Grant Budget Narrative: Draft It With AI and Reconcile Every Figure

Write the budget narrative from the final spreadsheet, not the other way round. Give AI each budget line with its quantity, rate and total, ask for a two-to-three-sentence justification per line using only those figures, then extract every number from the draft and match it to the sheet. Submit only when each figure ties.

The reconciliation is the part that protects you. Reviewers notice when a narrative says $10,376 and the budget says $10,736, and a mismatch makes them doubt the rest. AI is useful for the prose and risky with arithmetic, so the routine below keeps it away from the second.

Follow me on Instagram@sagnikteaches

Why most narrative errors are transposition errors

Budgets change late. A salary is revised, a rate drops, an evaluator is added. The spreadsheet gets updated and the narrative, written three days earlier, does not. Or a figure is retyped and two digits swap. Neither error is about money knowledge; both are about two documents drifting apart. Humans are poor at catching this by rereading, because the sentence reads fine. A systematic check catches what a reread misses.

Connect on LinkedInSagnik Bhattacharya

AI can make the drift worse if it is allowed to compute, because it may round, combine or recalculate silently. It can make it better if you use it for what it does well: wording, structure, and extracting every number in a document into a list you can test mechanically. This tutorial uses it in both of those ways and bans it from the third.

Subscribe on YouTube@sagnikteaches

An illustrative budget to work from

Say a small charity applies for a 12-month youth literacy project. The funder's template wants a narrative under each budget category. Here is the final spreadsheet, with every total calculated in the sheet:

You can verify the arithmetic: 31,200 + 17,600 = 48,800; 22% of that is 10,736; adding travel, packs, evaluator and venue brings direct costs to 68,756; ten per cent is 6,875.60; the request is 75,631.60. These are made-up figures for teaching, and the rates and percentages stand in for whatever your funder allows.

Step 1: lock the spreadsheet before drafting

Do not start the narrative until the budget is final, or at least frozen for the draft. Save a dated copy named "budget-v-final-for-narrative" and treat any later change as a reason to rerun the reconciliation below. Each line needs a stable label (the same wording in the sheet and the narrative), a basis (quantity x rate, or percentage x base) and a total. If any line lacks a basis, fill it in now. A line you cannot explain to a reviewer will be hard to justify in prose.

Also gather the funder's rules: the template, word or character limits, allowed and disallowed costs, the treatment of indirect costs and any required match or in-kind contribution. Paste the key instructions, not your whole application, into the prompt later. Funder guidance is usually public; client confidential figures are not, so check the plan's terms before you paste them, as covered in how to use AI to write grant applications, step by step.

Step 2: draft line by line with a locked prompt

The prompt does three things: it supplies the lines, it fixes the figures, and it sets the form of the justification. A tested version:

You are drafting a grant budget narrative. Follow the funder's rules below.
Use ONLY the figures in the table. Do not calculate, round, combine or infer any number.
For each line write 2-3 sentences: what the cost is, why the project needs it, and how
the figure was built (quantity x rate or percentage x base), quoting the table exactly.
Do not add costs, claims or outcomes that are not in the project summary.
If a line has no stated need, write [NEED MISSING] and stop.
Funder rules: [paste: allowed costs, limits, indirect costs wording]
Project summary: [paste 100 words on activities and outcomes]
Budget table: [paste lines with basis and total]

An illustrative output for one line reads: "Programme coordinator ($31,200). A coordinator at 0.6 FTE of a $52,000 full-time salary will schedule sessions, manage school partnerships and supervise tutors across the 12 months." Check it against the sheet: 0.6 x 52,000 is 31,200. Fine. The phrase "manage school partnerships" must appear in your project summary; if it does not, it is an invention and you strike it. That is the second job of review: every activity named in a justification must exist in the project.

Use the same process for in-kind and match lines. Ask for wording that separates cash from in-kind, and keep the labels the funder uses. If the funder's form has a column for "source of funds", mirror it. The related tutorial on ChatGPT prompts for grant writing shows prompts for the sections that surround the budget.

Step 3: word the personnel and fringe lines carefully

Personnel lines are the ones reviewers read hardest, because they are usually the largest. Each needs a role, a percentage of time, the basis of the salary and what that person does on the project. Avoid vague claims such as "oversees all aspects"; say "supervises the three part-time tutors and reports monthly to the steering group". Fringe should state its rate and its base: "22% of the combined salaries of $48,800". If the base excludes a role, say so.

A before and after. First draft from AI: "Staff costs support delivery of the programme." After the fix: "The coordinator (0.6 FTE, $31,200) and tutor lead (0.4 FTE, $17,600) deliver and supervise 36 weekly sessions; fringe benefits are 22% of those salaries ($10,736)." The second version names roles, effort, quantities and a figure for each part. A reviewer can see how the total was built and can check it.

