Can you use AI to write grant proposals?

The 80-Hour Proposal Problem

A single federal grant application takes 80 to 120 hours to complete manually, according to an analysis of AI grant writing tools. That's two to three full work weeks of one person's time, poured into a document that may never get funded. Multiply that across a pipeline of opportunities, and the math gets grim fast.

So where do those hours actually go? The research points to three time sinks: literature review, formatting, and boilerplate. These are the mechanical, constrained tasks that eat days without requiring much strategic thinking. They're also exactly where AI helps most. Studies cited across multiple platforms report roughly 60% time savings in proposal preparation, concentrated in those three areas, and user-reported data across tools shows a 40 to 70% reduction in drafting time.

That efficiency gain changes the funding math in a subtle way. When one proposal costs 100 hours, you can afford to submit a handful per year and hope one lands. When AI compresses the cycle, volume becomes the strategy. Researchers using these methods report submitting 10 proposals in the time it used to take to write 3, a 3.3x increase in output. One case study showed a Grantable user recycling 60% of a winning NIH grant into an NSF application, cutting prep from 30 days to 3.

Volume beats polish because grant funding has always been partly a numbers game. Even excellent proposals get rejected for reasons unrelated to quality: fit, timing, reviewer fatigue, budget cycles. The only reliable lever most applicants control is how many well-targeted submissions they put into the pipeline. Trained, strategic AI users report 22% higher success rates, which correlates with that increased volume.

The catch is that not all hours are equal. Reviewers can spot generic AI output, and one anonymous NIH study section member described the problem bluntly: AI-assisted proposals are "all bad in the same way. Same cadence. Same excessive confidence." Speed without substance gets rejected, as the Vercor analysis puts it. The hours you save on formatting and boilerplate should be reinvested in the parts that actually differentiate a proposal: the specific aims, the strategic framing, the authentic organizational voice.

This is where AI agents earn their keep. A well-configured agent can research funders, synthesize literature, draft boilerplate sections, and then run adversarial review against your own draft, attacking weak arguments the way a skeptical reviewer would. That last task, using AI to critique rather than generate, is the workflow Stanford researchers recommend. If you're building that kind of agent workflow, agent skills can automate the research, drafting, and review steps; you can browse and test free ones on SkillzBench, where every skill is security-scanned and benchmarked so you know which ones actually lift performance before you plug them into your agent.

What AI Can and Can't Write for You

Here's the section:

Most grant teams now use AI somewhere in the writing process, but the difference between a proposal that sails through review and one that gets flagged comes down to which tasks you hand over. Stanford Medicine researchers drew a clear line: use AI for efficiency, never as a substitute for the scientific process of asking impactful questions and testing hypotheses.

The safe zone is mechanical and research-heavy work. Literature synthesis tools like ResearchRabbit and Elicit handle the grind of scanning papers, while AI also excels at citation formatting, grammar, budget templates, and mock review that stress-tests your logic before a real panel does. A single federal grant application runs 80 to 120 hours manually, so these gains compound fast.

The risky zone is anything that carries your intellectual fingerprint. Specific aims, innovation statements, methodology design, and hypothesis generation belong to you. Stanford's rule is blunt: never let AI write the first draft of core scientific content. Draft it yourself, then use AI adversarially to critique and attack it. An NIH study section member described AI proposals as "all bad in the same way. Same cadence. Same excessive confidence."

That distinction matters because funder policies now diverge sharply:

• NIH (September 2025): "substantially AI-developed" proposals face automatic rejection, with a six-application-per-year cap when AI is detected
• NSF: requires transparency and disclosure
• ERC: permissive for brainstorming and summarization
• UKRI: evaluates case by case

Foundations are stricter still. Only 10% explicitly accept AI-generated proposals, and a 2024 Candid survey found 97% of U.S. foundations aren't using generative AI to screen applicants yet.

The workflow that works: let AI research, format, and edit, then ground every draft in your organization's own data. Generic output reads as generic, and reviewers notice. If you want to automate the repetitive parts of that pipeline, agent skills can handle the rest of the workflow, and you can browse and test free ones on SkillzBench before trusting them with your process.

Funder Rules You Have to Check First

Before you let an AI agent loose on a grant proposal, you need to know what the funder will tolerate, because the rules vary wildly and the penalties are real. The same draft that's perfectly acceptable for one agency can get your application thrown out by another.

The strictest policy comes from the NIH. As of September 2025, proposals deemed "substantially AI-developed" face automatic rejection, with a cap of six applications per year when AI use is detected and violations referred to the Office of Research Integrity, according to policy analysis of major funders. That's a career-level consequence, not a slap on the wrist.

Other funders take softer positions, and the differences matter for how you structure your workflow:

• NIH: restrictive. AI-developed proposals can be rejected outright, so keep human authorship over core scientific content.
• NSF: transparency required. You can use AI, but you must disclose it.
• ERC: permissive for brainstorming and summarization tasks.
• UKRI: case-by-case evaluation, so check before each submission.

Foundations are a murkier picture. A 2025 Candid survey found that only 10% of foundations explicitly accept AI-generated proposals, yet most admit they can't reliably detect AI-assisted submissions, as reported in guidance for nonprofits. Detection tools don't solve this either. False positive rates run from 7.2% to 26%, which means a fully human-written proposal can still get flagged. That gray zone is exactly why the safest move is disclosure rather than concealment.

