An AI roadmap around a day job — eleven courses, four stages

The market is loud about AI and quiet about where to start. Most published roadmaps are either a research syllabus that assumes linear algebra, or a list of tools that will be obsolete before you finish it.

What follows is neither. It is a sequence of short courses, ordered so that each one is usable without the ones after it, aimed at somebody with a job and a few hours a week.

Four rules before the first course

One: apply while you watch, not after. A course finished without a keyboard open produces recognition, not ability. You will recognise the concept in an interview and be unable to build it.

Two: keep the order. The sequence is not decoration. Stage three assumes stage two, and skipping to the agent courses because they sound interesting produces somebody who can wire a framework together and cannot debug it.

Three: build your own thing. After every stage, build something nobody assigned you. It does not have to work well. It has to be yours, because that is the only version you will remember.

Four: the course is the beginning. A completed course is a licence to start reading documentation, not a qualification.

Stage one — the ground floor

Prompting properly, and enough Python to be dangerous.

• ChatGPT Prompt Engineering for Developers — the shortest useful thing on this page. It replaces guessing with method.
• AI Python for Beginners — Python taught for this purpose specifically, rather than a general programming course you would abandon in week three.
• Collaborative Writing and Coding with OpenAI Canvas — working with a model rather than issuing instructions at it.

By the end of this stage you can automate something small in your own job. That is the checkpoint. If you cannot, do not move on.

Stage two — from a chat window to a system

The difference between using a model and shipping one.

• Building Systems with the ChatGPT API — chaining calls, handling failure, and the fact that a single prompt is almost never the answer.
• LangChain for LLM Application Development — the framework most job adverts name.
• Functions, Tools and Agents with LangChain — giving a model the ability to do things rather than only say them.

This is where the employability starts. A person who has finished stage two can build an internal tool that a company would actually use.

Stage three — retrieval and reasoning

Making a model answer from your data instead of from its memory.

• Building Agentic RAG with LlamaIndex — the single most requested pattern in real business work: answer this question from our documents, and cite where.
• Reasoning with o1 — what changes when the model thinks before it answers, and when that is worth the latency.
• Multi AI Agent Systems with crewAI — several agents with different jobs, and the coordination problem that comes free with them.

Retrieval is the stage that pays. Nearly every corporate AI project is, underneath, a retrieval problem wearing a chatbot.

Stage four — the specialist edge

Where you stop being interchangeable.

• Finetuning Large Language Models — when a model needs to be taught rather than told, and the far more common case where it does not.
• Red Teaming LLM Applications — attacking your own system before somebody else does. Rare, and increasingly asked for by anyone deploying to customers.

The last course is the one most people skip and the one that gets remembered in an interview. Everybody can demonstrate something that works. Very few can explain how theirs breaks.

What eleven courses do not give you

They do not give you a job. They give you the ability to build the thing that gets you one.

The difference between somebody who finishes this list and somebody who does not is rarely the courses. It is the three or four small projects built alongside them — the ones with a real user, even if the user is a colleague in accounts.

Start the first course this week and give it four hours. The whole path is short enough to finish and long enough to change what you are paid.

For what the labour market is actually doing to jobs over the same period, the WEF numbers are here. If you work in business systems and want to know where AI lands in that field specifically, start with what an ERP is and how its cycles produce the data any model would need.