TL;DR
Revenue is up 14%, margin is down 2 points, and the board pre-read needs one paragraph explaining both. AI is excellent at turning a driver tree you already built into a tight executive narrative, and bad at building the tree itself. Feed it the deltas at each node (not the raw GL), force a “structural vs. one-time” split, and cap the output at two drivers and 120 words. The prompt below does this in one pass.
The task
Revenue is up 14%. Margin is down 2 points. You need to write, in one paragraph, which two underlying drivers actually explain both — for a board pre-read where nobody will open the appendix.
When this is the right job for AI
- You already have a decomposition tree (volume × price × mix, or units × ARPU, etc.).
- You can hand AI the deltas at each node — not the raw general ledger.
- You want narrative, not modeling. AI is good at “explain this tree”; bad at “build this tree.”
If you have not built the tree yet, do that first. Models will trace dependencies inside a spreadsheet you give them, but they invent aggregations when asked to construct the decomposition from a trial balance.
What to feed the AI
- Top-line metric, base and current period (e.g. “Revenue: $38M → $43.4M”)
- The first two levels of the driver tree as a flat list:
- Volume: 1.2M → 1.32M (+10%)
- Price: $31.6 → $32.9 (+4%)
- Mix: B-tier share 18% → 24%
- Margin tree similarly
- Audience and a hard length constraint
Copy-ready prompt
You are writing the "drivers" paragraph for a board pre-read.
Revenue: $38M → $43.4M (+14%)
Drivers (level 1):
- Volume +10% (units 1.20M → 1.32M)
- Price +4% (ARPU $31.6 → $32.9)
- Mix: B-tier share 18% → 24% (higher-priced)
- One-time: enterprise renewal pulled forward = +$0.8M
Margin: 56% → 54% (-2 pts)
Drivers (level 1):
- COGS/unit +6% (input costs)
- Mix shift toward enterprise = -1 pt gross margin
- FX (CNY/USD) -0.5 pt
Write:
1. The two drivers (one for revenue, one for margin) that actually carry the story.
2. One sentence on which is structural vs. one-time.
3. One implication for next quarter.
Audience: board members, non-finance. 120 words. No bullets — one paragraph.Sample output structure
Revenue growth is real but two-thirds volume-driven, not price — sustaining 14% next quarter requires we keep the volume engine running, since pricing only carries 4 pts. Margin compression is the more structural story: COGS/unit is up 6% on input costs and is unlikely to reverse this quarter. The one-time pull-forward ($0.8M) flattered Q3 revenue by ~2 pts; back it out and underlying growth was closer to 12%. Implication: Q4 plan should assume 12% growth and a 54% margin floor unless we successfully renegotiate two top supplier contracts.
Which model to use (as of June 2026)
Any frontier chat model writes this paragraph well — the differences show up when you skip the manual flat-list step and want the model to read the workbook itself.
| Model | Plan | Reads your .xlsx/.csv | Best for |
|---|---|---|---|
| GPT-5.5 | ChatGPT Plus $20 | Yes — ChatGPT for Excel/Sheets went GA May 5, 2026; uploads up to ~50MB | Tracing “what drives this number” across tabs, then writing the narrative |
| Claude Opus 4.7 | Claude Pro $20 | Yes — Claude in Excel does cross-tab dependency tracing | Disciplined, source-anchored finance writing; tops the Finance Agent benchmark |
| Gemini 3.1 Pro | Google AI Pro $19.99 | Yes — 1M-token context | Pasting a long appendix and asking for the two-driver summary |
A practical default: build the tree in your model, paste the level-1 deltas into the prompt above (this keeps the model from hallucinating math), and let it write. If you want the model to pull the deltas itself, ChatGPT for Excel and Claude in Excel both trace assumptions through to outputs — but always spot-check two of the numbers it reports against your sheet before it goes in the pre-read.
How to refine
- AI gives you a list of five drivers — push back: “pick the one driver that explains 80% of the move.”
- Output too quantitative — add: “the audience does not have the deck open; do not cite percentages they cannot see.”
- AI conflates one-time and structural — add a hard rule: “label each driver
structuralorone-timein your head and tell me which is which.”
Common mistakes
- Feeding AI the raw GL trial balance and asking it to build the tree. It will hallucinate aggregations.
- Asking for eight drivers. The audience will remember zero. Two drivers, one paragraph.
- Skipping the “structural vs. one-time” split — this is the single most useful framing for execs.
- Letting AI invent benchmarks (“industry average is 6%”). Strip those unless you have a real source.
FAQ
- What if my driver tree is incomplete? Tell AI exactly that: “These are the only levers I have measured. Do not infer hidden drivers.” It will stay disciplined.
- Can I have AI suggest drivers I haven’t measured? Yes — separate prompt, separate output. Do not mix exploratory and reportable.
- How does this differ from variance analysis? Drivers explain “why the metric moved at all”; variance explains “why we missed plan.” Different audiences. See AI variance analysis.
- Can I just upload the spreadsheet instead of typing the deltas? Yes, with ChatGPT for Excel (GPT-5.5) or Claude in Excel, both of which trace calculations across tabs. But verify the reported deltas against the sheet — the board paragraph is too high-stakes for an unchecked AI read.
- Should I use a finance-specific tool instead? Dedicated commentary writers (e.g. Fathom) keep an auditable math trail and are worth it if you publish recurring management reports. For a one-off board paragraph, the general chat models above are faster.
Related
- AI variance analysis
- AI financial trend analysis
- AI sensitivity analysis
- AI executive summary
- Monthly Business Review Prompts for MBR Decks
External reference: OpenAI — ChatGPT for Excel and Google Sheets
Tags: #AI writing #Finance #Business analysis #KPI #Executive update