00Workshop material · Making sense of AI

From chat boxto colleague.

What’s actually behind ChatGPT, agents and MCP? This page breaks the AI landscape down into eight levels, from a simple chat to an agent that gets on with the work by itself. Plus a glossary and buzzword bingo for the workshop.

Go to the levels

01How we got here

Seventy years of run-up, three years of leap.

  1. 1956The term

    At a conference in Dartmouth, the phrase “artificial intelligence” is used for the first time. For decades, AI mostly means rules, written by hand.

  2. 1990s onMachine learning

    Instead of following rules, software learns from examples. That’s how spam filters, product recommendations and fraud detection come about.

  3. 2012Deep learning

    Large neural networks suddenly recognise images and speech reliably. The basis for voice assistants and translation services.

  4. 2017Transformer

    Researchers at Google publish the architecture that almost every language model is built on today. The “T” in ChatGPT.

  5. 2022ChatGPT

    AI goes mainstream: a chat window anyone can type into. Within about two months, more than 100 million people are using it.

  6. 2024 onAgents

    AI stops just answering and starts acting: searching, calculating, writing into other software. Late 2024 brings MCP, a shared standard for connectors.

  7. TodayDivision of labour

    Agents run in the background, on schedules and in teams. The question is no longer whether AI can do something, but how to plug it in safely.

02The eight levels

Every level gives the AI a little more reach.

Most people know level 1. The real payoff starts at level 3, when the AI stops just writing and starts doing. Each level builds on the one before.

Level 1

Chat

You ask, the model answers.

A language model (LLM) has learned how language works from vast amounts of text and writes the most likely continuation. It knows a lot, but only up to its training cut-off, and it knows nothing about you or your business.

How it fits together: You → Language model → Answer
Example
Wording a polite rejection, shortening a long text, thinking an idea through.
Today, for instance with
ChatGPT, Claude, Gemini, Microsoft Copilot, Le Chat
Where it stops
The answer always sounds convincing, even when it’s wrong. And anything the model should know, you have to paste in yourself.
Matching package
AI Kickstart

Level 2

Chat with your knowledge

The AI works with your documents.

You give the model your own documents and standing instructions: price list, guidelines, tone of voice. With large amounts of material, the system first looks up the relevant passages and only hands those to the model. That’s called RAG.

How it fits together: You → Your documents → Language model → Answer
Example
An assistant that answers customer questions using your actual terms and product data.
Today, for instance with
Projects in ChatGPT and Claude, custom GPTs, NotebookLM
Where it stops
The AI can read, but it can’t do anything. Copying, sending and filing is still on you.
Matching package
AI Kickstart

Level 3

Agent with tools

The AI works in rounds until the goal is reached.

An agent is a language model that’s allowed to use tools: search the web, run code, create files. It plans, acts, checks the result and keeps going until the job is done. An answer turns into a piece of finished work.

How it fits together: Goal → Agent → Search · code · files → Result
Example
“Compare the five biggest point-of-sale vendors and put it in a table.” The agent does the research and hands you the finished file.
Today, for instance with
ChatGPT agent, Claude with web search and code, Gemini Deep Research, Perplexity
Where it stops
The tools are generic. Your inbox, your calendar and your customer data stay out of reach.
Matching package
AI Kickstart

Level 4

Agent with connectors

The AI reaches into your systems.

Connectors give the agent access to the software you work with: email, calendar, file storage, CRM, shop. It can read there and, if you let it, write too. The shared standard for this is called MCP.

How it fits together: You → Agent → Connectors → Mail · calendar · CRM
Example
“Which quotes have gone unanswered for two weeks? Draft a follow-up email for each.”
Today, for instance with
Claude or ChatGPT with connectors to Gmail, Outlook, Google Drive, Notion, HubSpot
Where it stops
The agent only works while you’re chatting with it. Without you, nothing happens.
Matching package
AI Build

Level 5

Scheduled agent

The AI works while you do something else.

The agent starts by itself: every morning at seven, on every new order, on every incoming invoice. It handles the routine and only puts in front of you what needs a decision. That’s called human in the loop.

