If you write grants, you have probably felt the whiplash. One headline says AI is going to replace you by Friday. The next says a real professional would never touch it. Neither is true, and neither is much help when you have a deadline on Thursday.

The honest answer is that AI is genuinely good at some parts of grant writing and genuinely bad at others, and the writers who thrive in 2026 are the ones who know the difference. This is not a tutorial on prompts, and it is not a pitch to automate your job out of existence. It is a map. We are going to walk through the grant process one stage at a time and ask the same simple question at each step: should you let AI in here, yes or no, and why. Here is how we actually use it, and where we stop.

A quick note on tools. These days we run most of this through Notion AI, because it works right alongside the docs and data we already keep in Notion. Under the hood, Notion AI draws on several of the common models you have probably heard of, like ChatGPT and Claude. So when we say “ask AI” throughout this post, that is usually what we mean, and the principles hold no matter which tool you reach for.

Funder research and prospecting

AI? Yes, as a research assistant, not the final word.

This is one of the places AI saves us the most time, and we use it in two distinct passes.

First, the wide net. We give AI an organizational profile, a short summary of the organization, its mission, programs, and budget size, along with our parameters for geographic fit and timeline, and ask it to turn up a list of funders that might match. That gets us from a blank page to a working prospect list in minutes instead of days.

Second, the deep dive. Once we have a shortlist, we pull grantee information from each funder’s 990s and hand that to AI for a closer read. We ask it to analyze the funder’s giving by sector and geography, gauge how open they appear to be to new grantees versus the same names year after year, and estimate what a reasonable ask would be for an organization of that size. That last part is the difference between a hopeful application and a well-aimed one.

However, AI will confidently misremember a deadline or a funder’s focus, and it can read a pattern into a 990 that is not actually there. So everything it surfaces is a lead to confirm, not a fact to bank on. We check it against the funder’s own guidelines and filings before it shapes a real decision.

Reading and decoding the RFP

AI? Yes. This is one of the best uses there is.

High value, low risk. Paste in the public guidelines and ask AI to reverse-outline them into a requirements checklist: every question, attachment, word limit, and formatting rule, in order. It turns a forty-page notice of funding opportunity into a punch list you can actually manage.

Here is the specific move we recommend: ask AI to return the key details as a table, and to include the page number where it found each one. Deadlines, eligibility rules, budget caps, required attachments, word and character limits, submission format — one row each, with a page reference beside it. The page numbers are the whole trick. They turn the table from something you have to take on faith into a verification tool, because you can jump straight to the source to confirm a detail or pull more context the moment you need it.

You still read the RFP yourself, but you start from a map instead of a blank page, and you always know exactly where to look.

Organizing discovery and messy notes

AI? Yes. Think of it as your assistant here.

This is the stage where it helps to stop picturing AI as a writer and start picturing it as an assistant, the one who keeps track of the loose ends so your brain does not have to. Grant work carries a heavy mental load: who promised what on which call, the figure someone mentioned in passing, the three things you still need from the program team. That is exactly the kind of weight AI is good at carrying for you.

Our setup starts before the notes are even messy. We run a meeting transcriber or note-taker on discovery and check-in calls, so the conversation is captured without anyone trying to write and listen at the same time. Afterward, we hand the transcript to AI and ask for two things: a list of the open questions still to be answered, and a list of revision points, the spots where something we heard changes the draft or the plan.

That turns a sprawling, hour-long conversation into a short, workable to-do list. You are not leaning on memory or a page of half-legible notes. You are working from a clean account of what was said, what it changes, and what is still missing. And as always, the move is to have AI surface the gaps rather than fill them: ask it for the questions you still need to ask, not its best guess at the answers.

Strategy and the core argument

AI? No. This part is yours.

A clean paragraph is not a winning argument. Deciding what to emphasize for this funder, how to frame the need, which outcomes to lead with, and how to read the politics in the room, that is judgment, and it is the actual job.

There is also a practical, technical reason the strategy has to stay with you, and it is worth understanding. AI is genuinely good at the sentence and the single paragraph. Stretch it much past that and it starts to lose the thread. It does not hold a long argument in mind the way you do. It predicts what comes next from what is closest at hand, so across several paragraphs it tends to drift, repeat itself, and quietly forget the point it was supposed to be building toward. Ask it to write a whole narrative on its own and you often get prose that reads fine line by line but never adds up to a single coherent case.

That is exactly why the architecture has to be human. You hold the through line: the one argument the whole proposal is making, the order the points need to come in, the way each section sets up the next. Once that skeleton is yours, AI becomes useful again at the smaller scale, drafting an individual paragraph you have already framed, or pressure-testing a strategy you have already set by asking what a skeptical reviewer would push back on. The pattern that works is simple. You own the argument, AI helps with the paragraphs, and never the other way around.

Drafting the narrative

AI? Yes for the scaffolding, no for the voice.

This is where the last section pays off. Once you own the argument, AI is genuinely helpful at the paragraph level: roughing out structure and offering phrasing options for points you have already framed and know to be true. What we never do is hand it a blank page and ask for the finished narrative. Generic in, generic out. Left to its own devices, AI prose is competent and forgettable, and reviewers can feel it.