So what does this mean in practice? You can still use AI heavily for research, drafting, and editing. You just have to match the task to the funder. Literature synthesis, formatting, and mock review are safe almost everywhere. Core scientific content, hypothesis generation, and specific aims are where the NIH draws a hard line, and where Stanford Medicine researchers advise keeping humans firmly in charge.

If you're building an AI-assisted grant workflow with agent tools, the same discipline applies: verify what each tool actually does before you trust it with a compliance-sensitive task. SkillzBench's free directory benchmarks agent skills in a sandbox and publishes the measured Lift, so you can see which ones genuinely help with research and drafting before they touch anything submission-ready. Check the funder's current policy first, then choose your tools second. The order matters.

A Workflow That Holds Up to Review

The proposals that survive review aren't the ones AI wrote. They're the ones where AI played a supporting role and a human kept control of the argument. The difference shows up in the workflow.

The core rule comes from Stanford Medicine's peer-reviewed guidance on grant writing: never let AI write the first draft of core scientific content. You write the aims, the hypothesis, and the strategy yourself. Then you flip the relationship and use AI adversarially, asking it to attack your draft, find the weak logic, and flag the claims a reviewer would question. This matters because reviewers are already tired of the pattern one anonymous NIH study section member described: AI proposals that share "the same cadence, same excessive confidence," with none of the messy texture of real science.

Grounding is the second pillar. Generic AI output "reads as exactly what it is. Reviewers notice," as one evaluation of grant writing tools puts it. Feed the model your past proposals, needs assessments, program outcomes, and partner names before asking for anything. The tools that draft from your organization's actual data produce content that sounds like you. The ones drafting from generic training data produce content that sounds like everyone else's.

The third pillar is the "clean room" protocol used at the University of Bath: use AI for research and planning, then write the proposal itself in a separate, AI-free environment. This gives you a clear audit trail, which matters as funder policies harden. NIH now treats "substantially AI-developed" proposals as grounds for automatic rejection, and only 10% of foundations explicitly accept AI-generated proposals, according to a 2025 Candid survey.

A workflow that holds up looks like this:

• Research phase: AI handles literature synthesis, funder scanning, and formatting. This is where the reported 40-70% drafting time savings actually come from.
• Strategy phase: human only. You decide the aims, the framing, and the ask.
• Adversarial phase: AI critiques your draft, hunts for weaknesses, and runs mock review.
• Voice phase: a human editor rewrites anything generic, and someone not involved in drafting reviews the final version.
• Compliance phase: verify the specific funder's AI policy before submission, every time.

One practical note for teams running this kind of multi-step workflow: agent skills can automate chunks of it, like the literature scan or the mock-review pass. If you go that route, vet them first, since published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench's free directory benchmarks and security-scans skills so you can pick ones that actually lift your workflow. Then get back to the part only you can do: the science, and the voice.

Automating the Rest of the Pipeline

Writing the proposal is only half the work. A single federal grant application takes 80 to 120 hours to complete manually, and a large share of that time goes to everything around the writing: hunting for opportunities, logging activity in your CRM, chasing follow-ups, and producing post-award reports (Vercor's 2026 software guide). This is where agent skills shine, because these steps are mechanical, rule-based, and low-risk in a way that core scientific content never is.

Think about what actually eats your week. Grant discovery means scanning funder databases and matching eligibility. CRM logging means copying deadline dates, contact names, and submission statuses into records. Reporting means pulling the same outcome data into the same templates every quarter. None of it requires judgment about your mission. All of it rewards automation.

Agent skills (SKILL.md instruction packages that plug into Claude Code, Codex, Cursor, Gemini CLI, or chat-app agents) can run each of these steps on a schedule or on demand. A discovery skill can flag new opportunities matching your program areas. A CRM skill can log every submission touchpoint automatically. A reporting skill can draft the boilerplate sections that never change, leaving you to fill in the actual outcomes. The same logic that makes AI safe for formatting and citation management makes it safe here: the output is easy to verify, and a mistake is annoying rather than disqualifying.

The catch is quality control. Anyone can publish a skill, and published research has found that roughly 26% of skills carry at least one vulnerability. You would not install a random browser extension on the machine where your funder data lives, and skills deserve the same caution. That's the problem SkillzBench was built to solve. It's a free, independent directory that continuously crawls skills from GitHub, pins every version to a commit SHA, runs security scans and license checks, and benchmarks each skill in a sandbox with and without it, publishing the measured Lift in percentage points. Every skill carries a 0-100 trust score built from security, quality, license, maintenance, and community signals, and the Lift is always shown separately so money never touches the numbers.

Practically, here's how to approach the pipeline:

• Start with discovery skills, since bad matches waste the most downstream time.
• Check the trust score and benchmark Lift before installing anything, not after.
• Use the free MCP connector to give each agent only the skills you've chosen for its role.
• Keep human review on anything a funder will see, especially given diverging AI policies.

That last point matters more than it sounds. Only 10% of foundations explicitly accept AI-generated proposals, according to a 2025 Candid survey, and NIH now automatically rejects proposals it deems substantially AI-developed. Mechanical pipeline automation keeps you on the right side of that line, because it touches process, not prose. The writing itself stays yours, and the calendar reminders, CRM entries, and report scaffolds handle themselves.

You can browse and test the current crop of skills free at SkillzBench, including per-agent loadouts that let you toggle exactly which skills each agent can see. Pick the two or three that save you real hours, verify their scores, and let the mechanical half of grant work run itself.