How it fits together: Time or event → Agent → Your systems → Your sign-off
Example
Every morning a sorted inbox is waiting: draft replies to standard requests, urgent items at the top.
Today, for instance with
Scheduled tasks in ChatGPT and Claude, n8n, Make, Zapier with AI steps
Where it stops
You’re limited to what off-the-shelf platforms offer. If they don’t have it, you can’t do it.
Matching package
AI Build

Level 6

Your own app

The AI lives inside your own software.

Instead of someone else’s chat window, you get a tool shaped around exactly how you work. The AI is plugged in through an API, and custom connectors link it to systems nobody has built one for yet, such as your industry software.

How it fits together: Your app → Language model → Custom connectors → Your data
Example
A quote calculator that understands what the customer wants, prices it with your rates and drops the finished PDF straight into your inventory system.
Today, for instance with
Claude or OpenAI API, your own MCP server, a small web app
Where it stops
A single agent can do a lot, but not everything equally well. Long, many-sided tasks are too much for one.
Matching package
AI Build

Level 7

Agent team

Several agents share the work.

A lead agent breaks the task down and hands parts to specialists: one researches, one writes, one checks. Like a small team, each sticks to their own job. The result is better because someone reviews it.

How it fits together: Goal → Lead agent → Research · writing · review → Result
Example
One agent gathers relevant public tenders, one drafts the bid, one checks deadlines and required details.
Today, for instance with
Claude Code with subagents, OpenAI Agents SDK, LangGraph, CrewAI
Where it stops
So far these are individual tasks. Owning an entire process, exceptions and all, is the next level.
Matching package
AI Partnership

Level 8Outlook

Digital colleague

The AI owns a whole process.

Agents with memory take on a process from start to finish, remember what worked and learn from your feedback. Where there’s no API, they operate software the way a person does, with mouse and keyboard. You set goals and limits and decide the exceptions.

How it fits together: Goal → Agent with memory → Every system → You, for exceptions
Example
The entire order flow: understand the enquiry, write the quote, follow up, send the invoice. You only look at the cases that don’t fit the pattern.
Today, for instance with
Partly possible today, for example with computer use and agents that have memory. The full picture is still an outlook.
Where it stops
Trust. The more the AI decides on its own, the more you need clear rules, logs and oversight.
Matching package
AI Partnership

03What makes an agent

A chat answers once. An agent keeps going until it’s right.

That’s the whole trick: the model gets the results of its own actions back and decides what to do next. The loop runs until the task is done or the agent needs to ask you something.

  1. 1. Understand the goal
  2. 2. Make a plan
  3. 3. Use a tool
  4. 4. Check the result

04Around the edges

Nine topics that come up at every level.

  • 01

    Memory

    Normally a model forgets everything once the conversation ends. With memory, it keeps track of preferences, decisions and work in progress.

  • 02

    Multimodal

    Today’s models understand more than text: images, PDFs, speech and sometimes video. A photo of the delivery note will do.

  • 03

    Models that think

    Reasoning models work through a problem step by step before answering. Slower and pricier, but much better at tricky tasks.

  • 04

    Local models

    Open models such as Llama, Mistral or Qwen also run on your own machines. Your data never leaves the building.

  • 05

    Privacy and the EU AI Act

    GDPR and contracts govern which data may go into which model. The EU AI Act sorts AI applications by risk and, depending on the level, requires things like transparency.

  • 06

    Guardrails

    They define what an agent may do: read, yes; delete, no; move money, never without sign-off. The more reach, the more they matter.

  • 07

    Cost

    Billing is per token, meaning per chunk of text. A chat costs a fraction of a cent; an agent working for a long time costs noticeably more.

  • 08

    Measuring quality

    Evals test the AI against real examples to see whether it gets things right, before you hand it a task for good.

  • 09

    Security

    In prompt injection, hidden instructions in an email or web page try to hijack the agent. A good agent treats outside content as data, not as orders.

05Buzzword bingo

Five in a row and you shout bingo.

For the workshop: everyone gets a card. When a term comes up, mark it. Tap a square to see what the term means.

AI

Tap a term.

