Most finance teams do not have a modeling problem. They have a first-draft problem. Someone needs a rough unit-economics view before Friday's meeting, and the alternative is three hours in Excel before anyone writes a single formula. That gap — between a rough draft and an auditable model — is exactly where LLM prompts earn their keep, and exactly where they quietly destroy value if you confuse the two.
This roundup compares two workflows honestly. Prompts win when you need speed, framing, and a second pair of eyes. Excel-by-hand wins when you need a final model that survives an auditor, a board, or a due-diligence data room. Below are 16 prompts that cover the first workflow — with the mechanics of why each works, the failure modes, and the handoff points back to Excel.
Why a prompt can read a P&L better than you expect — and worse than you hope
An LLM is strong at pattern recognition across a schema: it spots that marketing spend grew while new customers did not, that a line item's name changed mid-year, or that gross margin moved for a reason unrelated to COGS. It is weak at arithmetic you did not show it. Ask it to compute a 37-row sum and it may drift; give it the rows and ask it to check a sum and it usually catches the mismatch.
So the rule that governs every prompt below: let the model reason about structure, and let Excel (or Python) do the math. Where a prompt must compute, make it show its work line by line so you can verify.
The handoff boundary
| Need | Prompt workflow | Excel by hand |
|---|---|---|
| First-draft unit economics | 10–20 min | 2–4 hours |
| Scenario framing (what to stress) | Excellent | Manual, slow |
| Anomaly triage in a messy export | Strong | Formula-heavy |
| Final model for audit / data room | Not acceptable | Required |
| Formula documentation | Good draft | Authoritative |
The prompts: from unit economics to anomaly hunting
1. The unit-economics skeleton
Task: turn a business description into a contribution-margin structure before you touch a spreadsheet.
You are a financial analyst. Business: [describe: product, price,
customer acquisition channel, delivery model].
Build a unit-economics skeleton with these lines:
Revenue per unit, variable COGS, gross margin, CAC,
payback period, LTV with an explicit churn assumption.
For each line, state the formula and the input you'd need.
Do NOT invent numbers — mark each input as [NEED].
Flag which inputs are usually the least reliable.
The last instruction is the valuable one: it forces the model to name which assumption will break the model first — often CAC or churn, rarely price.
2. Payback math, shown line by line
Task: compute CAC payback with visible arithmetic so you can verify.
CAC = 420, monthly gross margin per customer = 35.
Compute payback in months. Show the division.
Then recompute if churn rises so average lifetime
drops to 14 months — does payback still clear the
lifetime? State the break-even churn.
With those inputs, 420 / 35 = 12 months payback. If lifetime is 14 months, the margin for error is two months — thin. This is arithmetic you can check, not a statistic to trust blindly.
3. The P&L structure validator
Task: check that a P&L's line items roll up correctly and consistently.
Here is a P&L export [paste rows: line, amount, period].
1. Verify subtotals equal the sum of their components.
2. Flag any line whose sign is unusual for its category.
3. Flag lines whose name changed between periods.
Report mismatches with the exact rows. Do not fix — report.
4. Margin bridge (why did gross margin move?)
Task: decompose a margin change into price, volume, and mix effects.
Gross margin fell from 41% to 36% between Q2 and Q3.
Here are revenue and COGS by product line [paste].
Decompose the change into: price effect, volume effect,
mix effect. Show each step's arithmetic. State which
effect dominates.
This is where LLMs shine: a mix shift toward low-margin products is easy to miss in a summary view and obvious once decomposed.
5. Scenario setter: three cases, named honestly
Task: build base / stress / severe scenarios with explicit triggers.
Business: [describe]. Key drivers: price, volume, CAC, churn.
Build three scenarios. For each: the assumption change,
the P&L impact in %, and the TRIGGER that would tell
us we are in that scenario (a metric and a threshold).
Name each scenario after its trigger, not 'optimistic'.
Naming a scenario after its trigger — "churn above 4%/month" — is a small trick that changes behavior: people act on triggers, not adjectives.
