A grant writing AI agent can draft narrative sections from your notes, turn a funder's guidance into a requirements checklist, tighten answers to word limits and critique a draft against the scoring criteria. A person must still verify eligibility, every figure and claim, the funder's current rules and the declarations someone signs.
"Agent" is mostly a label. What works for a small firm is a reusable assistant inside a ChatGPT Project, a Claude Project or a Gemini Gem (becoming a skill from November 2026), loaded with your fact sheet, past bids and the call guidance. OpenAI is retiring custom GPTs, which are scheduled to stop running on 11 December 2026, so build in a Project instead. The example here is a business grant, not a charity pitch: a nine-person specialist food importer applying for equipment funding.
Assistant or agent: what a bid actually needs
An assistant answers inside a conversation, using the files and instructions you have loaded. An agent works through several steps across apps and files on its own. OpenAI's ChatGPT Work plans and carries out longer tasks across apps, files and a browser (full access from Plus, limited desktop access on Free and Go), and Anthropic's Claude in Chrome extension, on paid plans, reads, clicks and navigates websites for you. Both are real, and neither is what a grant bid mostly needs. The expensive mistakes in a bid come from wrong facts, not from slow clicking.
| Job in the bid | Assistant is enough | Agent may help | Human only |
|---|---|---|---|
| Turn the call guidance into a requirements matrix | Yes | No | Check the result |
| Draft the project description and need statement | Yes | No | Edit and approve |
| Collect past awards and similar funded projects | Partly | Yes, with sources listed | Verify every example |
| Gather attachments from your own drive | No | Yes, with approval of each action | Confirm they are current |
| Decide whether you are eligible | No | No | Yes |
| Build the budget and match funding | Wording only | No | Yes, in a spreadsheet |
| Sign declarations and submit in the portal | No | No | Yes |
That pattern holds on every platform: let the software do the reading and the first drafting, and keep the decisions, the numbers and the signature with a named person.
What it drafts well, ranked by the time it saves
These are the jobs where a loaded assistant earns its place in a small firm's bid, roughly from biggest saving to smallest. For each, the right-hand column says what you must check.
| Task | What you give it | What you check |
|---|---|---|
| First draft of the project description | Your fact sheet, project notes and the question wording | Every claim against the fact sheet; invented outcomes |
| Requirements matrix from the call | The guidance PDF and any FAQ | Mandatory items against the original text |
| Fitting answers to word limits | Your full answer and the limit | That nothing material was cut |
| Critique against the scoring criteria | Your draft and the published criteria | Whether its scores are plausible, not accurate |
| Work plan and timeline table | Milestones and dates you have confirmed | Dates against supplier lead times |
| Wording of the budget justification | The finished budget spreadsheet | Each number against the sheet |
| Risk register | Your own list of risks and mitigations | That each mitigation is something you will really do |
| Letter-of-support requests | Names, roles and what you need from each | Tone and who signs |
The critique job deserves a special mention. Ask the assistant to mark your answer against the funder's published criteria, one score and one sentence of reasoning per criterion, and to list the evidence missing from the answer. Its scores will not match the funder's panel, but the missing-evidence list is often exactly what you forgot to attach. The series covers the wider process in how to use AI to write grant applications step by step.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA reusable bid assistant: the project set-up to copy
Build one project per funder call, or one per firm if you apply often, and load five kinds of file. The fact sheet is the most important: one dated page of figures the firm stands behind (headcount, turnover band, trading start date, products, certifications, past grants). If a number is not in that file, the assistant has no business using it.
- Fact sheet: dated, one page, owned by a named person.
- The call guidance: the full PDF plus the FAQ and any scoring criteria.
- Past applications: successful ones first, marked with the funder and year.
- Budget workbook export: the final spreadsheet saved as a PDF or CSV.
- Style rules: plain English, your spelling, banned phrases and the firm's tone.
The platforms differ in what they allow.
| Platform | Where the set-up lives | Limits worth knowing | Sharing |
|---|---|---|---|
| ChatGPT Project | Project instructions plus uploaded files | Files per project: 5 on Free, 25 on Go and Plus, 40 on Pro, Business and Enterprise | Open to signed-in users, subject to plan and workspace settings |
| Claude Project | Project knowledge plus project instructions | Free accounts can create up to five projects; paid plans add retrieval that expands capacity | Team and Enterprise only |
| Gemini Gem | Instructions plus files | Becoming a skill: personal-account Gems turn read-only on 17 November 2026 and are removed on 17 December 2026, Workspace business accounts follow in March 2027, so build a new personal set-up as a skill | Recipients can see the instructions and files |
Paste something like the following into the project instructions, then edit it for your firm. It tells the assistant what it may not do, which matters more than what it may.
