ORBii.Academy
Module M7 · Agentic AI · Architecture, Governance & Controlled Deployment · 2026
Module 7 · 1 day
Agentic AI
in Banking
Architecture · Deployment · Guardrails · Banking Sector
Agentic AI is no longer a promise — it is in production within banking IT departments. This module provides the keys to understanding what an AI agent truly is, how multi-agent systems are architected, why 40% of agentic projects will be abandoned by 2028 (Gartner), and how to establish the essential governance before scaling up.
Learning Objectives
01Understand the architecture of an agentic system — from a single agent to a multi-agent fleet
02Master the 3 orchestration patterns (sequential, parallel, loop) and their banking use cases
03Identify the 4 critical guardrails: HITL, RBAC, sandbox, audit trail — and why they are not optional
04Apply the Anchor → Activate → Secure → Scale approach to prioritize agentic use cases
IT Architects
CIO
IT Managers
Compliance & Risk
Prerequisites: M4 + M5
Pejman Gohari · CDO · Chief AI Officer · ORBii
Advisor Agentic AI BPCE SI 2025–2026 · Hype Cycle + Tech Radar 25+ use cases · Agentic governance framework Executive Committee · Author DUNOD · IESEG
academy.orbii.tech
ORBii.Academy · M7 · Agentic AI · Architecture, Governance & Controlled DeploymentConfidential · 202601
ORBii.Academy
M7 · Agentic AI · 02
Section 1
What an AI agent really is — Beyond the chatbot
"An LLM responds. An agent acts. The difference is far from trivial: an agent can modify a database, send an email, execute code, call an external service — without a human validating each action. That is where governance becomes existential."
— Pejman Gohari · Advisor Agentic AI · BPCE SI 2025-2026 · CDO · Chief AI Officer · ORBii
LLM vs Agent — The fundamental distinction
| Dimension | Classic LLM | AI Agent |
|---|
| Mode of action | Generates text in response to a prompt | Plans and executes sequences of actions |
| Tools | No access to external systems | Calls APIs, databases, tools |
| Memory | Limited to the context window | Short-term + structured long-term memory |
| Autonomy | None — waits for each prompt | Iterates, corrects, delegates to other agents |
| Error risk | Textual hallucination | Hallucination + real action on systems |
| Oversight | Human reads and validates text | HITL mandatory for critical actions |
OPERATIONAL DEFINITION — BPCE SI GLOSSARY 2026
Agentic AI
AI capable of acting autonomously to achieve a business objective — by planning a sequence of actions, using external tools, iterating on its results, and delegating to other agents if necessary.
Agentic system
A structured environment enabling AI agents to operate within a controlled framework — with guardrails, traceability, and defined human oversight.
MCP (Model Context Protocol)
A standard enabling AI agents to access enterprise information systems through secure, standardized, and auditable interfaces.
The 4 components of an AI agent
1
Cognitive core (LLM + Reasoning)
The LLM at the center of the agent: it understands the objective, reasons about the necessary steps, plans actions (planning), and evaluates its own results before continuing (ReAct, CoT).
2
Memory (short-term + long-term)
The context of the current session (short-term) + structured and persistent memory across sessions (long-term). Includes forgetting mechanisms to avoid accumulation of irrelevant context.
3
Tooling & Integration (MCP / APIs)
Catalog of tools the agent can call: databases, business APIs, IT tools, external services. The MCP (Model Context Protocol) standard secures and normalizes these accesses. Each tool has a permission scope controlled by RBAC.
4
Knowledge base (RAG / Knowledge Graph)
The knowledge base the agent draws upon to enrich its responses: documents, procedures, reference data. The quality and classification of this base determines the quality — and security — of the agent's actions.
⚠️
The systemic risk of agentic AI: Unlike an LLM that generates fallible text, an agent can execute real and irreversible actions — modify data, send messages, trigger processes. A hallucination in an agent is not an erroneous text to correct — it is a concrete action to undo.
ORBii.Academy · M7 · Agentic AI · Architecture, Governance & Controlled DeploymentConfidential · 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