Put a Real Analyst Inside Your Excel

The analysis is already sitting in your exports

Most businesses are drowning in reports nobody interrogates. Sales by store, inventory aging, parts velocity, ordering reports — they get glanced at and closed. The data that would change your ordering, your staffing, or your product mix is usually sitting there the whole time. What's missing is the hours to question it properly.

That's the part AI actually fixes. You don't need a data team. You need a spreadsheet you already have, a $20-a-month AI account, and one good prompt. By the end of this guide you'll have controller-grade analysis of your own numbers, in one sitting.

This guide uses a running example the whole way through: 24 months of parts sales across 3 locations. Swap in your own data — the steps are identical.

Step 1: Get your export

You need one spreadsheet export from whatever system you already use — QuickBooks, your POS, your DMS, your ERP, even a report someone emails you monthly.

Good candidates:

• Sales by month, by product, by location
• Inventory aging (item, quantity on hand, last sold date, cost)
• Order history (date, item, quantity, vendor)

A good export looks like this:

• One header row at the top. Column names like Month, Location, Part Category, Units Sold, Revenue. No merged cells, no title rows above the headers, no logo.
• One row per record. One month of sales for one part category at one location = one row. Not a pivot table, not a formatted summary with subtotals mixed in.
• 12+ months of history. You can't see seasonality with 6 months. 24 is better.

The running example: a file called parts-sales.csv with columns Month, Location, Part Category, Units Sold, Revenue. Three locations, 24 months, about 40 part categories. Roughly 2,800 rows. That's a small file. AI handles it easily.

How to get it: in most systems, find the report you want, look for Export, and choose CSV or Excel (.xlsx). Either works. If your only option is a formatted report with subtotals, export it anyway — you'll have the AI flag the mess.

Step 2: The privacy prep (do not skip this)

Two rules before anything leaves your computer.

Strip customer names and PII. Open the export. Delete any columns with customer names, phone numbers, emails, addresses, VINs tied to a person. For sales analysis you don't need to know who bought — you need what, when, where, and how much. Delete the column, save the file.

Use a business-tier account with training turned off. On Claude, that's a Claude for Work plan (Team or Enterprise) — commercial terms mean your data isn't used to train models. On ChatGPT, that's Team or Enterprise, or at minimum go to Settings → Data Controls and turn off "Improve the model for everyone" on a paid personal plan. The business tiers cost around $25–30 per user per month. If your numbers are sensitive enough that you'd not email them to a stranger, they're sensitive enough to spend $25 on the right account.

One honest note: uploading files and analyzing spreadsheets requires a paid plan on both Claude and ChatGPT. The free tiers will frustrate you here.

Step 3: The analyst master prompt

Go to claude.ai (or chatgpt.com), start a new chat, and click the + (paperclip on some layouts) to attach your file. Then paste this prompt. All of it.

You are a skeptical financial controller reviewing this data for the owner of
a small multi-location business. You are thorough, you show your work, and you
never invent numbers. If something can't be answered from the data, say so.

I've attached a spreadsheet export: [describe it — e.g., "24 months of parts
sales across 3 locations, one row per month/location/part category, columns:
Month, Location, Part Category, Units Sold, Revenue"].

STEP 1 — PROFILE THE DATA FIRST. Before any analysis, tell me:
- Every column, what it appears to contain, and its data type
- The date range covered, and whether any months are missing
- Gaps, blanks, duplicates, or oddities (negative quantities, zero-revenue
  sales, categories that appear and disappear)
- Anything about the structure that could mislead the analysis
Wait for nothing — do this profiling, then continue to Step 2 in the same
response.