You help [company name] draft grant applications. Rules:
1. Use only facts that appear in the fact sheet, the call guidance or the
budget file. Never invent figures, dates, jobs created or outcomes.
2. After every paragraph, list the source file for each claim in brackets.
3. If a needed fact is missing, write [CHECK: what is missing] instead of
guessing.
4. Never state an award amount. Quote only what the call guidance says,
and flag any calculation for a person to confirm.
5. Keep within the stated word limit and answer the question asked.
6. British English, plain words, short sentences, no sales language.
7. At the end, list every number you used and the file it came from.If you built something similar as a custom GPT, copy its instructions and files out now: existing GPTs are scheduled to stop running on 11 December 2026, and OpenAI's own route is a "Migrate to plugin" option where your account offers it. The guide to free AI plans for grant writing covers what the free tiers allow if you are starting without a paid plan.
A worked bid: a food importer and a cold-store grant
Take a hypothetical call, invented for illustration. It funds up to 50 per cent of eligible costs with a maximum award of $20,000. Applicants must have fewer than 50 employees and at least two years of trading. Equipment and installation are eligible, staff training is not, and any cost incurred before the award date is ineligible. The applicant funds the rest in cash and finishes within 12 months.
The applicant is a specialist food importer with nine staff, trading for six years, replacing its cold store. The equipment costs $41,000, installation $4,000 and staff training $3,000, a project cost of $48,000. Last month it paid a $5,000 deposit on the equipment. The correct sum runs like this. Eligible costs are $41,000 plus $4,000, which is $45,000. Remove the $5,000 pre-award deposit and eligible costs are $40,000. Half of that is $20,000, exactly the cap. The firm asks for $20,000 and funds $28,000 itself.
Now see what an assistant that was not given the call's rules in a form it could apply wrote in its first draft, and what a person has to fix.
| Claim in the AI draft | The check | Corrected position |
|---|---|---|
| "We request $24,000, which is 50% of our $48,000 project." | The percentage applies to eligible costs only, and the award is capped | Request $20,000 |
| "We will contribute $24,000 from reserves." | Project cost $48,000 less the $20,000 award | Contribution is $28,000, backed by a bank statement |
| "Staff training is included in this request." | The call says training is ineligible | Show the $3,000 as firm-funded, outside the request |
| "Work will begin once the award is confirmed." | A $5,000 deposit was paid last month | Disclose it as pre-award spend, excluded from eligible costs |
| "The upgrade will create four new jobs." | Not in the fact sheet or project plan | Delete, or keep only if the firm commits and can evidence it |
Five errors in five sentences, and every one of them reads smoothly. That is the real limit of a drafting agent: it is fluent whether or not it is right, and a grant panel marks the figures, not the fluency. A food importer would catch the $24,000 only by doing the sum above, which is why the budget lives in a spreadsheet and the assistant only words what the spreadsheet says.
Eligibility is a human check, done before any drafting
Do this test first, in writing, before a single paragraph is drafted. An assistant can extract the criteria from the call into a table, but a person fills in the answer column and attaches the evidence. For the importer it looks like this.
| Criterion in the call | Our answer | Evidence to attach | Result |
|---|---|---|---|
| Fewer than 50 employees | 9 | Latest payroll summary | Pass |
| Trading for at least two years | Six years | Incorporation certificate | Pass |
| Project finished within 12 months | Five-month install | Installer's programme in the quote | Pass |
| No other support for the same costs | None | Written confirmation from the owner | Pass |
| No costs before the award date | $5,000 deposit already paid | Supplier invoice and payment date | Partly: exclude $5,000 |
| Match funding of at least half | $28,000 in cash | Bank statement | Pass |
The row that failed is the kind a model reading a long brief can easily miss. Ask the assistant for the matrix, then verify each row against the original text yourself. The same discipline applies in other sectors, as the series' farm grant application tutorial shows with a land-based business.
The requirements prompt, and what you would fix in its answer
The first job to hand the assistant is reading the call. The prompt matters because it forces the model to quote rather than paraphrase, which makes every row checkable in seconds.
Read the attached call guidance. Produce a table with these columns:
requirement, exact wording from the guidance, section number, mandatory or
optional, and the evidence we must supply. Quote the guidance exactly. If
anything is unclear, write [CHECK] and say why. Do not decide whether we
are eligible.This is the kind of output you will get back, shortened and illustrative.