Two things keep our drafts sounding like a specific organization instead of like everyone else’s AI. One is a standard grant narrative: a master version of the organization’s core story, its mission, need, programs, outcomes, and approach, written once and kept current. We give that to AI as the source of truth so it works from the organization’s real language and evidence instead of inventing a plausible-sounding version. The other, when the voice needs to be especially distinct, is a style guide: a short set of notes on tone, the words this organization uses and avoids, sentence rhythm, and how formal or plainspoken to sound. Hand AI that guide and its output stops sounding like a default chatbot and starts sounding like you.

Even with both in hand, the draft AI gives back is a starting block, not a finish line. You take it the rest of the way for nuance, judgment, and the parts of the story only your organization can tell.

Data, outcomes, and evidence

AI? Yes for working with evidence, never for inventing it. This is the section that needs the most care.

Left unsupervised, AI invents statistics, misattributes sources, and will happily produce an outcome number that looks plausible and is completely made up. That risk is real, and it is the part of the process people get burned on. But the lesson is not to keep AI away from evidence entirely. It is to know which jobs are safe to hand it, and to put the right guardrails around each one.

Start with the safest use: organizing evidence you already trust. When you hand AI a curated collection you have vetted yourself, your statistics, study findings, past outcomes, and supporting quotes, it is very good at parsing that material and placing it where it belongs. It will match the right statistic to the right section, pull a supporting detail into the argument it strengthens, and flag where a claim still has no backing. You are not asking it to find truth here, only to organize what you already trust, and that is AI doing what it does best: working with what it is given rather than reaching for what it does not have.

A step riskier, but still useful, is asking AI to help identify new evidence, and here the whole difference is in the instructions. Ask a vague question and you get confident fiction. Give it tight constraints instead, what kind of source actually counts, what time frame is acceptable, a requirement that every claim arrive with a citation you can open and read, and an instruction to tell you when it cannot find something rather than papering over the gap, and it becomes a capable research assistant for surfacing studies and supporting context that you then verify.

Either way, two habits keep this safe. First, the guardrails go in up front, not as an afterthought. Second, you check everything it returns against the original source before it goes anywhere near a proposal. Used that way, AI points you toward evidence faster, but it never gets to be the evidence. And the hard lines never move: it does not supply your outcome numbers, which come from your client or your records, and you never paste confidential or personally identifiable information into a public tool to chase a statistic.

Budgets and budget narratives

AI? Not for the numbers, but yes as a cross-check.

The numbers are yours. A language model is not a spreadsheet, it cannot price your staff time or your supplies, and it should never be the source of a dollar figure. That part does not move.

What AI is good at here is catching what humans miss, and two cross-checks have earned a permanent spot in our process. One is a completeness check: give AI the program narrative and ask it to generate a list of items that might need to be in the budget, things like personnel for the activities described, materials, travel, and evaluation. It is a fast way to catch the line item you wrote about in the narrative but forgot to fund. The other is a consistency check: put the budget and the budget narrative side by side and ask AI to surface discrepancies between them, the figure that does not match its explanation, the cost with no narrative justification, the activity that shows up in one but not the other. Reviewers notice those gaps. Better that you catch them first.

AI can also take a first pass at the words around the math, drafting budget narrative language that you then correct line by line. But treat every figure it touches as a draft to verify, never an answer to trust. Its job here is to be a second set of eyes on your numbers, not the source of them.

Editing, clarity, and the compliance check

AI? Yes. Back in the green zone.

This is the stage AI was practically built for. As a line editor it tightens wordy paragraphs, flags jargon, checks reading level, and re-runs your draft against that requirements checklist from earlier so nothing slips through. A dedicated editing tool like Grammarly fits right in here too.

It also handles the squeeze every grant writer knows, trimming a section to fit a tight word or character limit without gutting the substance, or pointing out where you have room to say more. One caveat worth remembering: AI is not a reliable counter, and it will confidently tell you a passage is under the limit when it is not. Make the cuts with AI, then confirm the real count in your word processor or the application portal before you submit.

Its most valuable editing job is a step up from the sentence: scoring your whole draft against the funder’s own criteria. If the RFP includes a scoring rubric or any guidance on how reviewers will evaluate applications, paste it in alongside your draft and ask AI to assess the draft against it, criterion by criterion, calling out where you are strong and where you are thin. It will surface the requirement you answered in passing but never really made the case for, or the section a reviewer could dock you on. That gives you a reviewer’s-eye read before a reviewer ever sees it.

Through all of it, final judgment on tone and truth stays with you. AI makes the polishing pass faster and sharper, but you are the one deciding what to keep.

Final review and submission

AI? No. The last eyes are human.

Before anything goes out, every claim has to be confirmed true, every number sourced, and the whole application checked against exactly what the funder asked for. The thing that matters most at this stage is accountability: someone has to own that final pass. Not the tool, and not a vague sense that it was probably handled, but a specific person who has actually read it through and signed off. Decide who that someone is well before the deadline, not in the scramble at the end.

So, can AI write your grant?

No. But it can lift real weight off the parts of the process that drain your hours, which frees your energy for the parts that actually win: strategy, story, and judgment. The skill in 2026 is not drafting. It is knowing where to let the tool help and where to keep your hands on the wheel.

Ready to turn this into a system that runs?

This post is the map. If you want the built version, the prompts, templates, and workflow wired together so you are not reinventing the wheel on every application, that is what we put inside the AI Grants Hub™.