Large language models (LLMs) have learned not only to write text, but also to analyze tabular data, build reports, and uncover insights. You can give a model a log file, a financial table, or a raw CSV — and get:

• a summary of metrics,
• automatic visualization,
• natural-language recommendations.

📘 What You’ll Learn

• How to process tables with an LLM
• How to transform raw text into a structured report
• How to build automated reporting for CRM, logs, and surveys
• How to use Google Sheets together with ChatGPT or Gemini

📚 Key Terms

• Code Interpreter — A module in ChatGPT that can run Python code, process CSV and Excel files, build charts, and analyze data.
• EDA (exploratory data analysis) — Exploratory analysis: summaries, pattern detection, charts, distribution and correlation analysis.
• Pivot Table — A tool for aggregating data by groups. Used for counts, sums, averages, and other metrics.
• CRM — Customer Relationship Management system. Examples: amoCRM, HubSpot, Bitrix24. Stores data about customers, deals, and activity.
• Insight — A useful observation or conclusion from data analysis that helps make a decision or identify a problem.

🔄 How the Process Works

An LLM (the same GPT model with Code Interpreter or via integration with Google Sheets) can work with real data like a universal analyst-assistant. And it doesn’t matter what form the data arrives in — a table, raw logs, or unstructured text. The model adapts and takes the first steps of analysis for you.

1. 📊 CSV, Excel, Google Sheets

The most typical case: you have a table — an export from CRM, sales, budget, attendance, etc.

How it works:

• You upload a CSV or Excel file to Code Interpreter (or use an add-on in Sheets),
• The model understands the structure,
• And on the fly builds:
  • column summaries (e.g., “average order value,” “max/min sales”),
  • conditional filters (e.g., “all deals over 500$”),
  • charts: by day, region, category.

📌 Prompt example:

Here’s a file with sales by region. Build a summary for each manager: number of deals, average order value, and a bar chart.

📎 Analogy: Before, you opened Excel, set filters, wrote formulas, and drew charts by hand. Now it’s like having an assistant you simply tell: “Show sales by manager, compute the average — and build a chart.”

2. 📝 Raw Text

Data often comes as plain text — a client email, a chat message, a note. Previously you had to manually spread such lists across table rows. An LLM can recognize structure in the unstructured and convert it to tabular form.

📌 Prompt example:

Here’s a list of payments for May: Smith — 1200 $, Wood — 1450 $, Peterson — 1100 $. Make a table and calculate the total.

The model will:

• extract names,
• recognize amounts,
• format everything as a table,
• and immediately compute the total.

📎 Analogy: As if you dictated to an assistant: “Here’s the list — put it into a table and give me the total.”

3. 📁 Raw Logs

For more technical tasks: if you have log files (e.g., Nginx, Apache, CRM activity reports) — an LLM can look for patterns and anomalies.

📌 Prompt example:

Here’s a week of website login logs. Find peak hours and IPs that accessed suspiciously often.

The model will:

• parse timestamps,
• build a distribution by hour,
• highlight IPs with anomalous frequency,
• and propose a visualization or a tabular output.

📎 Analogy: Instead of poring over logs and building filters in Splunk or Excel, you have an analyst who “reads” them, understands them, and explains where “something is off.”

Here’s a detailed expansion of this section with analogies:

📊 Prompts for Analysis

A strong side of LLMs is quickly turning data into conclusions. The model can do the routine: aggregate, filter, summarize, and propose hypotheses. You just need to craft the prompt correctly.

1. 📑 Tabular Analytics (EDA — exploratory data analysis)

If you have a table (for example, sales.csv), you can immediately ask for a basic analysis:

Load this CSV and show:

Total number of rows

Number of unique customers

Average order value

Most frequent payment method

📌 You can additionally request:

• “Build a histogram for the amount column” — to see the distribution of amounts.
• “Find outliers in order values” — abnormally large or small values.
• “Filter all deals > 10000 $ and in San Diego” — like an Excel filter, but in plain text.

📎 Analogy: Imagine sitting with an analyst and saying: “Check how many customers we had, what the average order value is, and plot a chart of payments.” Now it’s a model that does it instantly.

2. 📊 Pivot Tables

When you need to aggregate by groups — managers, regions, products — use a pivot table.

Build a pivot table: for each manager — sum and count of orders.

The model will:

• group by the required column (manager),
• compute sum total and record count,
• return a table suitable for a report or presentation.

📌 You can specify:

• “First sort by descending sum”
• “Add a column with share (%) of the total volume”

📎 Analogy: It used to take 10–15 minutes in Excel: “Insert → Pivot Table → Group…”. Now you just write to ChatGPT: “Show me top managers by sales.”

📌 In Google Sheets it works the same way:

• export the table,
• paste it into the prompt,
• get a =QUERY(...) or =PIVOT(...) formula.

3. 💡 Insights and Hypotheses

One of the most useful functions is data interpretation. The model can tell you what’s important that’s going on:

Here are the data for March. What 3 conclusions can we make? Any anomalies? What should we check next?

📌 The model can:

• compare with the previous period, if you provide it,
• find spikes or drops,
• highlight anomalous customer groups or order values.