1. Size | "fewer than 50 employees" | 2.1 | Mandatory | Payroll summary
2. Trading | "at least two years of trading" | 2.1 | Mandatory | Incorporation certificate
3. Eligible costs | "Equipment and installation are eligible; staff training is not" | 3.2 | Mandatory | Itemised quotes
4. Timing | "Costs incurred before the award date are ineligible" | 3.4 | Recommended | Dated invoices
5. Award | "up to 50% of eligible costs, maximum $20,000" | 4.1 | Mandatory | Budget workbookRow 4 is the typical slip. The model softened a rule to "Recommended", although the sentence it quoted plainly says costs before the award date are ineligible. You catch it by reading the quoted wording, which is why the prompt demands exact text and a section number: they turn the check into a ten-second lookup instead of a re-read of the whole guidance. If the call is a short form with six questions, you do not need any of this. Draft the answers yourself and keep the assistant for polishing.
Figures: every number gets a source or a flag
Keep the arithmetic out of the model. Build the budget, the match funding and the award calculation in a spreadsheet you control, export it, and load the export into the project. Then make the assistant earn its keep on wording. After it drafts, run a reconciliation: ask for a table of every number in the narrative, then compare it with the budget line by line. This is the table for the importer's first draft.
| Item | Budget file | Number in the narrative | Match? |
|---|---|---|---|
| Cold-store equipment | $41,000 | $41,000 | Yes |
| Installation | $4,000 | $4,500 | No: the draft rounded up |
| Staff training | $3,000 | $3,000 | Yes, but it sits in the wrong column |
| Total project cost | $48,000 | $48,500 | No: follows the installation error |
| Award requested | $20,000 | $24,000 | No: the percentage was applied to the wrong base |
Two habits make this routine. First, the instruction to write [CHECK] instead of guessing turns silent errors into visible gaps, which are far cheaper to find. Second, never accept a figure from the assistant that did not come from a file: if it says the sector is growing by a stated percentage, ask for the source, and expect that it cannot give one. The series covers the discipline in detail in its tutorial on drafting a budget narrative with AI and reconciling every figure.
Where the draft goes wrong: five patterns and their fixes
Invented outcomes. The importer's draft promised four new jobs because models tend to supply a measurable result whether or not one exists. Fix: outcomes come only from the fact sheet or the project plan, and the firm commits to each in writing.
Generic need statements. "Customers increasingly demand fresh, ethically sourced produce" could be pasted into any bid. Fix: replace it with the firm's own evidence, such as the number of spoiled pallets last winter or the orders it turned down for lack of cold space, and cite the file.
Mismatch with the scoring criteria. A fluent answer can spend 400 words on history when the criteria give most marks for delivery risk. Fix: give the assistant the criteria with their weightings and ask where each paragraph earns marks.
Stale rules. A model may describe last year's call from memory. Fix: use only the current guidance loaded in the project, and read the original for deadlines, caps and eligible costs.
Unsourced statistics and references. Models can produce a plausible figure with a plausible source that does not exist. Fix: delete any statistic you cannot trace to a document you have opened, and prefer the firm's own data.
Funder rules on AI: read the call, declare when asked
Funders are not uniform on AI, and their rules are still changing. A call may ask you to declare AI use, limit it, or say nothing at all. Read the current call for any such statement, and keep a simple log of which sections the assistant helped with, so you can answer honestly if the panel asks. The series' guide to using AI for grant writing without funders rejecting it goes further on how panels react.
Whatever the funder says, a named person signs the declarations and is accountable for the whole application, including any sentence an assistant wrote. That is the practical meaning of "checked by hand".
Where an agent should stop: portals, declarations and signatures
If you do use an agent, give it the preparation jobs and nothing that commits the firm. ChatGPT Work checks in when judgement or approval matters, and OpenAI lists confirmations for high-impact actions among its agent safeguards, so never tell it to skip them. Anthropic's safety page for Claude in Chrome advises strongly against using it on financial accounts, legal documents or contracts, and sensitive work accounts, and a funding portal login with your company details is exactly that kind of account. Keep portal passwords out of any agent's reach, let it assemble the attachment pack, and have a person log in, read each declaration and press submit.
Plan the submission itself as a human task with a time buffer. Portals fail on deadline day, files exceed size limits, and a mandatory field appears that the guidance never mentioned. A bid submitted two days early with a confirmation email beats a perfect one submitted at 4:59 pm.
Testing a dedicated grant tool against your own call
Several dedicated grant-writing tools rank for this search, and they may suit you, particularly if you apply often. Test any of them the same way you would test a Project. Upload the same call and fact sheet, and ask for the requirements matrix. Check it against your own, and note which items it missed or invented. Ask it to draft one answer and see whether it cites the file behind each claim. Find out where your data is stored, whether it is used to train models, and whether you can export your drafts if you cancel. If a tool cannot show its sources for a claim, it has the same limits as the plain assistant, with a higher bill.
