The fastest way to learn AI is to use AI tools on your real work every day while you build understanding in small, steady steps. This 90-day roadmap has two tracks, one for non-technical people and one for developers, with free courses, weekly practice tasks and a final project, so you finish with skills you can show.
- Learn by using: spend at least 20 minutes a day applying an AI assistant to real tasks from your job or studies.
- Weeks 1 to 4 build foundations (how language models work, prompting, limits); weeks 5 to 8 apply AI to your work; weeks 9 to 12 are a project.
- Non-experts do not need math or code; Elements of AI's first part requires neither. Developers should learn the API, retrieval and agents before deep learning theory.
- Free resources are enough: Elements of AI, Google's Machine Learning Crash Course, fast.ai, Hugging Face courses and Karpathy's Zero to Hero.
- Finish with one project you can show, plus a short write-up of what worked and what did not.
How should you learn AI in 2026?
Learn AI by using it on real work every day and adding understanding in small steps. Most people stall because they start with theory (linear algebra, neural network math) and never connect it to anything they do. The opposite mistake is collecting tool demos without learning why outputs go wrong.
A balanced plan does three things in parallel:
- Use: apply an AI assistant to your own tasks daily, so skills build on real problems.
- Understand: learn enough about how models work to predict when they fail.
- Build: finish one project that you can show to someone else.
First decide which of two goals you have, because they lead to different roadmaps:
| Goal | You are | You need to learn | You can skip |
|---|---|---|---|
| Use AI well at work | Marketer, manager, analyst, student, founder | Prompting, tool choice, verification, privacy, automation | Code, math, model training |
| Build with AI | Developer or technical person | APIs, retrieval, agents, evaluation, coding agents | Training models from scratch (for now) |
Research careers in machine learning are a third path with years of math and statistics ahead. This roadmap covers the first two, which is what most people mean when they ask how to learn AI.
What do you need before you start?
You need a computer, one AI assistant account and about 30 to 60 minutes a day. No degree, math background or paid course is required.
Set up before day one:
- One assistant. ChatGPT, Claude or Gemini on a free tier is enough for weeks 1 and 2. If you use it heavily, a paid plan around $20 a month removes most limits. The ChatGPT vs Claude vs Gemini comparison helps you pick.
- A learning log. A single document where you paste prompts that worked, outputs that failed and why. This log becomes your personal prompt library.
- A list of ten real tasks from your work or studies that take time: summarizing reports, drafting emails, analyzing a spreadsheet, writing code, studying for an exam.
- Developers only: Python 3 installed, a code editor, a GitHub account and an API key from one model provider.
Weeks 1 to 4: what are the AI foundations to learn?
The first month builds a working mental model of language models and the core skill of prompting. By the end, you should be able to explain why a model got something wrong and rewrite a prompt to fix it.
| Week | Learn | Practice |
|---|---|---|
| 1 | What a large language model is: tokens, training, context windows | Use an assistant for 3 of your 10 tasks; log results |
| 2 | Prompting: context, examples, format, iteration | Rewrite your 3 worst prompts; compare outputs |
| 3 | Limits: hallucinations, outdated knowledge, bias | Fact-check 5 answers against sources; log errors |
| 4 | Privacy and responsible use; files, projects and custom instructions | Set up a project or custom instructions for your main use case |
Resources for this month:
- Non-experts: Elements of AI, a free course from the University of Helsinki and MinnaLearn. Part 1, Introduction to AI, needs no programming or complicated math.
- Everyone: the plain-English guide to what a large language model is and the practical prompt engineering guide.
- Developers: the "Intro to Large Language Models" module in Google's Machine Learning Crash Course, which goes from tokens to Transformers, plus the chapters on embeddings and neural networks.
The single most useful habit this month is noticing failure patterns. When an answer is wrong, write down why: missing context, a vague request, a question about recent events, a calculation done in its head. Most "AI is useless" conclusions come from a handful of fixable patterns. The post on AI hallucinations covers the main ones.
Ask the assistant to teach you. "Explain context windows to me with an analogy, then quiz me with three questions and correct my answers" turns any assistant into a tutor.
Weeks 5 to 8 for non-experts: how do you apply AI to your work?
In the second month, non-technical learners turn one-off prompts into repeatable workflows. The goal is a handful of tasks where AI reliably saves you time, with a check step you trust.
| Week | Focus | Practice |
|---|---|---|
| 5 | Writing and editing | Build a voice guide and a reusable drafting prompt |
| 6 | Research and documents | Use a research mode on a real question; verify every key claim |
| 7 | Data and spreadsheets | Analyze a real CSV: summaries, charts, formulas; check the math |
| 8 | Automation | Connect two apps with a no-code tool so one task runs without you |
Concrete exercises:
- Week 5: take three documents you wrote recently, ask the assistant to describe your style, edit that description, and use it as the start of every writing prompt.
- Week 6: pick a question from your job that needs ten sources. Use a deep research feature, then open every source behind the main conclusions and mark which claims held up.
- Week 7: upload a spreadsheet and ask for the three most important patterns. Ask for the formulas or code behind each number and recompute one by hand.
- Week 8: automate one weekly task with Zapier, Make or n8n, such as summarizing a form's responses into a weekly email.
Weeks 5 to 8 for developers: how do you build with AI?
Developers spend the second month building small programs on top of model APIs, then working with coding agents. You learn more from one small working app than from three courses on theory.
| Week | Focus | Build |
|---|---|---|
| 5 | Model APIs: messages, system prompts, streaming, structured output | A command-line tool that turns notes into JSON |
| 6 | Retrieval (RAG) and embeddings | A Q&A bot over a folder of your own documents |
| 7 | Tool use and agents | An agent that calls two tools (search, a calculator, your own API) |
| 8 | Coding agents and evaluation | Ship a feature in a real repo with a coding agent; write tests for your week 6 bot |
Resources:
- The Hugging Face LLM Course, free and covering Transformers, Datasets and Tokenizers. Hugging Face also offers an Agents Course on its learning hub.
