Last spring I made a spreadsheet of eleven AI classes. I finished three. The other eight are still sitting there with a little progress bar at 12%, which is roughly where most people land when they pick a course by reputation instead of by fit.
The best AI classes aren’t the ones with the biggest names or the longest curriculum. They’re the ones matched to something you’re building in the next ninety days. Everything below is a filter for finding those, plus a short list of offerings that survive the filter.
Start with the free classes, because they’re not the warm-up act
The gap between free and paid AI education is much narrower than the price tags suggest. DeepLearning.AI publishes hour-long short courses on prompt engineering, RAG, and agent building. Hugging Face has a full NLP course with runnable notebooks. fast.ai’s Practical Deep Learning for Coders is free and still one of the best-taught technical courses anywhere. Microsoft Learn and Google Cloud Skills Boost both give away learning paths that map directly to their certification exams.
Paid courses earn their money in three specific ways: structure you’d never impose on yourself, feedback on your work, and a deadline. If a course offers none of those, the free version is usually the smarter buy. There’s a longer breakdown of that trade-off in how to tell what an AI course is actually worth paying for, and it’s worth reading before you enter a card number.
Run a 20-minute check before you spend anything
Most bad course purchases are knowable in advance. Open the course page and look for four things:
- The final project. If the syllabus doesn’t tell you what you’ll have built by the end, the course is probably a video library with a quiz at the bottom.
- The last update date. Anything on LLM tooling that hasn’t been touched since 2023 is teaching you a museum piece.
- A public repo or sample lesson. Real instructors publish something you can inspect. Fake ones publish testimonials.
- The instructor’s shipped work. Someone who has deployed a model to production teaches differently than someone who has only explained one.
If a course fails two of those, move on. If you want the full version, there’s a structured six-step method for choosing an AI program that takes you from goal to shortlist without drowning in options.
The AI classes that hold up, by what you’re trying to do
If you want the foundations to be solid
Andrew Ng’s Machine Learning Specialization on Coursera remains the default recommendation for a reason: roughly 60 hours, math kept practical, and labs that force you to write the gradient descent rather than import it. Budget $49–79 per month, so the real cost depends on your pace. Pair it with fast.ai if you learn better by building first and reading theory after.
If you want to build with LLMs this month
DeepLearning.AI’s short courses are 60–120 minutes each and drop you into a working notebook by minute ten. Take three in a row and you’ll have a functioning mental model of the API layer. For chains and agents specifically, LangChain Academy’s free intro course is a reasonable next step, though it assumes you’ve already shipped something small.
If your employer is paying and wants a certificate
Google Cloud Skills Boost and AWS Skill Builder both run on subscriptions in the $29–39 per month range, and their associate-level exams cost $99–125. The material is dry but the labs are real consoles, which is more than most Udemy courses can say.
If accountability is the actual problem
Cohort programs from Maven or similar run $1,000–2,500. That’s a fair price for weekly live sessions and someone reviewing your work. It’s a terrible price if you’re the person who watches replays at 2x speed and never shows up live. Be honest about which one you are.
Test the syllabus against your actual job
A course can be excellent and still useless to you. Before enrolling, write down three tasks you did at work last week that AI could touch — cleaning a messy export, drafting replies to inbound leads, summarising a 40-page contract. Now search the syllabus for those exact tasks. If none of them appear, the course is teaching you a hobby.
This is the core of a five-step test for finding AI programs that fit your real work, and it’s the single highest-leverage thing you can do before paying. Twelve minutes of syllabus reading beats three weeks of enthusiastic but irrelevant coursework.
The project is the class
Here’s what I’ve noticed after watching a lot of people learn this stuff: the course matters less than what you build during it. Someone who takes a mediocre $15 course and ships a working tool that answers questions over their company’s PDFs will out-learn someone who audits the best $2,000 bootcamp and never deploys anything.
Pick the build before you pick the class. A small document Q&A tool, a classifier for support tickets, an agent that drafts weekly reports from a spreadsheet — any of these will force you through the parts of the course that matter and let you skim the parts that don’t. If you’ve never shipped anything, this step-by-step guide to a first real AI project is the right starting point, and it’s shorter than most course intros.
What a schedule that survives actually looks like
Most AI classes fail in week three, not week one. Week one is novel, week two is fine, and week three is the Tuesday you get home at 8pm and the notebook won’t run.
Five hours a week is a realistic load for someone with a job. That’s about one 90-minute session midweek plus a longer Saturday block, and it puts a 40-hour course at eight weeks. Protect the Saturday block. Skip a weekday session instead. Plans that break the long session tend to break entirely, which is the whole premise behind this eight-week plan for learning AI online.
Red flags that separate a $400 mistake from a $40 win
Watch for these in the sales copy. A 60-hour video library with no project. “AI mastery” bootcamps priced at $3,000 with no code review. Certificates that require no submission. Curricula that never mention embeddings, evaluation, or cost-per-token, which are the three things you’ll hit in your first real week of building. Any course claiming to make you an expert in a weekend is selling the feeling of progress, and the feeling fades by Tuesday.
The reverse is also true: a plain-looking, cheap course with a hard project and responsive instructors is almost always the better buy than a glossy one with a completion badge.
Pick one this week and put it on the calendar
Choose a single class. Not three, not a “learning path.” Block the Saturday sessions through the end of next month. Write down the one thing you’ll have built by then, in a sentence specific enough that a friend could check it — “a Slack bot that answers HR policy questions from our handbook” rather than “better AI skills.”
Then start the free intro module tonight, before the enthusiasm settles into a to-do list item. The best AI class in the world is the one you’re still working through in week three, and that’s entirely a function of what you picked and what you cleared space for.