A time budget for a business bid, with and without an assistant
Set expectations honestly, because the saving is real but modest. The hours below are illustrative, for a bid with a 3,000-word narrative and a budget workbook.
| Task | Manual hours | With a loaded assistant | Why |
|---|---|---|---|
| Fact sheet and evidence gathering | 5 | 5 | The assistant cannot invent your evidence |
| Reading the call and building the requirements matrix | 3 | 1 | It extracts, you verify against the original |
| Eligibility test | 2 | 2 | A human decision |
| Budget workbook and match funding | 6 | 6 | Spreadsheet work, done by a person |
| First draft of the narrative | 12 | 5 | Drafted from the fact sheet |
| Reconciling figures and claims | 2 | 4 | Longer, because every number is traced to a file |
| Edit, second reader and revisions | 5 | 5 | Unchanged |
| Attachments, portal entry and submission | 3 | 3 | Unchanged |
| Total | 38 | 31 | About 18 per cent saved |
Seven hours back on a 38-hour bid is worth having, and it comes from reading and drafting. A vendor that promises a 90 per cent saving is describing the first draft only. The checking time goes up, which is the price of using a tool that writes fluently whether or not it is right, and it is the part of the budget you should protect when a deadline gets tight.
A final review before anyone submits
- Eligibility table completed, with evidence attached for every row.
- Deadline, time zone and submission route confirmed from the current call.
- Every figure in the narrative reconciled to the budget file.
- Award amount calculated by a person, applying the percentage, caps and eligible-cost rules.
- Every claim about the business traced to the fact sheet or an attached document.
- Supplier quotes in date, and any pre-award spend disclosed.
- Word limits met without cutting required content.
- Letters of support received and signed by the right people.
- A statement about AI use added if the funder asks for one.
- A second person has read the whole application cold.
- The signatory has read every declaration.
- Confirmation email saved after submission.
The assistant can draft the checklist, but it cannot tick a single box for you. That is the line between a useful drafting tool and an application that survives a panel. If you write bids on behalf of others, the companion tutorial on grant writer rates when AI drafts faster shows how to price the checking work that AI makes more important.
Grant writing AI questions from small firms
Can ChatGPT, Claude or Gemini write a whole grant application?
They can draft most narrative sections from your notes, but they cannot confirm that you are eligible, that your figures are right or that the funder's rules have not changed. Treat the output as a first draft that you check line by line against the call guidance and your own records, then have a second person read it before anyone signs.
What happens to my custom GPT?
OpenAI is retiring custom GPTs: they are scheduled to stop running on 11 December 2026, with approved deferrals for some Enterprise workspaces to 11 February 2027, and OpenAI offers a migration to plugins where your account shows it. Otherwise move the instructions and files into a ChatGPT Project, a Claude Project or a Gemini Gem (becoming a skill from November 2026), and copy your instructions out of the GPT before it stops.
Do I need a paid plan to set this up?
Not to build it: Projects exist on free ChatGPT and Claude accounts, with limits such as five files in free ChatGPT Projects and five projects on free Claude. For real client or financial data, a business plan is safer, because ChatGPT Business and Claude Team do not train on business content by default, while consumer plans need a privacy setting switched off.
Should I tell the funder I used AI?
Read the call first. Funder policies on AI differ and are still changing, so look for any statement on AI use in the guidance. If the guidance is silent, keep a log of which sections AI helped with, so you can answer honestly if asked. The person who signs the declaration is responsible for the whole application either way.
Can an AI agent submit the application for me?
It should not. Submission involves declarations, a signatory and a funder portal login, and the platforms' own safeguards point the same way: OpenAI lists user confirmations for high-impact actions among its agent safeguards, and Anthropic strongly advises against using Claude in Chrome on financial accounts, legal documents or contracts, and work accounts with sensitive company data. Let the agent prepare files and have a person submit.
Further reads
- ChatGPT Prompts for Grant Writing: Needs to Outcomes, Fact-Checked — Prompts for the needs and outcomes sections, fact-checked.
- Best AI App for Grant Writing: What Each Does Well on a Funding Bid — How dedicated grant tools differ on a funding bid.
- Logic Model Template: Build One With AI From a Client's Programme — Build the outcomes framework many funders ask for.
- AI Grant Writing Software: What Each Costs a Freelance Grant Writer — Price Grantable, Instrumentl and a general assistant for a one-person practice, check client-data handling, and test each on one past bid.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI's custom GPT retirement FAQ, its Projects help article and ChatGPT Work pages, Anthropic's help articles on Claude Projects and on using Claude in Chrome safely, Google's help pages on the move from Gems to skills and on sharing a Gem, and vendor plan pages checked on 4 November 2026.