- Anthropic's courses, including Building with the Claude API, Introduction to Model Context Protocol and Claude Code 101.
- Step-by-step guides on this blog: the retrieval augmented generation walkthrough for week 6.
Week 8 is where you see how much of software work agents can do. Give a coding agent like Claude Code or Codex a real, well-scoped ticket, review its pull request like a colleague's, and note where it needed correction. Writing clear tasks for an agent is a skill you will keep using.
Weeks 9 to 12: what project should you build?
Spend the last month on one project that solves a real problem for you or someone you know. A finished small project teaches more than an ambitious unfinished one, and it gives you something to show.
Project ideas by track:
Non-experts
- A documented AI workflow for a recurring task at work, with before-and-after time measurements.
- A custom GPT, Claude project or Gemini Gem loaded with your team's documents and instructions, tested on 20 real questions.
- An automation that turns incoming requests (forms, emails) into sorted, summarized tasks.
Developers
- A RAG assistant over a real document set with an evaluation file of 30 questions and a measured accuracy.
- A small web app built mostly by a coding agent, deployed, with tests and a README.
- An agent that runs a weekly chore end to end: collects data, analyzes it, posts a report.
Run the project in four steps, one per week:
- Week 9: define the problem, the user and what "works" means in one paragraph. Build the simplest version.
- Week 10: test with real inputs and log every failure.
- Week 11: fix the top three failure types. Add checks (tests, validation, human review).
- Week 12: write a one-page summary: problem, approach, results, what you would do next. Share it with your team or publish it.
If your project involves a coding agent running long tasks, you will notice they stop when your laptop sleeps. Orkio is one option: it runs Claude Code on an always-on cloud box with your own Claude account, so you can start a task and check the result from your phone.
Which free AI courses are worth your time?
A small set of free resources covers everything in this roadmap. Pick one per stage instead of starting five at once.
| Resource | Best for | Prerequisites | Cost |
|---|---|---|---|
| Elements of AI | Non-technical foundations | None for Part 1; basic Python for Part 2 | Free |
| Google ML Crash Course | Core ML concepts through intro to LLMs | Some programming and basic math | Free |
| fast.ai Practical Deep Learning | Hands-on deep learning | About a year of coding, high school math | Free |
| Hugging Face LLM Course | Working with open models | Python | Free |
| Karpathy's Zero to Hero | Building a GPT from scratch | Solid Python, intro calculus | Free videos on YouTube |
| Anthropic courses | Claude, the API, MCP, AI fluency | Varies by course | Account required |
The deeper courses are optional for both tracks. Take fast.ai (9 lessons of about 90 minutes in Part 1) or Karpathy's series after day 90 if you want to understand training, not before. Karpathy's series builds a small autograd engine, a character-level language model and finally a GPT model and tokenizer, which is the clearest way to see what happens inside the models you have been using.
What should you avoid when learning AI?
Avoid the three traps that stall most learners: too much theory too early, too many tools, and no verification habit.
- Theory first. Starting with linear algebra before you have used a model for a real task is the most common reason people quit. Learn theory when a practical question makes you curious.
- Tool hopping. New AI tools launch weekly. Pick one assistant (and, for developers, one coding agent) and learn it deeply for 90 days.
- Paid certificates. For most roles, a project you can demonstrate counts for more than a certificate. Spend money on a paid assistant plan before a paid course.
- Trusting outputs. Every learner needs a verification habit: check facts, rerun calculations, run the code and the tests.
- Letting AI do the learning. If you are learning to program, asking the agent for answers without understanding them skips the learning. The guide to learning to code with AI has a weekly plan that keeps you doing the thinking.
Your first week checklist
Start today with these steps and you will have completed week 1 by Sunday.
- Pick your track: use AI at work, or build with AI.
- Create one assistant account and a learning log document.
- Write down ten real tasks you could try with AI.
- Try three of them and log what worked and what did not.
- Start Elements of AI Part 1 (non-experts) or the LLM module of Google's crash course (developers).
- Block 30 minutes a day in your calendar for the next 12 weeks.
- Decide on a rough idea for your week 9 project, even if it changes later.
Frequently asked questions
Can I learn AI without coding?
Yes. Most people who use AI at work never write code: they learn to prompt well, choose the right tool, check outputs and build simple automations. Elements of AI's introductory part needs no programming or complicated math. Learn to code only if you want to build AI products or customize systems.
How long does it take to learn AI?
You can become a confident, effective AI user in about a month of daily practice. Building AI-powered apps takes a developer roughly two to three months of steady work. Training your own models and doing research takes years of math, statistics and engineering.
Do I need math to learn AI?
Not to use AI or to build apps on top of AI models through an API. You need math, mainly linear algebra, calculus and probability, if you want to understand how neural networks are trained or work in machine learning research. fast.ai, for example, asks only for high school math.
What is the best free AI course for beginners?
For non-technical beginners, Elements of AI from the University of Helsinki and MinnaLearn is free and needs no programming for its first part. For people with some Python, Google's Machine Learning Crash Course covers core concepts up to an introduction to large language models. For coders, fast.ai's Practical Deep Learning is free and project-based.
Should I learn machine learning or generative AI first?
Start with generative AI if your goal is to use AI at work or build apps, because that is where the practical tools are today. Learn classical machine learning later if you move toward data science or model training. The two share concepts like training data, evaluation and overfitting, so neither is wasted.
