AI is a good guesser. And a bad calculator.

Files, prompt, and exercise from the workshop. Grab what you need.

Step 1

Grab the files

Same workbook and dataset we used in the live build. Download both.

exec-summary-template.xlsx

The workshop workbook

The same example workbook we built live. Reference it for structure, assumptions, and decision logic — not as a universal template for every business.

Download

coffee-shop-data.csv

The demo dataset

The shared Jan–Mar 2026 coffee-shop data we used in the live build and the exercise.

Download

Step 2

Copy the prompt

Works with Claude, ChatGPT, Gemini, or Copilot. Paste your question at the bottom. Let the AI interview you first.

prompt-pattern.txt

This is Maya. She runs a small coffee shop.

January 2026 was cafe-only. On Feb 3, 2026 she started a wholesale channel — beans and pastries sold to a few local accounts on a Tuesday / Thursday / Friday cadence. By the end of March she had two channels running side by side: cafe and wholesale.

The CSV I just uploaded is her Jan 1 - Mar 31, 2026 transaction data, with both channels in it.

She has two questions:
1. What actually changed after wholesale started?
2. Can she afford to hire a helper for 15 hours a week right now?

I'm going to help her build a real Excel workbook that answers both. You're going to help me build it.

Before you write ANYTHING — no spec, no plan, no formulas, no file — interview me until you have a crystal clear brief.

Two things you need to be clear on before building:
1. The model. How I define the economics — what counts as raw data versus what counts as an assumption I provide. You do not get to invent assumptions. If something isn't on the table, ask.
2. The output shape. What level of formula complexity, who reads the result, how the final decision should be presented, and what "done" looks like.

Ask in plain conversation. Do not put me into a structured questionnaire form. Talk like a person. Whatever rhythm feels natural — one question or a small batch — is fine.

Once you believe you have a crystal clear brief, stop and do this:
- Repeat the brief back to me as a short numbered list.
- Show me the exact assumptions you'll use, with the exact values I gave you.
- Wait for me to say "go" before building anything.

Then think carefully before you build. Do not skip the gate.

The workbook structure is fixed: four tabs, in this order.
1. Raw_Data
2. Monthly_Summary
3. Assumptions
4. Executive_Summary

When you build:
- Generate a real downloadable .xlsx file. If your tool can't, say so honestly — do not fake it with a formula dump.
- Use beginner-safe Excel formulas only. SUMIFS, IFERROR, direct cell references. No SUMPRODUCT acrobatics, no array formulas, no hidden helper sheets.
- Every cell must be traceable to either Raw_Data or Assumptions. No magic numbers anywhere in the workbook.
- Keep the raw rows visible on Raw_Data. Don't drop, hide, or filter anything.
- Make it visually obvious which cells on Assumptions are inputs I can edit and which cells are calculated outputs.

Hard rules:
- Don't invent numbers, rates, definitions, or assumptions. If you don't know, ask.
- Don't substitute a different economic model than the one I gave you.
- Don't proceed past the gate without my explicit "go."
- Don't claim you built a file unless an actual .xlsx artifact exists.

Start by interviewing me.

Step 3

Pick one question

One question. Not three. Interview first, build the workbook second, verify in Excel, then explain. Not the other way around.

Prompt 1

Revenue growth source

How much of the January-to-March revenue increase came from wholesale rather than the cafe?

Prompt 2

Revenue vs labor demand

After wholesale started, did revenue grow faster than direct labor demand?

Prompt 3

Revenue minus COGS trend

What happened to monthly revenue minus COGS once wholesale was added?

Prompt 4 · limit case

The under-supported one

Which wholesale product should Maya raise prices on first: cold-brew concentrate or cookie dough?

This one is under-supported on purpose. That's the lesson. The honest answer is probably: we don't have the product-level data to decide.

The Why

The move

Don't ask AI for the answer. Ask it to build you a tool that gives you the answer.

Here's the trap. Drop a messy spreadsheet into a chatbot, ask it a business question, and take the number it gives you. That number is almost always wrong. AI hands you the wrong answer with total confidence, which is the worst way to be wrong.

The move is different. Use AI to figure out the real question, sketch the sheet, and write the formulas. Excel does the math. Then ask AI what the verified numbers mean. That's it.

The workbook is Maya's coffee shop example. Use it as a reference, not a universal template.

Next Monday

Try this on your own data

When you sit down with your own numbers next week, here's the sequence that works.

1

Export one clean CSV

One file. One sheet. No blank columns. No merged cells.

2

Pick one narrow question

Not five. One. Narrow beats broad every time.

3

Ask AI to interview you first

Before it builds anything. Let it surface the assumptions you forgot to name.

4

Tell it to generate a real workbook

An Excel file if it can. Otherwise exact tab layout, formulas, and cells.

5

Open the workbook and verify

Check the cells that matter. Trust the sheet, not the chat.

6

Put assumptions in visible cells

Not hidden in prose. On the sheet, where you can change them.

7

Only then ask AI to explain

Once the numbers are verified. Not before.

Good first questions to try

What changed month to month?
Which channel drove the change?
Did revenue grow faster than labor?
Can I afford the next hire under clear assumptions?

Keep this rule

If the data can't support it, say so.

Sometimes the honest answer is: we need a better assumption, or cleaner data, or more detail before we decide. That's still a useful answer.