What Conversational AI Can and Cannot Do With Your Data

A chart can look impeccable and, at the same time, be completely wrong. You get it in a few seconds by asking the AI for a visualization, you drop it into a presentation for leadership, and if a faulty aggregation sits behind it, nobody will challenge it anymore — because it looks convincing. This is the real stake of data analysis with AI: not the speed at which you reach a result, but the ability to know whether that result is true. The very same tool that can make you several times more productive can also lead you, with utter serenity, to a wrong conclusion. The difference between the two lies in what you know to ask of it and what you know to verify — exactly what we build in this course.

Why this lesson comes first

Most materials about "data analysis with AI" start with an impressive trick: you upload a file, ask for a chart, get a chart, applause. The problem is that the initial enthusiasm masks precisely the risks that matter in a professional context. A beautiful chart generated from a wrong aggregation is more dangerous than no chart at all, because it looks convincing and ends up in a presentation for leadership, where nobody has time to challenge it anymore.

So our pedagogical contract is simple and firm:

• AI accelerates the mechanical work of cleaning, describing, and visualizing data.
• AI does not guarantee the correctness of the results it produces.
• The final responsibility for the numbers remains yours, no matter how articulate the answer sounds.

You will notice throughout the course that we work exclusively with synthetic data — invented datasets, with no real people — precisely so we can experiment freely, without exposing confidential information. This is not a cosmetic detail, but a professional rule, as you will see in the section about GDPR. All the example files in this course (for instance synthetic_orders.csv) are fabricated specifically for learning and do not correspond to any real person or company.

Who this course is for

The course is built for non-programmers: people who work with data every day — in sales, marketing, operations, administration, operational finance — but who do not write code and do not want to learn a programming language just to do their job. You do not need knowledge of Python, R, or advanced SQL. You need two things: curiosity about what your data says and the discipline to verify what the AI answers. The first one you already have, since you are here; the second one we build together.

What conversational AI does WELL

There is a clear category of tasks where an AI assistant like ChatGPT (with the GPT-5.5 model in 2026) saves you entire hours of repetitive work. We group them into four zones of real competence.

1. Fast data cleaning

Real-world data is dirty: extra spaces, inconsistent date formats, values written as "YES/Yes/yes", chaotically named columns, numbers stored as text. AI excels at proposing and executing cleaning transformations, especially when you describe the problem in natural language.

Example prompt (synthetic data):
"I uploaded a CSV file with 1,200 synthetic orders. The
'payment_status' column contains the values: 'paid', 'Paid', 'PAID',
'settled', 'unpaid', 'Unpaid'. Normalize these values into exactly two
categories: 'Paid' and 'Unpaid'. Show me the mapping table before you
apply it and tell me how many rows fall into each category."

Notice the structure of the prompt: you ask for the explicit mapping and the row counts before applying the change. That gives you a checkpoint — exactly the mentality we cultivate in the next lesson. The difference between a naive user and a competent one is not the tool, but the prompt: the competent one always asks for intermediate evidence.

2. Descriptive statistics

AI quickly produces descriptive summaries: mean, median, minimum, maximum, standard deviations, distributions by category. These are well-defined and deterministic calculations, and the execution environment (such as ChatGPT Advanced Data Analysis, which runs Python code in a sandbox) computes them accurately, rather than "guessing" them.

A well-requested descriptive summary gives you, in a few seconds, an X-ray of your data quality before any serious analysis. Always ask first "how many values are missing, and in which columns?" — the answer tells you how much you can trust everything that follows.

3. Visualizations from a prompt

One of the most appreciated capabilities: you describe what you want to see, and the AI generates histograms, bar charts, scatter plots, time series. For an initial exploration, this drastically shortens the distance between question and image.

Example prompt:
"Generate a histogram of the values in the 'order_value_eur' column
(synthetic data). Use 20 bins. Add a vertical line at the median.
Tell me explicitly what the value of the median is."

Visualization is not a goal in itself, but a discovery instrument: a histogram instantly shows you whether the distribution has a long tail, whether there are outliers, or whether the data clusters into two peaks (which would suggest two mixed populations).

4. Formulating queries

Here AI becomes an excellent translator between your intention in natural language and technical syntax. You tell it what you want to find out, and it proposes an SQL query or an Excel formula. The "text-to-SQL" landscape has matured exactly this use case: you no longer need to memorize the syntax, but to understand the logic.

