Generative AI for Analysts: How It Works, Prompting & AI vs GenAI vs Agentic AI

The AI family tree

AI vs GenAI vs agentic AI

How an LLM works (without maths)

An LLM is trained on vast amounts of text to predict the next token (a word or word piece). Repeating that prediction one token at a time produces sentences, code and tables. It has learnt statistical patterns of language, facts and reasoning styles, but it does not look things up unless it is given a tool or documents.

Prompt engineering for analysts

A prompt is an instruction plus the material the model needs. The most reliable structure has five parts:

Weak vs strong prompt

Weak: "Analyse my sales data."

Strong:

Role: You are a senior retail data analyst for an Indian omnichannel retailer.
Context: Table superstore(order_id, order_date, region, category, sub_category, sales, discount, profit),
         one row per order line, Jan 2023–Dec 2025, amounts in rupees. Margin = profit / sales.
Task: Explain why overall margin is only about 9%. Look at category, sub-category and discount band.
Format: (1) a 3-bullet summary for the CFO, (2) a table of the 5 worst sub-categories (sales, profit, margin),
        (3) the SQL Server queries you used.
Constraints: Use only the columns listed. State any assumption. If a number cannot be computed from the table, say so.

The strong prompt names the grain, the metric definition, the output format and the rules for uncertainty. The result is far more likely to be right and checkable.

Core techniques

Prompt templates for common analyst jobs

Choosing and using AI tools

Never paste

1. Personal data (names with phone numbers, emails, Aadhaar/PAN, health, salary of identifiable people) unless your company's approved tool and policy allow it.
2. Confidential business data (unreleased financials, customer lists, contracts, source code, credentials, API keys) into public tools.
3. Anything you are not allowed to share: client data under NDA.

Use enterprise versions with data-protection terms, mask or sample the data (replace names with IDs, share the schema and 5 fake rows), and follow your organisation's AI policy. In India, personal data is also covered by the Digital Personal Data Protection Act, 2023.

A verification habit for every AI answer

Here is why verification matters. Suppose an AI writes the following query for "average order value per region". It looks right, runs without error and returns plausible numbers, but it averages order lines, not orders:

import duckdb, pandas as pd
con = duckdb.connect()
con.execute("CREATE TABLE superstore AS SELECT * FROM read_csv_auto('superstore_sales.csv')")
wrong = con.execute("SELECT region, ROUND(AVG(sales), 0) AS avg_order_value FROM superstore GROUP BY region ORDER BY region").df()
right = con.execute("""SELECT region, ROUND(AVG(order_total), 0) AS avg_order_value
                       FROM (SELECT region, order_id, SUM(sales) AS order_total FROM superstore GROUP BY region, order_id) GROUP BY region ORDER BY region""").df()
pd.DataFrame({"region": wrong.region, "AI query (per line)": wrong.avg_order_value, "Correct (per order)": right.avg_order_value})