06Glossary

The terms, each explained in brief.

AI (artificial intelligence)
Umbrella term for software that handles tasks that used to need human thinking. In everyday use it now usually means generative AI.
Generative AI
AI that creates something new: text, images, code, music. As opposed to AI that only sorts or recognises things.
LLM (large language model)
A neural network that has learned to continue language from enormous amounts of text. The engine behind ChatGPT, Claude and Gemini.
Prompt
What you give the AI: a question, a task, some context. The clearer the prompt, the better the result.
System prompt
Standing instructions that apply before every conversation, such as role, tone and rules. The person chatting usually doesn’t see them.
Token
The unit models read and write in: a chunk of a word, on average shorter than a word. It’s also what you’re billed by.
Context window
How much text the model can keep in view at once, measured in tokens. Modern models can hold whole books. Whatever doesn’t fit doesn’t exist for the model.
Training cut-off
The date up to which a model has seen data. It only knows about anything later if it’s allowed to look it up.
Hallucination
The model makes up something plausible but wrong: a source, a number, a court ruling. Your own data, citations and checking help.
Temperature
A dial for randomness. Low means reliable and consistent; high means creative and surprising.
Parameters
The learned settings inside a model, often billions of them. More parameters usually means more capability, but also more computing power.
Training and inference
Training is where a model is made; it takes months and costs millions. Inference is using it afterwards, meaning every single answer.
Reasoning model
A model that thinks step by step internally before answering. Stronger at maths, logic and planning, but slower.
Multimodal
A model that understands several kinds of input: text, images, audio and sometimes video.
Embedding
A piece of text turned into a row of numbers that captures its meaning. Similar meaning, similar numbers. That’s how AI finds the right passage even without the exact word.
Vector database
A store for embeddings. Given a question, it finds the passages closest in meaning within a fraction of a second.
RAG (retrieval-augmented generation)
Before answering, the system looks up relevant passages in your documents and hands them to the model. It answers from your knowledge and can cite sources, without retraining.
Fine-tuning
Training a finished model further on your own examples so it masters a style or a specialised task. Rarely needed; good instructions and RAG are often enough.
Agent
A language model that uses tools on its own and works in rounds until it reaches a goal.
Tool use (function calling)
The model decides to call a tool, such as a search or a calculator, and gets the result back.
API
Application programming interface. The way programs talk to each other, without a screen or mouse.
MCP (Model Context Protocol)
An open standard that lets AI applications plug into other systems. Think USB for AI: build it once, use it anywhere.
Connector
The link between the AI and a specific program, such as Gmail or your CRM. Often built on MCP.
Workflow
A fixed sequence of predefined steps in which AI handles individual tasks. More predictable than an agent, but less flexible.
Multi-agent system
Several specialised agents working together, steered by a lead agent. The steering is called orchestration.
Human in the loop
A person checks or signs off before anything important happens. The AI prepares, you decide.
Guardrails
Technical and content limits on what an AI may do and say.
Prompt injection
An attack in which hidden instructions in text, emails or web pages try to get the AI to do something you don’t want.
Evals
Tests for AI: a set of real examples with expected results, used to measure whether the system is good enough.
Open-weight model
A model you can download freely, such as Llama, Mistral or Qwen. It also runs on your own hardware.
Copilot
AI riding along inside a program, such as Word, Excel or a code editor. Microsoft turned the word into a product name.
Computer use
An agent operates a computer the way a person does: it sees the screen, clicks and types. Useful where there’s no API.
Vibe coding
Building software by describing to the AI what you want instead of writing code yourself. Fast for prototypes, risky without review.
Memory
The ability to remember things beyond a single conversation, such as preferences or earlier decisions.
EU AI Act
EU law that classifies AI applications by risk. The higher the risk, the stricter the obligations, up to an outright ban.
AGI (artificial general intelligence)
A hypothetical AI that thinks as broadly and flexibly as a person. A popular buzzword with no agreed definition.

07AI Kickstart

Which level are you on?

Most businesses use level 1 today and suspect there’s more to it. In the Kickstart, we work out together which level pays off first for you.