TL;DR
You already built the sensitivity table; AI’s only job is the narrative. Paste the grid plus the base case, demand one headline + three scenarios + one watch-item + one recommendation, and tell it not to interpolate. For internal numbers, use a no-training tier (ChatGPT Business, Claude for Work, or any consumer chat with training off). For the heavy lifting on a dense grid, Claude Opus 4.7 (1M-token context) is the safest pick as of June 2026.
The task
You ran a sensitivity table — price ±10%, volume ±10%, COGS ±5% — and need to turn it into a one-page narrative your CEO can actually read. The numbers are right; the story is missing. This is a writing-and-framing job, not a modeling job, which is exactly where a current chat model earns its keep.
When this is the right job for AI
- The math is already done in Excel / Sheets / Looker — you only need narrative.
- You can hand AI a clean table (no raw transactions, no PII).
- The audience is non-finance and needs “what to do”, not the matrix itself.
If you are still building the model, do that first. AI cannot recover a broken sensitivity table, and it will happily narrate wrong numbers with total confidence.
What to feed the AI
- Base-case revenue, gross margin, EBIT
- The sensitivity table itself (rows = price scenarios, cols = volume scenarios — or vice versa)
- The 1-2 assumptions you suspect are most fragile (e.g. “Q4 churn assumed at 4%”)
- Audience: “CFO + 3 non-finance VPs”
Copy-ready prompt
You are writing a one-page sensitivity-analysis narrative for non-finance executives.
Base case: revenue $42M, GM 58%, EBIT $6.1M.
Most fragile assumptions: [paste 2]
Sensitivity table:
[paste table - price scenarios x volume scenarios -> EBIT]
Rules: use only the cell values I provide. Do not interpolate or invent
scenarios that are not in the table.
Write:
1. One-sentence headline: which lever moves EBIT the most.
2. Three bullets: best case, worst case, most likely case - each with the lever that drives it.
3. One "watch this" - the assumption I should pressure-test before locking the plan.
4. One recommendation in plain English: which lever to defend in the plan.
Tone: direct. No "it depends." 300 words max.Sample output structure
Headline: EBIT swings ±$2.4M on price, only ±$0.9M on volume — pricing is the lever to defend.
- Best case ($46M rev / GM 60%): if we hold list price and volume comes in at +5%, EBIT lands at $7.8M.
- Worst case ($38M rev / GM 55%): a 5% price concession plus flat volume drops EBIT to $3.7M.
- Most likely ($42M rev / GM 58%): base case ships if we don’t discount.
Watch this: the Q4 churn = 4% assumption. At 6% churn the worst case becomes the most likely.
Recommendation: defend list price. Volume incentives are cheaper than discounts if growth slips.
Which model to use (June 2026)
All three flagships read a pasted 2D grid fine. The differences show up on big multi-tab workbooks, dense filings, and “don’t lose a single cell” reliability.
| Model | Context window | Best for this task | Notes |
|---|---|---|---|
| Claude Opus 4.7 | 1M tokens | Dense grids, multi-quarter context, “no interpolation” discipline | Benchmark leader for long-document financial analysis; strongest at multi-needle retrieval, so it loses fewer cells |
| GPT-5.5 (Thinking) | ~320 pages in-app (Plus); 1M API | Math-heavy framing, tight executive prose | Default ChatGPT model since ~Apr 23, 2026; full 1M in-app window only on the $200 Pro plan |
| Gemini 3.1 Pro | 1M tokens (Google AI Pro) | Pasting straight from Google Sheets | Lives next to your Sheets data; solid on numeric reasoning |
For a single 2D table that fits on one screen, any of them is fine — pick the one your data already lives in. For a 200-row scenario dump or a workbook with several tabs, Claude Opus 4.7’s 1M-token window and long-context retrieval make it the safer choice. See the full lineup in ChatGPT vs Claude vs Gemini.
Keep your numbers off the training set
For anything with real internal figures, use a tier where your inputs are not used to train the model:
- ChatGPT Business — $25/seat/month, or $20/seat/month billed annually (2-seat minimum). Business data is excluded from training by default. (OpenAI pricing)
- Claude for Work / Team / Enterprise — commercial data is never used for training. (Anthropic privacy)
- Consumer ChatGPT Plus / Claude Pro ($20/mo) — you must turn training off in settings; it is not off by default. Treat these as personal tiers, not approved channels for company numbers.
Either way, strip customer names and absolute dollar figures if your policy is strict — the ratios and deltas are what drive the narrative anyway.
How to refine the output
- Output too generic? Paste 1-2 prior-quarter narratives from the same company so AI matches voice.
- AI hedging (“could go either way”)? Replace prompt line 1 with: “Pick the single most important lever; do not list two.”
- Wrong audience tone? Add: “The CFO already knows the math — do not re-explain sensitivity. Focus on the call.”
Common mistakes
- Pasting raw transactional data instead of the already-built sensitivity grid.
- Letting AI invent numbers between the cells you provided — always tell it “do not interpolate, use only my values.”
- Asking for “what if customers churn at 10%?” without giving the row for that scenario. AI will hallucinate one.
- Skipping the recommendation. A sensitivity report without a call is data, not analysis.
FAQ
- Can I paste my sensitivity table into ChatGPT? Only on a no-training tier for real numbers. ChatGPT Business ($25/seat/mo, or $20 annually) and Claude for Work exclude your data from training by default; on consumer ChatGPT Plus or Claude Pro ($20/mo as of June 2026) you must switch training off yourself first. Strip customer names and dollar absolutes if your policy is strict — ratios are usually fine.
- What if my sensitivity is 3-dimensional (price × volume × COGS)? Collapse to two 2D tables (price × volume at base COGS, then price × COGS at base volume). Every model narrates 2D far more reliably than a 3D cube.
- Which model handles a big multi-tab workbook best? Claude Opus 4.7 — its 1M-token context and long-context retrieval mean it loses fewer cells than a model you have to feed in chunks. For a single screen-sized grid, any flagship is fine.
- Does this work for non-P&L sensitivities (e.g. churn × ARPU)? Yes — same template. Replace “EBIT” with the target metric and keep the “no interpolation” rule.
- How do I stop it inventing numbers? Put the rule first (“use only the cell values I provide; do not interpolate”), and give it the exact rows for any scenario you want named. If a scenario isn’t in your grid, it isn’t in the answer.
Related
- AI financial trend analysis
- AI variance analysis
- AI executive summary for management reporting
- AI business drivers breakdown
- Excel / Spreadsheet Analysis Prompts
Tags: #AI writing #Finance #Sensitivity analysis #Business analysis #Executive update