6. Sensitivity table generator
Task: produce a two-way sensitivity grid to paste into Excel.
Drivers: price (rows: -10%, -5%, 0, +5%, +10%) and
volume (cols: same). Base gross margin 38%.
Produce a Markdown table of resulting gross margin,
assuming variable cost per unit is fixed at 62% of base price.
Show one worked cell so I can verify the method.
7. Anomaly hunter on a messy export
Task: triage a raw ledger export for outliers and duplicates.
Here is a transaction export [paste ~50 rows].
Find: duplicate entries (same vendor, amount, near-date),
outliers (>3x median for that category), and entries
with a category that contradicts the vendor's usual one.
List each with row number and why it's flagged.
Do not delete anything.
8. Formula documentation back into Excel
Task: convert a plain-English model description into Excel formulas you can paste.
For a model with inputs in B2:B6 (price, units, cogs_pct,
cac, churn), write Excel formulas for: gross margin,
contribution, LTV, payback. Use named ranges.
Explain each formula in one line.
9. The "what would have to be true" inverter
Task: work backward from a target to the assumptions it requires.
Target: reach 20% net margin next year. Current: 11%.
List every lever (price, mix, CAC, opex, churn) and
for each, the magnitude of change required to close
the gap alone. Rank by plausibility and say why.
10. Cohort sanity check
Task: test whether retention assumptions are internally consistent.
Cohort retention: M1 82%, M3 61%, M6 44%.
Check: does this imply a plausible monthly churn curve?
Compute the implied monthly churn between each pair.
Flag any inconsistency.
11. The red-flag checklist for a P&L
Task: generate a review checklist tuned to the business model.
Business model: [SaaS / marketplace / retail].
Give me a 10-point P&L red-flag checklist specific to
this model — the things that hide problems.
For each: what it looks like and what it usually means.
12. Variance commentary draft
Task: turn actual-vs-budget numbers into a first-pass narrative.
Actual vs budget [paste]. Write variance commentary:
for each line >5% off, state the likely driver and
whether it's timing or structural. Mark anything
you cannot explain from the data as [INVESTIGATE].
13. The audit-trail handoff
Task: convert the prompt-built draft into a spec for the Excel model.
Take this unit-economics draft [paste]. Produce a build
spec: sheet names, input cells, formula logic per cell,
and a list of assumptions needing a source.
Format so a modeler can build it without asking questions.
This prompt is the bridge. It is how you stop using prompts and start using Excel without losing the reasoning.
14. Break-even and runway, computed visibly
Task: compute runway from burn and cash with the math shown.
Cash = 900,000. Monthly net burn = 75,000 and rising
8% per month. Compute runway in months, showing each
month's cash balance. State the month cash hits zero.
15. The devil's-advocate reviewer
Task: attack your own model before someone else does.
Here is my model's assumptions [paste]. Act as a skeptical
investor. List the three assumptions most likely to be
wrong, the evidence that would confirm each, and the
P&L line each one damages most.
16. Prompt hygiene check
Task: audit your own prompt library for ambiguity.
Review this prompt [paste]. Identify where it allows the
model to invent numbers, where it lacks a verification
step, and where output format is ambiguous. Rewrite it
tighter.
Where this breaks — and how to spot it early
Three failure modes recur. Silent arithmetic drift: the model sums a column wrong in a long context; catch it by always asking for a worked example cell. Schema blindness: it treats a renamed line as a new category; catch it with prompt #3's consistency check. Confident framing: it presents a plausible scenario as if it were forecast — a prompt asking for triggers, not adjectives (#5), keeps it honest.
A non-obvious insight worth keeping: the most valuable prompt in this list is #13, the handoff. The purpose of prompting a financial model is not to replace Excel — it is to arrive at Excel with the structure already argued and the assumptions already named. That is where the hours actually get saved.
If you want to sharpen the underlying skill, the prompt engineering learning plan covers the mechanics, and the HR analytics dashboard guide shows the same draft-then-lock discipline applied to a different domain.
Start with prompt #1 and #13 on your next model. You will feel the difference in the first hour.