Most unit-economics spreadsheets are read once, by the analyst who built them, and never again. The board wants a story (what changed, why, and what to do), not a wall of cells. This workflow turns the numbers into a half-page narrative the CEO can read in 90 seconds and act on the same day.
TL;DR
- Don’t paste a screenshot. Upload the workbook so the model reads the actual rows, then make it cite the source cell for every headline number.
- For reading a multi-tab financial model, use Claude for Excel (Opus 4.7, cell-level citations). For computing payback or charting from a CSV, use ChatGPT data analysis (GPT-5.5 runs Python). For models that already live in Google Sheets, use Gemini 3.1 Pro in the side panel.
- The prompt below forces one headline, three bullets, a root-cause paragraph marked
[hypothesis]where it’s inferred, and a “verify” footer. - Anchor every metric to a 2026 benchmark: healthy B2B SaaS CAC payback is under 12 months (median ~6.8), and a healthy LTV:CAC sits between 3:1 and 5:1. Read payback next to gross margin or it means nothing.
The task
You have a unit-economics workbook with CAC by channel, contribution margin per cohort, LTV computed two ways, payback period, and a deck slot due tomorrow morning. You want the AI to read the numbers, find the one or two cohort shifts that actually matter this quarter, and write a narrative that says “paid social CAC went from 42 to 71, payback stretched from 7 to 14 months, the rest is noise”, not a recap of every cell.
When this is the right job for AI
AI is strong at reading a structured table, comparing periods, and producing a narrative in your tone. It is weak at deciding which metric you actually care about and at catching unit errors (mixing gross and net, mixing trial-converted and total signups). Hand it a clean sheet, name the metric definitions inline, and verify every headline number against the source row before sending. If the sheet is messy, fix the sheet first; the model will dutifully narrate garbage.
Pick the right tool for the file you have
The three assistants are not interchangeable here. The split (as of June 2026) is between reading a workbook and running code on it:
| Tool | Model | Reads the actual file? | Best for | Limits |
|---|---|---|---|---|
| Claude for Excel | Opus 4.7 / Sonnet 4.6 | Yes, with cell-level citations and cross-tab dependency tracing | Auditing a multi-tab model, tracing a number to its source cell, narrating a clean summary tab | 30 MB/file, 20 files per chat, 1M-token context |
| ChatGPT data analysis | GPT-5.5 | Yes, and runs Python in a sandbox | Computing payback from raw inputs, charting a cohort curve, cleaning a messy CSV | CSV/XLSX up to 512 MB per file |
| Gemini in Sheets | Gemini 3.1 Pro | Yes, native side panel | Models that already live in Google Sheets, with Workspace context | 1M-token context |
Rule of thumb: if the workbook is the source of truth and you must not re-derive anything, use Claude for Excel so it cites the cell. If you need a number computed (payback from a cash-flow row, a fitted retention curve), use ChatGPT’s data analysis so the Python is visible and checkable. Don’t ask a model to eyeball arithmetic it could run.
What to feed the AI
- The unit-economics table, at minimum CAC, gross-margin-adjusted LTV, payback in months, by channel and by cohort
- Two comparison points: last quarter, and the same quarter last year
- Your metric definitions, e.g., “LTV here is 24-month, gross-margin-weighted, churn-adjusted”
- The audience (board, exec team, or finance review); each wants a different lens
- The decision the narrative should support, e.g., “should we cut paid social spend”
- Any known one-offs: a holiday promo, a pricing test, a campaign you paused mid-quarter
Copy-ready prompt
You are reading a unit-economics workbook for a SaaS / consumer business.
Write a half-page narrative for the audience and decision below.
Audience: [board / exec team / finance review]
Decision the narrative should support: [one sentence]
Metric definitions (use exactly these, do not redefine):
CAC: [definition]
LTV: [definition, including window and margin treatment]
Payback: [definition]
Known one-offs this quarter: [list, or "none"]
Data (current quarter / last quarter / same quarter last year):
"""
[paste the table — keep column headers]
"""
Return:
1. Headline (one sentence) — the single most important shift this quarter.
2. Three bullets: what changed, by how much, and in which channel or cohort.
Use the metric definitions above. Do not invent any number not in the table.
3. One paragraph on root cause — only what the data supports.
Mark anything you are inferring as [hypothesis].
4. One paragraph on the decision: what the data suggests we should do,
and what we would need to see to reverse the call.
5. A short "verify" footer listing the 3-5 numbers a reader would want to
double-check in the source before quoting.
Tone: direct, plain-English, no MBA hedging. Numbers in absolute terms first,
percent change second. Do not paraphrase a 12% move as "meaningful."For the follow-up pass: “Now rewrite the headline three ways: one for the board, one for the exec team, one for the finance committee. Same numbers, different lens.”
