Course overview
AI Basics for Everyday UsersAI in the Real World
Part 2 · Beyond the single model

How AI works in the real world.

Part 1 showed how one model thinks. This part zooms out: where AI sits in the wider family of technologies, how it safely connects to your company's data, how it can act on its own, and where it actually runs. Still plain language, still hands-on.

Nesting RAG MCP Agents Edge & Cloud
Section 1 · The map

1The AI family, nested

“AI” gets used for very different things. These aren't rival technologies — each one lives inside the one before it. Click any layer to see what it is and what it adds.

Artificial IntelligenceAI
Machine LearningML
Deep LearningDL
Generative AIGenAI
Large Language ModelLLM
AI

Artificial Intelligence

The whole field of getting computers to do tasks that normally need human thinking — deciding, recognising, predicting, generating.

What it adds

The umbrella term that covers everything else here.

At work

Spam filters, route planners, recommendation feeds and chat assistants.

An LLM isn't competing with “AI” or “ML”. It's a highly specialised engine sitting deep inside the family — a small box inside a much bigger nesting doll.
Section 2 · Connecting to your data

2How AI reaches real company information

On its own, a model only knows general patterns from training — not your files, prices or policies. These are the ways it gets connected to real data, safely.

API

The hard-wired pipe

A fixed connection between two specific software systems. Reliable, but built one link at a time.

RAG

Read first, then answer

The model searches your private documents first and answers using what it actually found.

Vector DB

Search by meaning

A store that finds the right text by meaning, not exact keywords — the engine behind RAG.

MCP

One universal plug

A single standard that lets a model securely connect to many tools and data sources at once.

Pick a route and watch where your question — and your data — actually travels.

You
your question
Cloud AI model
general training
Your documents
vector database
Connected tools
via MCP

Plain prompt

Your question goes straight to the cloud model. It answers from general training only — it can't see any of your company's information.

Sees your private data No access
Answers grounded in real sources Not grounded
Risk of made-up details Higher
RAG doesn't make a model perfect, but grounding answers in your real documents makes it far less likely to invent things. MCP is what lets one model reach many systems through a single secure standard instead of a separate pipe for each.
Section 3 · Tools that act

3From following steps to working on its own

Some tools just run a fixed script. Others can plan, hit a snag, fix it themselves, and keep going — but stop for a human before anything risky. Run the same job both ways and watch the difference.

AI Agent

Narrow & linear

A script that uses a model to finish one fixed task. Fast — but it breaks the moment the task changes.

Agentic AI

Plans & self-corrects

Models that split work into steps, check their own results, and adapt in a loop until the job is done.

Human-in-the-loop

The approval gate

A hard stop that pauses the system for a person to approve before any high-risk action goes ahead.

The job: process an incoming vendor invoiceRead the details, then schedule payment.
⛔ Human approval required

The system is ready to send a wire transfer. By design, it stops here for a person to confirm before any money moves.

More independence means more has to go right on its own — so the riskiest steps are built to pause for a human, not to push ahead.
Section 4 · Where it runs

4Modalities, security and the trade-offs

The same kind of model can be deployed in very different places — and each choice trades cost, speed, control and risk. First the key terms, then a live look at the trade-offs.

Multimodal AI

Text, image & audio together

Handles several input types — words, pictures, sound — in a single pass instead of one at a time.

Diffusion models

The image & video engine

Builds pictures by starting from random noise and refining it step by step into a clear image.

!
Shadow AI

The hidden risk

Staff quietly using public consumer tools for company data — convenient, but a serious leak risk.

Sovereign AI

Kept fully in-house

Running and training models entirely inside your own servers or your country's borders for full control.

Edge AI

Runs on the device

A compact model running directly on a laptop, phone or camera — fast, private, even fully offline.

Choose where a model runs and watch the trade-offs move. There's no “best” — only the right fit for the data.

Setup & resource cost higher = heavierLow
Response speed higher = fasterHigh
Data-leak risk higher = riskierHigh
Independence higher = works aloneLow
The riskiest pattern is Shadow AI: pasting company data into a public chatbot on your own. It carries the highest leak risk with none of the controls the other options give you.

The rule that keeps AI safe at work: match the tool to the data

The more sensitive or high-stakes the task, the more grounding, control and human oversight it needs.

Know where your data goesBefore you paste anything in, check whether it leaves your control.
Ground company answersFor real facts and figures, use tools connected to your own sources.
Keep a human on riskAnything costly or irreversible should pause for a person to approve.