STEP 2 — ANSWER THIS FIXED QUESTION SET:
1. Trends: Is total revenue and unit volume growing, flat, or declining?
   Show the numbers by year and by location.
2. Seasonality: Which months are reliably strong or weak? Is the pattern
   consistent across both years and all locations?
3. Top and bottom performers: Top 10 and bottom 10 part categories by
   revenue AND by units. Note where the two lists disagree.
4. Concentration risk: What share of revenue comes from the top 5
   categories? From the single biggest location?
5. Anomalies worth investigating: List the 3-5 strangest things in this
   data — sudden drops, spikes, a location that behaves differently.
   For each one, say what it might mean and what question I should ask
   my team about it.

RULES:
- Show your reasoning. When you state a number, say how you got it.
- Flag every data-quality problem you find, even small ones.
- Never invent or estimate numbers that aren't derivable from the file.
- Write for an operator, not an analyst. Plain language, no jargon.

Fill in the bracketed description with your own file's shape, hit enter, and read what comes back.

The profiling step matters more than it looks. Real exports are messy in ways that quietly poison analysis: a location missing a month because of a POS migration, a part category renamed midway through the data so it looks like one product dying and another launching. A human analyst catches that. Generic AI use doesn't — unless you make it profile first. That's the difference between an answer and an analysis.

Step 4: Five follow-up prompts that go deeper

The first pass tells you what's happening. These tell you what to do. Paste them one at a time, in the same chat, so the AI keeps the full context.

1. The cut list:

Based on this data, which part categories would you cut or stop stocking,
and why? Be specific about the revenue we'd give up and what we'd gain.
Then argue the other side: give me the best reason to keep each one.

2. Reorder points:

Build a simple reorder point for each of the top 20 categories: average
monthly units, peak monthly units, and a suggested reorder trigger assuming
a 2-week lead time. Show it as a table I can hand to a parts manager.
State your assumptions clearly.

3. Questions for the team:

Take the three biggest anomalies you flagged earlier. For each one, write
the exact questions I should ask my store managers to figure out whether
it's a data problem, a process problem, or a real business change.

4. The location comparison:

If Location 2 performed like Location 1 on a per-category basis, what would
its revenue have been over the last 12 months? Where are the biggest gaps,
and which ones look fixable versus structural?

5. The one-page summary:

Write a one-page summary of everything we found, for me to bring to my
Monday manager meeting. Three sections: what's working, what's broken,
what we should decide this month. Plain language. Include the specific
numbers.

Step 5: Verify before you act

AI models are good at this work and they still make arithmetic mistakes. Treat the output like a first draft from a smart new hire.

• Pick 3–4 numbers from the analysis and check them against the raw file. Open Excel, filter to the category and month, sum the column. Does it match?
• The more a number matters, the harder you check it. "December is our weak month" — quick glance. "Cut these 8 categories" — verify every one before you touch inventory.
• If a number doesn't match, tell the AI: "Your figure for Location 2 in March doesn't match my file — recheck it." It will usually find its own error. That's fine. That's the process working.

Never invent numbers was a rule in the prompt. Verification is how you enforce it.

Step 6: Make it routine

The real payoff isn't one analysis. It's the monthly habit.

1. Same export, same prompt, first week of every month. Save the master prompt in a note so you're not rewriting it.
2. Keep a running "questions for the team" list. Every month the anomaly section feeds it. Bring it to your manager meeting.
3. After three months, add one line to the prompt: "Compare against the issues we found in prior months — what got better, what got worse?" (Paste in last month's one-page summary so it has the context.)

The wins in data-heavy operations rarely come from one heroic analysis. They come from asking the same hard questions of the same data every month until the answers change. This gets you that habit for the price of a lunch. If the analysis keeps surfacing manual work behind the numbers, the manual-work inventory is how you size it.

Tool names, plans and prices in this guide were checked in August 2026. AI products change fast — verify anything before you buy it.

Where Hoven fits

If you'd rather have someone find where this actually pays off in your business — and where it won't — that's what I do. Most clients start with a $999/day audit; most take 1 to 5 days. You get a prioritized plan plus a "don't build this" list, and if it isn't worth more than you paid, you don't pay. Details on the pricing page, or get in touch.