Sample output structure
Headline: Paid social payback stretched from 7 to 14 months; everything else held.
- CAC on paid social: $42 → $71 (+69%), Q4 vs Q3, all cohorts.
- LTV on paid social: $620 → $585 (-6%), driven by a 9% drop in month-3 retention.
- Payback: 7.2 → 13.8 months. Organic and partner channels unchanged.
Root cause: CPM inflation on the two largest creative sets we have not refreshed since Q2 [hypothesis — confirm with the ads team]. LTV drop is too small to fully explain payback; the issue is CAC.
Decision: Cut paid social spend to a 30% floor pending creative refresh. Reverse if 4-week rolling CAC drops below $55 with retention flat.
Benchmark every number before it ships
A narrative that says “payback is 14 months” is useless until the reader knows whether that is good. Anchor each headline metric to a current benchmark so the board reads a verdict, not a value. Healthy ranges as of June 2026 (B2B SaaS):
| Metric | Healthy | Watch | Source of truth |
|---|---|---|---|
| CAC payback | Under 12 months (median ~6.8) | 12-18 months | Your own blended, fully loaded CAC |
| LTV:CAC | 3:1 to 5:1 (median ~3.2:1) | Below 3:1 = overspend; above 5:1 = under-investing in growth | Gross-margin-adjusted LTV only |
| Gross margin | 70-85% | Below 70% reprices every other metric | Your P&L, not the model |
| Net revenue retention | Above 100% | Below 100% means the cohort is leaking | Cohort retention table |
Two notes that catch people out. First, B2C and enterprise read differently: consumer payback under ~6 months is normal because CAC is low and activation is fast, while enterprise can run 15-18 months and still be healthy if gross margin clears 70% and NRR is above 100%. Second, an LTV:CAC above 5:1 is not a victory lap; it usually means you are leaving growth on the table by under-spending. Have the model write the ratio and the read, never the bare number. The SaaS Metrics 2.0 essay by David Skok is still the standard reference for why these ratios move together.
How to refine
- If the narrative recaps every cell: “Cut to the one shift that moves the decision. Everything else is a footnote.”
- If it invents a number: “Use only numbers in the table. If you want to cite a derived metric, show the arithmetic.”
- If it hedges every sentence: “Strip ‘may,’ ‘could,’ ‘potentially.’ Either the data supports it or mark [hypothesis].”
- If root cause is hand-wavy: “Name the specific cohort or channel. ‘Engagement dropped’ is not a root cause; ‘month-3 retention on the Sept paid-social cohort dropped from 64% to 55%’ is.”
Common mistakes
- Mixing gross and net CAC across periods without flagging the change
- Letting AI compute LTV from raw inputs rather than reading your sheet’s definition
- Hiding the one big shift inside a paragraph of small ones
- Forwarding to leadership without the verify footer: bad numbers ruin credibility for a quarter
- Treating a 4-week blip as a trend; require two consecutive periods to call a shift
- Quoting a payback or LTV:CAC number with no gross margin next to it
FAQ
- Can I trust the LTV the AI computes for a board decision?: No. Have it read both methods (cohort LTV and predictive LTV), write both values, and flag the gap as a range. Decisions run on the lower bound, not on a single number the model produced.
- CAC went up: is it a channel problem or a market problem?: Feed the model that channel’s CAC trend alongside every other channel’s trend over the same window. Only this one rising means a channel issue (creative fatigue, bids, targeting). All of them rising means a market issue (seasonality, a new competitor, ad-auction inflation).
- What payback period counts as healthy?: Under 12 months for B2B SaaS (the median across thousands of tracked companies is around 6.8), under ~6 for consumer, up to 18 for enterprise, but only if gross margin clears 70%. Always pair “payback X months” with “gross margin Y%” in the narrative; payback alone is meaningless.
- My unit economics don’t match the P&L. Which wins?: The P&L is the real ledger; unit economics is a model, so the P&L wins. If the narrative diverges sharply from P&L figures, recheck the cost-allocation logic in the unit-economics sheet first. Don’t bend the P&L to the model.
- Which tool should I actually use?: Claude for Excel to audit and narrate an existing workbook (it cites cells), ChatGPT data analysis to compute payback or chart a curve from raw data, Gemini in Sheets if the model already lives in Google Sheets. See the table above.
Related
- Board deck narrative AI: pair this with deck building
- Financial report summary AI: adjacent 10-K summarisation workflow
- Business data analysis AI: broader business-data narrative
- Pricing experiment AI: pricing test design
- AB test summary AI: translate experiment output to narrative
- Chart table explanation AI: make the underlying table digestible