TL;DR
Feed your AI a clean budget table (this year vs. last, by category, with absolute and percentage change) plus your top priorities and cuts, then use the prompt below to get a 300-400 word narrative. Three paragraphs: totals and year-over-year change, top investments with rationale, cuts and one real risk. Use Claude for Excel or ChatGPT data analysis to compute the variances first, then have the model write the narrative. Verify every number before it leaves your screen.
The task
You finalized next year’s budget. The spreadsheet is closed. Now leadership wants a narrative: what changed, why, what increased, what got cut, and what risks you took on. The narrative is short (300-400 words), but it is the only artifact most of the executive team will read. Get it wrong and the budget gets unbundled in committee. Get it right and the spreadsheet ships intact.
This is a writing task with a data-integrity trap underneath it. The narrative has to be both persuasive and factually airtight, and AI is good at exactly one of those two things by default. The job here is to make it do both.
Which AI to use for each step
Budget narrative work splits into two jobs: computing the variances and writing the story. As of June 2026, different tools win each.
| Step | Best tool | Why |
|---|---|---|
| Compute variances from a live workbook | Claude for Excel (Anthropic add-in) | Cell-level citations, cross-tab dependency tracing, and a “Claude Log” tab that records every action. Available to all Claude Pro users ($20/mo) since Jan 24, 2026; went GA on May 7, 2026. |
| Analyze an exported CSV/XLSX and chart it | ChatGPT data analysis (GPT-5.5) | Runs Python (pandas) on uploaded files. Handles CSV/XLSX up to ~50MB, though files above 30MB can parse slowly or partially. Plus ($20/mo) and up. |
| Native Google Sheets formulas (QUERY, ARRAYFORMULA) | Gemini in Sheets (Gemini 3.1 Pro) | Strongest on Sheets-native functions; included with Google AI Pro ($19.99/mo). |
| Write the narrative itself | Claude Opus 4.7 or GPT-5.5 | Both produce confident, executive-register prose. Claude tends to hedge less; GPT-5.5 follows tight word limits more reliably. |
A clean two-step flow: have the spreadsheet tool produce a verified variance table, paste that table into the writing prompt, and never let the writing model recompute a number.
What to feed the AI
- Budget table: this year vs. last, by category, with absolute and percentage change
- Top 3 strategic priorities driving spend
- 1-2 cuts with the reason behind each
- The audience (CFO, CEO, board)
- Tone preference (confident, cautious, candid)
- Anything that is not up for re-negotiation, so the model does not invite it
Copy-ready prompt
Write a budget narrative.
Audience: [CFO / CEO / board]
Tone: [confident / cautious / candid]
Strategic priorities driving spend: [list]
Cuts and reasons: [list]
Not up for re-negotiation: [list]
Budget data (this year vs last):
"""
[paste verified variance table]
"""
Return:
1. A 300-400 word narrative with: total $ + YoY change, top 3 areas of
increased investment with rationale, top 2 cuts with rationale, and 1
risk if the budget assumption is wrong.
2. A "what stakeholders will challenge" block: 3 predicted objections, each
with the prepared response and the specific person likely to raise it.
3. A "summary line": one sentence that captures the whole budget for
executive consumption.
4. A redline on any number you cannot trace to the source table: flag [VERIFY].
Tone constraint: assume the reader supports the strategy. Confidence, not
justification. Do not state a number without its rationale.For board-facing narratives, append: Add a 100-word "why now" framing that anchors this budget to the multi-year plan.
Where AI helps and where it does not
AI is strong at structuring a narrative around variances, naming year-over-year changes, and framing trade-offs in business language. It is weak at internal politics. If a cut hits a specific exec’s team, the model will not soften the language correctly, because it does not know who reads it. Read every draft against your stakeholder map. The narrative is part finance, part diplomacy, and only the finance half is automatable.
The other failure mode is invented arithmetic. A general-purpose chat model will happily fabricate a YoY percentage that “feels right.” This is why the workflow above computes variances in a tool that cites cells (Claude for Excel) or runs real code (ChatGPT data analysis), and only then hands a frozen table to the writer.
Recommended output structure
Three paragraphs:
- Macro picture — total spend and YoY change, framed against the strategy.
- Investments — top 3 increases, each with a one-line rationale.
- Cuts and risk — top 2 cuts (named, not buried) plus the single biggest risk if a core assumption is wrong.
Put the summary line at the top. Keep the “predicted objections” block in your prep notes, not in the narrative itself.
How to check the output is usable
- Every number traces to the source table. No invented totals, no fabricated percentages.
- Cuts are not hidden in passive voice. They are named, explained, and owned.
- The “risk if the budget is wrong” is a specific assumption that could break, not boilerplate.
- Predicted objections name the specific person likely to raise them.
- The summary line is repeatable. Your CFO can quote it in committee without edits.
Common mistakes
- Numbers without rationale. Execs reverse-engineer the rationale themselves, badly.
- Hiding cuts. They get found anyway; owning them beats being caught.
- Letting the model invent percentages. Compute YoY in Excel or with code, then paste the frozen figures.
- Defensive tone. Assume the reader supports the strategy.
- Missing the strategic frame. Without it, the budget reads as accounting, not a plan.
FAQ
How long should it be? 300-400 words is the sweet spot. Shorter reads as thin; longer goes unread. Board decks sometimes want an extra 100-word “why now” paragraph, but keep the core to one page.
Which model writes the cleanest narrative? Claude Opus 4.7 and GPT-5.5 are both strong. Opus 4.7 hedges less and reads more confident; GPT-5.5 holds a tight word count more reliably. Try the same prompt in both and keep the better draft.
Can the AI compute the variances too, so I skip Excel? It can, but verify it. For a live workbook, Claude for Excel cites the cells it used; for an exported file, ChatGPT runs pandas and you can ask it to show the code. Never ship a number the tool could not trace.
Should I include charts? Reference them, do not embed them in the narrative. Keep the prose a narrative and let the deck carry the visuals.
What if leadership pushes back on a cut? Pull the prepared response from your “predicted objections” prep notes. That block exists precisely for the committee meeting, not the document.
Related
- Financial trend analysis — multi-year trend writing
- KPI weekly report AI — same structure for weekly updates
- Manager update email AI — when the narrative is an internal update
- Financial report summary AI — same compression pattern for 10-Ks
- Business data analysis AI — adjacent business analysis
- Roadmap planning AI — when budget supports a roadmap
- Career Pivot Narrative Prompts for a Convincing Story
- AI Pricing Strategy Memo: Tier Design, Anchor, Discount Floor
- AI Monthly Investor Update Without Sounding Defensive