Step 4: indirect cost wording that holds up

Indirect costs are a common stumbling point, because funders differ and the wording must match their rules. Three situations cover most cases. If the funder sets a cap or a flat rate, state the rate and the base it applies to, and show the calculation. If your organisation has a negotiated rate agreement, name it, give its rate and its dates, and attach what the funder asks for. If the funder does not allow indirect costs, do not hide overhead in other lines. Put the legitimate, directly attributable costs in their own lines, such as a proportion of rent or a time-recorded share of finance staff, only where the funder permits it.

For our example the wording is: "Indirect costs are calculated at 10% of total direct costs ($68,756.00), giving $6,875.60, in line with the funder's published cap." The phrase "in line with the funder's published cap" is only true if you have read the cap in the guidance. Check it, and tell AI nothing about rates you have not confirmed. If you are unsure whether an item counts as direct or indirect, ask the funder's programme officer, and follow the answer. Do not rely on a model's view of what a funder allows, because funder rules are specific, they change, and the model may be recalling a different funder's.

Step 5: build a figures register and reconcile

This is the check that stops mismatches. Paste the finished narrative into a prompt that only extracts numbers:

List every number, percentage, quantity and currency amount in the text below,
one per row, in the order they appear. Columns: sequence, exact text as written,
the sentence it appears in. Do not calculate or correct anything.

Put the result in a Narrative tab of the budget workbook, with a column for the matching sheet cell. Match each row to its source with a lookup, or by eye for a short list, and mark Match or Mismatch. In our example, a sample result is a row "$10,376" with the sentence "fringe benefits are 22% of those salaries ($10,376)". The sheet says $10,736, so the row is a Mismatch. The error is a swapped pair of digits, a classic.

Then run three tests in the spreadsheet itself, with ordinary formulas:

1. Every narrative figure is in the sheet. A figure that appears in the narrative but nowhere in the budget is either an error or an invented number.
2. Every sheet line is in the narrative. A line with no justification will attract a reviewer's question. Count lines with no matching sentence.
3. The arithmetic holds. Recompute each total from its basis in cells beside it. For instance =0.6*52000 should equal the line total, and =SUM of the direct lines should equal the stated total direct costs.

Finish by reading the narrative aloud against the sheet, one line at a time. It takes ten minutes for a budget of this size, and it catches wording that is numerically right but misleading, such as "full-time" where the sheet says 0.6 FTE.

Three more lines drafted: travel, evaluator and supplies

Personnel gets the attention, but the smaller lines are where vague wording creeps in. Run the same locked prompt over the rest of the table and read each output against its basis. Illustrative outputs, with what a reviewer would fix:

• Travel ($760). Draft: "Travel to schools is needed for outreach." Fixed: "Twenty visits to partner schools at $38 per visit ($760) allow the coordinator to meet teachers, agree referrals and observe the first session in each school." The fix adds the quantity, the rate and the purpose, so a reviewer can see why 20 and not 40.
• External evaluator ($5,100). Draft: "An evaluator will assess the project." Fixed: "An independent evaluator will work for 60 hours at $85 ($5,100) to design a pre- and post-programme reading check, analyse results and write the final evaluation report." If your funder requires a competitive process for contracts above a threshold, say how the evaluator was or will be selected. Do not claim a procurement process that has not happened.
• Learner packs ($1,740). Draft: "Packs provide resources for learners." Fixed: "120 learner packs at $14.50 ($1,740), each containing a reading book, a journal and stationery, ensure every enrolled child has materials to keep." The contents must match your quote or supplier price, so check them.

The pattern is the same each time: what the cost is, the quantity and the rate, and why it serves the outcome. If the model produces a paragraph without a quantity, rerun with an instruction to quote the basis, or write the sentence yourself and use AI only for tightening.

A reconciliation sheet you can copy

The register tab is simple enough to build in ten minutes. An illustrative layout, with three of its rows filled in:

Add a final cell that counts Mismatch rows with =COUNTIF(F2:F60,"Mismatch*") and refuse to submit until it reads zero. Row 3 shows why the no-rounding rule exists: the request was reduced to a round figure in the summary paragraph and now disagrees with the sheet by $31.60. A funder's portal may reject a form in which the narrative total and the entered total differ, so a $31.60 gap can cost more than the $31.60.

Multi-year and match-funded budgets

Longer projects add two traps. First, annual figures rise with salary uplifts or inflation assumptions, and narratives must say what assumption you used, for example "a 3% annual uplift on salaries from year two", only if your budget applies one. Second, match funding and in-kind contributions are easy to double count. Keep separate columns for the funder's share, your cash match and in-kind value, and have AI write the narrative per column from the same table. In the register, add the year and the funding source so a $12,000 figure in year two is not matched against the year-one line.