You can add to the prompt:

• “Write like a business analyst”
• “Make a short summary for management (in 3 bullet points)”
• “Propose 2 hypotheses to grow sales”

📎 Analogy: Previously you needed a BI specialist or consultant. Now it’s like consulting a senior analyst, but in 15 seconds and without fluff.

🗂 Converting Text to a Table

Sometimes you don’t have a table — only text from a chat, an email, meeting notes.

Convert this into a table with columns: name, comment, sentiment.

📌 The model can:

• extract structure — who said what, to whom it relates, date/time,
• determine sentiment — positive, neutral, negative,
• output the result as a table, CSV, or Markdown.

Example:

“Smith: everything works great. Wood: the signup form won’t submit. Peterson: the interface is too complex.”

The model will transform it into:

📎 Analogy: It’s like you dictated notes to an assistant who then created a table with flags and counts of who complained how many times.

📤 Automatic Reports

One of the most underrated LLM features is automatic generation of textual reports based on tables. Especially useful for:

• managers,
• marketers,
• analysts,
• instructors.

🔁 How it works

1. You provide a table: .csv, .xlsx, copied data from Google Sheets, or a text table.
2. You specify the format and goal of the report.
3. The model produces text that draws conclusions, compares periods, highlights important points, and gives recommendations.

✏️ Prompt example:

Build a weekly report:

– Which products sold best

– Which regions dropped in revenue

– Which employees are in the top

Use a businesslike yet clear style. The report should be in English.

📎 Analogy:

Previously you:

• exported data manually,
• wrote a summary in Word or Notion,
• spent a long time wording conclusions.

Now it’s as if you have an assistant who:

• sees the table,
• identifies the interesting numbers,
• writes a clear text with figures, comparisons, and even charts.

🌐 Language and style — your choice

You can specify:

📘 Sample report topics:

• Sales by region: what grew, what fell
• Marketing: channel effectiveness, cost per lead
• Team: who performed best against KPIs
• Finance: month-over-month revenue comparison, forecast

🧠 Why this is useful

• 📉 Automation — daily and weekly reports without manual work.
• 🎯 Focus on the essential — the model filters noise and highlights key metrics.
• 📣 Communication — generate a report “for SEO,” “for a client,” or “for a presentation.”

🛠 Tools

Working with data via an LLM is possible in your familiar environment — e.g., in Google Sheets or via ChatGPT’s chat interface. Below is a breakdown of tools with examples of when and how to use them.

🧠 OpenAI Code Interpreter (Advanced Data Analysis)

This is a module in ChatGPT Plus that can run Python code, analyze files, and build visualizations.

📌 What it can do:

• 📁 Load and read .csv, .xlsx files
• 📊 Use pandas, numpy, matplotlib — like a full Jupyter Notebook
• 📈 Build histograms, pie charts, and line graphs
• 🧾 Compose textual reports and conclusions

Analogy: Imagine you have a built-in Python analyst inside ChatGPT. No setup needed — you just provide a file; it plots charts, computes, and explains.

📍 Available only with an active ChatGPT Plus subscription → “Advanced Data Analysis” tab

🌐 Google Sheets + ChatGPT

If you work in Google Sheets but don’t memorize formulas — ChatGPT can be your personal Sheets consultant.

📌 What you can do:

• Upload a table to Google Sheets
• Copy the necessary part (e.g., the orders table)
• Paste it into the prompt and ask a specific question

📝 Prompt example:

Here’s an orders table. Help calculate the average order value by channel. Give me the formula and where to paste it.

📎 Analogy: It’s like talking to an assistant who knows Google Sheets well and says: “Here’s the formula, here’s where to put it, and here’s why it works.”

🔧 For advanced users: You can connect the ChatGPT API to Google Apps Script and generate reports automatically right inside the sheet. This is no-code analytics with a custom integration.

🔁 Gemini (with Sheets)

Google’s new Workspace-integrated toolset — great for quick work with documents, spreadsheets, and presentations.

📌 Highlights:

• Integration inside Google Sheets: write directly in a cell or in Gemini’s side panel
• Can read tables, suggest formulas, analyze data series
• Can generate report texts, explain metrics, propose visualizations

Sample task:

Explain why B2C orders dropped in March and which channels should be strengthened.

📎 Analogy: It’s like hiring a Google analyst who sits right inside your sheet and says: “Here’s what’s wrong, here’s where to dig, here’s a formula, here’s the insight.”

✅ When to Use What

🧨 Common Mistakes

• Poorly structured data — First ask the model to “clean” and normalize
• Tables that are too long — Use excerpts or upload the file
• No task formulation — Add: “Show trends,” “Find anomalies,” “Make conclusions”

❓ FAQs

Q: How much can we trust LLM analytics?

A: About 80%. The model can make logical mistakes, but as a “first draft” of a report — it’s powerful.

Q: Can this be used in production?

A: Yes, with post-verification. Especially for internal reports, feature analytics, and email digests.

Q: Any alternative without a ChatGPT subscription?

A: Try Gemini, Claude, or Google Sheets + Apps Script.

📌 Summary

• LLMs can read and process tabular data
• You can generate reports, charts, and insights from CSV, Excel, or CRM
• It’s a great way to automate analytics and save time

👉 In the next section you’ll learn how to use LLMs for content review and localization, as well as generating email sequences.