If I had to learn AI and automation from scratch today, this is exactly how I'd do it.
Three months. Zero to actually shipping things.
The reason most people fail with AI is they learn it in the wrong order. They jump into RAG and vector databases on day one, get lost, and quit by week three. Or they sit in ChatGPT for six months and never connect it to anything that matters.
Here's the order I'd actually follow.
I write deep-dives like this when I find a topic that's worth the long version. Want the next one?
Step 1: Foundations
First, the words.
What is AI? What is an LLM? What is a prompt? What is RAG? What are vectors and embeddings?
These sound complicated until somebody explains them in plain English. They aren't complicated. Spend a few hours here. Not weeks. Just enough that you can follow a conversation and read documentation without getting lost.
You don't need a "Complete AI Course" yet. You don't know what to look for.
Step 2: Automation
AI is the brain. Automation is the legs.
AI can think, but it can't actually do anything on its own. It needs a way to log into systems, move data between tools, trigger actions, and connect platforms. Automation is what turns AI from smart text into a machine that runs your business.
This is where most people stop. They master prompts and have nothing to plug them into. So they live in a ChatGPT tab, copy-pasting answers back into their actual work.
The tool I teach: n8n. Visual, free if you self-host, and it does everything you actually need.
Once you can build a workflow that fires on a trigger, calls an AI model, and drops the result somewhere useful, you've crossed the line that 80% of "AI learners" never cross.
Step 3: Build something small and real
Pick a real problem. Not a tutorial problem. Something boring in your actual life or business.
- Read your inbox and triage emails by importance
- Pull action items out of Zoom transcripts
- Watch a Google Sheet for new rows and trigger downstream actions
- Connect an AI agent to your CRM and have it write personalized follow-ups
- Build an internal assistant that answers questions about your company
Your first system will be ugly. That is the point. The first version teaches you that AI plus automation can plug into almost anything. Once you see that, everything else becomes plumbing.
Step 4: RAG
Now you're ready for RAG. Not before.
RAG (retrieval-augmented generation) is how you customize AI with your own documents, your notes, your PDFs, your processes. It's how you build accurate systems instead of chatbots that make stuff up.
Most courses teach RAG first because it sounds impressive. I teach it fourth because it only makes sense once you've already built something that needs it.
You'll know you need RAG the day you build a system that has to answer questions about your specific data, and a regular AI call keeps making things up.
Step 5: Agents, Claude Code, MCP
This is where AI starts feeling like a junior employee instead of a toy.
Multi-agent systems. Instead of one giant prompt that tries to do everything, you split work across specialized agents. A research agent, a writing agent, a fact-checking agent. They pass work to each other.
Claude Code. Anthropic's coding agent. The cleanest example of an agent that does real work. It reads your codebase, edits files, runs commands, and stops to ask when it matters. Once you understand Claude Code, you understand what a production AI agent actually looks like.
Human in the loop. At critical decision points, the system pauses and asks you before continuing. You want this anywhere being wrong is expensive. Sending an email to a customer. Posting publicly. Spending money.
MCP (Model Context Protocol). The open standard for giving agents access to your tools and data. It's what lets your agent talk to your codebase, your database, your APIs without you rewiring everything from scratch.
Most people will never get to this step. The ones who do can offload 60-80% of their actual work.
Step 6: Apply it ruthlessly
Once you can build all of the above, you can automate almost anything.
Pick the most boring repetitive thing in your work. Automate it. Then pick the next one. Then the next.
People who do this for 6 to 12 months become quietly unstoppable. They have offloaded the part of their job that doesn't need a human, and they have time to do the part that does.
Your first 30 days: what to actually do
Reading a roadmap is not the same as walking one. If you're starting from zero today, here's a concrete first month.
Week 1 — Foundations + your first workflow
- Days 1-2. Read Anthropic's docs intro section. Play with Claude.ai for an afternoon. Watch one explainer on what an LLM actually is. Stop. Don't binge content.
- Days 3-4. Sign up for n8n cloud or install it locally. Build the "Hello World" workflow. Then build one that calls Claude and writes the result to a Google Sheet.
- Days 5-7. Pick the smallest annoying problem in your life. Automate it badly. Ship it badly. Don't optimize.
Week 2 — First real system
- Days 8-10. Pick one of these and build it end-to-end: email triage, lead enrichment, inbox summarizer, CRM follow-up writer.
- Days 11-14. Break it. Fix it. Break it again. Every working system has been broken three times.
Week 3 — Make it actually useful
- Days 15-17. Connect it to a tool you use daily (Slack, Gmail, Notion, your CRM). Add error handling. Send yourself a notification when it runs.
- Days 18-21. Use it for real for a week. Track how much time it actually saved. If it's zero or negative, kill it and pick a better problem.
Week 4 — Step into RAG
- Days 22-25. Upload your most-referenced documents into Claude Projects. No code. Just chat with your own data.
- Days 26-28. When Claude Projects hits a limit, set up a real RAG pipeline. Start with 50 documents, not 5,000.
- Days 29-30. Look back at what you built. Pick the next 30-day target.
By day 30, you've shipped something real, broken it, fixed it, and built your first RAG pipeline. That's further than 95% of people who say "I'm learning AI" ever get.
The 7 courses I built for this exact path
I teach this inside the Glitch community. Seven courses, every lesson built so a complete beginner can follow it, and the order matches the 6-step roadmap above.
So if you're starting from scratch in 2026
Pick the foundations first. Don't skip them. Don't jump to RAG. Build something tiny and real in your first two weeks. Connect it to one tool you use every day. Then keep going.
The people who become genuinely good at AI in 2026 aren't the ones who read every paper. They're the ones who shipped something ugly in week two and kept shipping for a year.
If you do that, by this time next year you'll know more about real-world AI than 95% of the people calling themselves AI consultants on LinkedIn. That's not a flex. That's the actual bar in 2026.