Say a three-year version of our project grows each year. Year one is $75,631.60, year two applies a 3% salary uplift to the $48,800 salaries giving $50,264.00, and fringe at 22% becomes $11,058.08. Run those through the sheet, never the model, then paste the sheet's results as the table. The narrative only describes them. A good habit is to ask the model afterwards to list every assumption it saw in the text, then confirm each is in your budget notes.

Pre-submission checklist

1. Spreadsheet final and saved with a date; narrative drafted from that version.
2. Funder's template, headings and limits followed; word counts checked.
3. Every line has quantity, rate or percentage, and purpose.
4. Indirect costs worded as the funder's guidance says, with base and calculation shown.
5. Figures register shows zero mismatches; totals recomputed in cells.
6. No activity, outcome or commitment in the narrative that is missing from the project plan.
7. Someone other than the drafter has read the narrative against the sheet.
8. Version saved with the submission, so you can answer later queries from the same figures.

Realistic ways this goes wrong

• A "helpful" rounding. The model writes "about $75,600" for the request. A reviewer reads that as a different number. The prompt bans rounding for this reason; the extraction step catches the cases that slip through.
• A drifting label. The sheet says "External evaluator" and the narrative says "independent assessor" in one place and "evaluation consultant" in another. Reviewers cross-check names. Fix by pasting the labels as fixed text.
• An invented justification. AI adds "to ensure sustainability beyond the grant period", which nobody planned. It sounds good and creates a commitment. Strike it unless the project really includes it.
• A late change. The evaluator's hours go from 60 to 50 on submission day. Compare the sheet with the dated "for narrative" copy (version history shows what moved), rerun the register, and allow time for it in the schedule.

Fitting the funder's format and limits

Funders vary in how they ask for a narrative: a section per category, a single page, a table with a justification column, or a character-limited box. Paste the format into the prompt and tell the model to respect it. For a limited box, ask for the shortest version first, with the build of each figure, and expand only if you have room. If the funder gives a scoring rubric, ask AI to check the draft against the criteria and list gaps, but decide for yourself whether a criterion is met.

Keep the logic of the budget aligned with the rest of the bid. The activities in your narrative should match the work plan, and the outcomes the money buys should appear in the logic model. Building a logic model with AI from a client's programme helps there, and tracking real spending against these lines later is covered in tracking grant expenses by budget line with AI.

Time, cost and what to tell the client

For a budget with a dozen lines, the routine costs about 20 minutes to draft and 30 to reconcile once you have the register tab set up, against perhaps two hours writing by hand. Those are rough planning figures. A larger multi-year budget with sub-awards takes proportionally longer and deserves a second human reader. Where you are a freelance writer, tell the client in writing that you use AI for drafting and that every figure is reconciled by you. The register tab is also something you can show the client: it turns "I checked it" into a list they can see. For pricing implications, see what to charge when AI drafts faster.

If you would like to set this up for your own bids, with templates that match your usual funders, that is what my AI implementation consultation covers, at the contact page.

Budget narrative questions grant writers ask

Should AI ever see the budget figures?

Yes, as inputs it must not change. Paste the final spreadsheet lines and tell the model to use only those figures, never to calculate or round. Then reconcile the finished narrative back to the spreadsheet yourself. The model drafts the words; the sheet owns the numbers.

How do I word indirect costs?

Follow the funder's guidance exactly: say whether you use a negotiated rate, a funder-set cap or a fixed percentage, name the base it applies to, and show the calculation. If the funder disallows indirect costs, show the overhead as direct lines it permits, and never rename costs to slip them through.

What if the funder gives its own template?

Use the funder's template, headings and word limits as the structure and let AI fill your content into it. A narrative in your own format risks being scored down or returned unread. Paste the template's instructions into the prompt so the draft follows them.

How detailed should each line be?

Enough that a reviewer can see what the cost is, why the project needs it, and how the number was built: quantity times rate, or percentage of effort times salary. One to three sentences per line is usual unless the funder sets other limits.

Can I put the funder's guidance into a chatbot?

Funder guidance is normally public, so pasting it is fine. Your client's confidential figures need a business plan with no training on content by default, and permission from the client where your agreement requires it.

Further reads

• Grant Report Template: Draft Funder Reports With AI From Project Data — The reporting side once the award is made.
• Can You Use AI for Grant Writing Without Funders Rejecting It? — Whether funders mind AI-assisted bids.
• Can AI Write a Sports Club Grant Application? What Funders Expect — What funders expect in a small club application.
• Grant Writing AI Agent: What It Can Draft in a Business Grant Bid — What an AI agent can and cannot draft in a bid.
• Best AI App for Grant Writing: What Each Does Well on a Funding Bid — Which AI app suits grant drafting best.
• AI Grant Writing Tools Compared for Small Non-Profits — Tools compared for small non-profits.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general funder guidance conventions; always check the specific funder's guidelines and budget template. All figures in the examples are illustrative.