-- Query proposed by AI from a natural-language request (conceptual)
-- "How many orders over 100 EUR do we have per month?"
SELECT
    strftime('%Y-%m', order_date) AS month,
    COUNT(*) AS order_count
FROM synthetic_orders
WHERE order_value_eur > 100
GROUP BY month
ORDER BY month;

Important: the AI gives you a verifiable starting point, not an absolute truth. You validate that the query reflects the real question. A wrongly placed filter (> 100 instead of >= 100) or a confused column changes the result completely, without the query looking wrong.

What AI does POORLY or RISKILY

Now, the part that many materials hide. There are tasks where blind trust in AI costs you dearly.

1. Complex calculations "in its head"

There is an essential technical distinction. When the AI executes code (for example, in ChatGPT Advanced Data Analysis, where it runs Python on your file), the calculations are reliable. But when the AI answers directly, in text, without running code, it "predicts" the numbers based on language — and it can get a sum or a percentage wrong with disarming serenity.

Golden rule: For any number that matters, make sure the AI has executed a calculation (ran code or a formula), not merely asserted a result in prose.

This distinction comes back throughout the course, because it is the boundary between a trustworthy tool and a convincing illusion. Always ask yourself: was this number calculated or narrated?

2. Aggregations on large datasets

Conversational tools have real limits on file size and the number of files you can upload. According to the public documentation in the OpenAI Help Center, file uploads in ChatGPT are subject to limits on size and on the number of files per conversation. When you work with large datasets, you risk the AI processing only a part (a sample), truncating, or summarizing — without clearly warning you. The conclusion: for millions of rows, the right instrument remains a database or a dedicated engine, with AI as an assistant for formulating the query.

3. Causal inferences

This is the gravest intellectual trap. AI can observe that two variables grow together and can formulate a sentence that sounds like causality. But correlation is not causation, and a language model has no way to establish a cause-and-effect relationship from numbers alone. If, in a synthetic dataset, ice cream sales and drowning incidents rise together, the common cause is summer, not the ice cream. Confusing the two leads to absurd decisions: you would stop selling ice cream "to save lives".

4. Confidential and personal data

This is both a technical risk and a legal one. We talk about it explicitly below, because a mistake here is not fixed by rephrasing a prompt — it can carry legal consequences.

The course's positioning map

Throughout the course you will work with three families of tools. Here is their map, at the paradigm level (we will go into detail in the lesson about the landscape):

Remember the commercial caveat, because we repeat it every time: Microsoft 365 Copilot is not free. The COPILOT() function and the Copilot assistant in Excel require a paid M365 Copilot license, on top of the regular Microsoft 365 subscription. That is why we always teach the ChatGPT alternative as well, so you do not get stuck behind a license.

Legal guardrail: GDPR and your data

Before uploading any file into a cloud-hosted AI tool, stop and ask yourself a simple question: does this file contain personal data or confidential information?

• Do not upload personal or confidential data into a public/free tier of a cloud tool without a legal basis and without the necessary contractual guarantees.
• For real business data, use Enterprise tiers that offer a Data Processing Agreement (DPA) and guarantees that your data is not used for model training.
• Alternatively, work with anonymized or synthetic data — exactly what we do in this course.

In addition, the EU AI Act introduces, through Article 50, transparency obligations regarding interaction with AI systems, with staged application (the relevant transparency provisions become applicable from 2 August 2026). For you, as an analyst, the practical message is: treat AI tools like collaborators to whom you do not give uncontrolled access to sensitive information.

This course does not provide regulated legal, financial, or accounting advice. For decisions with legal or fiscal implications, consult an authorized specialist.

The right mindset: AI as an amplifier

The fundamental conclusion of this introductory lesson: conversational AI amplifies your analytical productivity, but does not substitute your critical thinking. Used well, it eliminates the boring work and lets you focus on interpretation. Used poorly, it offers you an illusion of competence that collapses at the first serious check.

Think of AI as a very fast and very well-read intern, but one without experience and without a sense of the stakes: it does in a few seconds the work you would need an hour for, yet it does not know on its own when it has made a mistake. Your value does not disappear; it shifts from execution to supervision, asking good questions, and verification.

Recap in three ideas

1. AI does well: cleaning, descriptive statistics, visualizations, query formulation.
2. AI does riskily: calculations "in its head", aggregations on large datasets, causal inferences, handling confidential data.
3. You remain responsible for correctness; you work with synthetic/anonymized data and you comply with GDPR.

By now you have the map: where AI shines, where it betrays you, and why the final responsibility for the numbers remains yours. What comes next is exactly the instrument that separates the enthusiastic amateur from the trustworthy analyst — the verification mindset — which we build step by step, then put to work on concrete files, lesson after lesson, until it becomes a reflex.