Module M4 · Understanding AI · Demystification & Use Cases · 2026
Module 4 · Half day

Understanding
Artificial Intelligence

Demystification, real limitations & concrete banking use cases

AI is not magic. Nor is it inherently dangerous. It is a tool — powerful, useful, but limited — that does not replace human judgment. This module gives every employee the keys to understand what AI actually does, what it does not do, and how to use it wisely in a banking context.

Learning Objectives
01Understand what an AI model truly is — without code or mathematics
02Distinguish the types of AI used in banking and their maturity levels
03Identify hallucinations, bias and real limitations — and how to detect them
04Recognize suitable vs. unsuitable AI use cases within your business scope
All Employees Business Managers Compliance & Risk C-Suite Prerequisite: completed M3
Pejman Gohari · CDO · Chief AI Officer · ORBii
25 years banking experience · DataLab SG · Data Factory Bpifrance · BPCE SI · Author DUNOD · Professor IESEG
academy.orbii.tech
ORBii.Academy · M4 · Understanding AI · Demystification & Use CasesConfidential · 202601
M4 · Understanding AI · 02
Section 1

What AI is — and what it is not

"AI is not intelligence. It is a highly sophisticated statistical system that recognizes patterns in massive data. What it does very well: synthesize, classify, predict on data similar to what it has seen. What it does not do: understand, judge, reason autonomously."
— Pejman Gohari · CDO · Chief AI Officer · ORBii · Author DUNOD 2022/2024

The cooking recipe metaphor

Imagine a chef who has read 500 million recipes. He can reproduce any dish with remarkable precision. He can even invent recipes that "look" good. But he doesn't know if the dish will be good for you specifically — he doesn't know you. And if you ask him to cook a dish from a country that doesn't exist in any of his recipes, he will invent something that looks right — but with confidence.

This is exactly what an LLM does. It has "read" a massive amount of text. It predicts the most likely word following your sentence. It generates text that resembles the text it learned from — without "knowing" if it is true.

How an LLM Works
Statistical token prediction
An LLM (Large Language Model) is trained on billions of texts. For each query, it calculates the statistical probability of the next word (token) to generate, considering context. It does not access any external database in real time (unless connected via RAG). It does not "search" for the truth — it generates what is statistically probable. The coherence and fluency of the produced text does not imply that its content is accurate.

AI: fact or fiction — 8 common misconceptions

FALSE

"AI understands what it says." No. It generates statistically coherent text. There is no comprehension, no consciousness, no intent.

FALSE

"If AI says something confidently, it must be true." This is the very definition of hallucination: a false answer stated with certainty.

FALSE

"AI will replace my job." It automates repetitive and analytical tasks. It does not replace judgment, client relationships, or accountability.

FALSE

"AI is neutral and objective." It reproduces the biases present in its training data. It can discriminate without "knowing" it.

TRUE

"AI can process large volumes of text very quickly." Summarizing long documents, extracting information, classification — this is where it excels.

TRUE

"Credit scoring AI can be audited." Mandatory under the EU AI Act for high-risk systems. Explanation of the decision is a right.

TRUE

"Data quality directly impacts AI output quality." "Garbage in, garbage out." AI fed with poor data produces poor decisions.

NUANCE

"AI is dangerous." Neither dangerous nor harmless — it is powerful and poorly managed. It is the absence of governance and training that creates the risk.

ORBii.Academy · M4 · Understanding AI · Demystification & Use CasesConfidential · 202602
Protected Content

You have viewed the preview of this module (first 2 pages).
To access the full content, enter your access code or request access.

6 pages remaining Personal link · Valid 24h