than once.
- Getting past a blank page. A mediocre first paragraph you rewrite beats an empty document you stare at.
Notice the pattern. Every one of those tasks starts from information you already own and already verified. The AI is rearranging, not sourcing.
The four things it will quietly get wrong
Now the part that costs people money.
It invents numbers
This is the one that ends relationships with funders. Ask an AI to strengthen your need statement and it will hand you a statistic with a confident source attached. Sometimes the statistic is real and the source is wrong. Sometimes both are invented, and it reads perfectly either way.
A funder does not experience that as a technology problem. They experience it as your organization putting a number in front of them that is not true. Program officers move between foundations and agencies, and the nonprofit world in any given county is small. That is not a mistake you get to make twice.
So the rule in my practice is absolute: AI never produces a number and never produces a citation. Every figure in a proposal comes from your own program data, a named public dataset, or a source I have opened myself.
It does not know your organization
Left alone, an AI will write that you serve "over 1,000 individuals annually" because that is the shape of the sentence that usually goes there. It has no idea whether you served 1,000 or 140. It does not know your service area is three zip codes rather than a county. It does not know which of your programs is actually funded by the contract you just cited.
Everything specific enough to be persuasive is exactly the material the AI does not have.
It cannot do the budget
Budgets are where applications quietly break, and the break is arithmetic, not prose.
I have worked inside a state budget workbook whose own template miscalculated composite totals across program areas. The formula was wrong in the file the funder distributed. No language model reading the narrative was ever going to catch that. Somebody had to open the workbook, foot the columns by hand, and reconcile them against the narrative.
The ordinary version of the problem is just as costly. The narrative promises a full-time coordinator and the budget funds a half-time one. The indirect rate gets applied to a line the funder excludes. The personnel percentages do not add up across a three-year period. Reviewers find these, and they are the fastest way to lose points that had nothing to do with your program.
It does not see the eligibility gates
This is the most expensive blind spot, and the one nobody writes about.
Most applications that fail do not fail on the writing. They fail at a screen that happens before anyone reads a word of narrative. The organization was not on a required state registry by the deadline. The headquarters address on file did not match the county the funding served. The entity type was wrong for the strategy. A required attachment was missing. The portal closed at 4:59 p.m. and the signed budget arrived at 5:10.
Grants are lost at the screening table more often than they are lost at the scoring table.
An AI will happily help you write a beautiful application for a program you are not eligible for. It has no way of knowing, and it will not think to ask.
The third of the job AI cannot touch at all
Even with a clean draft, a verified budget, and every eligibility box checked, the work that decides the outcome is still human.
Someone has to decide which opportunity is worth your staff's time in the first place. Most organizations lose more to pursuing the wrong grants than to writing weak ones. A $75,000 award that requires a match you cannot raise, imposes reporting your systems cannot produce, and pays for staff you do not have is a net loss even when you win it.
Someone has to call the program officer before the deadline and ask the question the notice does not answer. That conversation routinely changes the application, and occasionally it is the reason you do not submit at all.
And someone has to run the award after it lands. I have taken on an organization where a departed employee had never reported hundreds of thousands of dollars of federally funded work that was actually performed. We pulled every document, reconstructed what was done and by whom, built the record that should have existed, and filed a formal appeal. That is the other half of grant funding, and no drafting tool has anything to say about it.
How I actually use it
For anyone who wants the workflow rather than the warning:
- I read the funding notice myself, all of it, before any tool touches it.
- AI produces a summary of requirements, scoring criteria, and attachments. I check it against the original.
- I decide whether we are eligible and whether the opportunity is worth pursuing. That decision is never delegated.
- I assemble the real material: program data, budget figures, previously approved language.
- AI helps shape and tighten drafts built from that material.
- I verify every number, every citation, and every claim against a source.
- The budget is built and checked in a spreadsheet, by a person, twice.
- I read the whole thing out loud once at the end, because AI prose is fluent enough to hide a sentence that says nothing.
If you take three rules from this, take these. AI never produces a number. AI never produces a citation. You read the funding notice yourself every time, even when you have applied to the same program before.
So can it write a winning grant?
It can write a competent one. Competent applications lose all the time.
What wins is a specific program described in specific numbers, aimed at a funder whose priorities you actually understand, submitted by an organization that clears every eligibility gate, with a budget that ties to the narrative line by line and a plan for the reporting that follows. AI helps with the sentences. It does not help with anything else on that list.
Which is fine. Use it for the typing. Just keep it away from the judgment.
Common questions
Will a funder know I used AI to write the application?
Sometimes, and increasingly they check. AI prose has tells: uniform paragraph length, abstract nouns doing the work of concrete ones, and claims with no source attached. A program officer who reads two hundred applications a cycle develops a good ear for it. The bigger risk is not detection though. It is that generic writing scores in the middle of the pack, and the middle of the pack does not get funded.
Is it against the rules to use AI on a grant application?
It depends on the funder, and the instructions are the only place to find out. A growing number of funders now address AI directly in their guidance, and some ask you to disclose whether it was used. Read the notice of funding opportunity every time rather than assuming last year's rules carried over. Where the guidance is silent, using AI as a drafting aid is generally treated the way a spell checker is. Submitting content you have not verified is a different problem entirely, with or without AI.
Can AI find grant opportunities for me?
Partly. It is useful for summarizing a long funding notice, comparing two programs side by side, or making sense of a funder's 990 filings. It is not reliable for discovery. Ask a chatbot which foundations fund your work and it will produce a plausible list containing real names, dead programs, wrong geographies, and deadlines that passed two years ago. Discovery still runs through funder databases, 990 filings, and the people who know your field.
What tools do professional grant writers actually use?
Less exotic than you would expect. A funder research source, whether that is Candid, Instrumentl, ProPublica's nonprofit explorer, or Grants.gov and your state portal. A spreadsheet for the pipeline. A narrative library, which is just an organized folder of previously approved language, budgets, and attachments. AI sits on top of that stack as a drafting and summarizing aid. The narrative library does more for turnaround time than any AI tool, and it costs nothing but discipline.
If AI can draft it, why hire a grant writer at all?
Because drafting was never the expensive part. The expensive part is knowing which opportunity is worth your staff's time, whether you clear the eligibility gates, what the reviewer is scoring against, whether your budget survives contact with the funder's workbook, and what happens after the award when the reporting starts. AI compresses the typing. It does not compress the judgment, and judgment is what you are paying for.