⚠️ Privacy: an orders export contains customers' personal data (names, phone numbers, addresses). Before uploading it to a cloud AI, anonymise it (drop or replace the personal columns) or work on a local model (see the combo "Local / private"). That's a legal requirement (152-FZ in Russia).
A ready-made setup: you understand your sales without an analyst and without formulas. You upload a CSV or Excel file of orders — the AI calculates revenue, trends, top products, best customers and average order value, and builds the charts itself. What you get is a short "what's working, what needs fixing" report in plain language.
What you end up with
- The key numbers: revenue, order count, average order value, month-over-month trend.
- Top products and top categories (what makes money and what's dead weight).
- Customer segments: who buys more often or spends more, and the share of repeat purchases.
- Charts (trends, ABC analysis) — with no Excel formulas.
- A short summary report: what to double down on, what to check, and the first 3 things to do.
What you'll need
Tools from the library (by name): ChatGPT/Claude (data analysis and charts — the top-tier ones can run code and read files) or Ollama locally (private, for data that isn't anonymised), the prompt "Break down data/a dashboard: what matters and what to do", the prompt "Customer persona / ICP from data" (describing segments), the Filesystem MCP (giving a local assistant access to the export file). Not in the library but relevant: the orders export itself (CSV/Excel from your CRM, store, marketplace or bank). Accounts/money/hardware:
- Free route: ChatGPT/Claude (free) — for an anonymised export; any PC.
- Private route: Ollama locally (non-anonymised data never leaves your machine; you need a capable PC/GPU — see the combo "Local / private").
- Fast route: a paid model (file analysis, code execution, charts out of the box, large row counts).
Step by step
- Prepare the export. What to do: export your orders (date, product, category, amount, customer, status). Anonymise personal data if it's going to the cloud (drop or replace names, phone numbers, addresses), OR get ready to crunch it locally. How you know it worked: the file opens, the columns make sense, and no personal data is going to the cloud.
- Describe the data to the AI. Tool: ChatGPT/Claude (with file upload) or Ollama + the Filesystem MCP. What to do: upload the file and give it context (what the business is, what the columns mean). How you know it worked: the AI has understood the structure correctly (which column is the amount, the date, the product) and asked about anything unclear.
- The core analysis. Tool: the prompt "Break down data/a dashboard". Ready-made prompt:
Here's an orders export (CSV). Columns: <list them and what they mean>. Business: "<what we sell>".
Calculate and show me: 1) revenue and order count in total and by month (trend, direction);
2) average order value and how it's changing; 3) top 10 products and top categories by revenue (ABC);
4) repeat purchases: share of customers with 2+ orders; 5) anomalies (spikes/dips, refunds).
Build charts wherever you can. Show the calculation/logic first, the conclusion after. Don't invent data that isn't in the file.
How you know it worked: the numbers are calculated from the file (you can see the logic), the charts are built, and nothing has been filled in from thin air. 4. Segments and customers. Tool: the prompt "Customer persona / ICP from data". What to do: ask it to describe who your best customers are (by frequency and spend) and to profile the buyers of your top products. How you know it worked: you can see who brings in most of the revenue (often 20% of customers → 80% of the money) and can target them. 5. Sanity-check the numbers. What to do: verify one or two key figures by hand (monthly revenue, order count) — the AI does sometimes miscount. How you know it worked: your control values match; if they don't, clarify the columns or filters and recalculate. 6. Report and actions. What to do: ask it to boil everything down. Ready-made prompt:
Condense the analysis into a one-page report: 1) the 5 headline numbers; 2) what's growing and what's falling;
3) top products/customers; 4) 3 problems or opportunities; 5) 3 concrete actions for the coming month.
Plain language, no jargon, written for a business owner.
How you know it worked: the report fits on one screen and you can make a decision from it without rereading the spreadsheet.
Free route vs fast (paid)
| Aspect | Free / private | Faster / more powerful (paid) |
|---|---|---|
| Analysis | ChatGPT/Claude free (anonymised) | paid model: files + code + charts out of the box |
| Privacy | Ollama locally (data stays put) | cloud — anonymised data only |
| Volume | split large files up | large row counts in one go |
A one-off analysis is genuinely doable for free on an anonymised export. A paid model is more convenient with large files and builds charts out of the box. For customer data that isn't anonymised — local only (Ollama).
Common problems and fixes
- The AI got the maths wrong. LLMs make arithmetic mistakes — ask it to show the calculation or code, and check one or two control figures by hand (step 5).
- The file is too big or everything crawls. Split it by period or pre-aggregate it (a monthly pivot), then analyse; or use a model with a large context or file mode.
- Column soup and broken dates. Clean up the export format (one date column in ISO, numbers without spaces or currency symbols) and explain every column to the AI.
- Nervous about customer data. Don't upload personal data to the cloud — anonymise it or crunch it locally (Ollama + the Filesystem MCP); that's the law, not paranoia.
- Vague conclusions. Give it business context and demand specifics with numbers and actions (step 6), not "sales look fine".
Time and money
- Time: the first analysis — 30–60 minutes (preparing the file + prompts + verification); a recurring one (same template, same export) — 10–15 minutes.
- Money: free (ChatGPT/Claude free on anonymised data, or Ollama locally) or a little for a paid model to handle large files and charts.
- Honestly: AI replaces manual formulas and builds charts, but it does NOT replace common sense — check the key numbers and remember the privacy rules around customer data. This is a tool for seeing the picture fast, not an auditor's report.