Search for artificial intelligence courses and you’ll get millions of results, a hefty share of them promising to make you job-ready in six weeks. Some of that is marketing. A smaller portion is genuinely true, and the gap between the two comes down to one question: does the course end with you having built something you can show someone?
I’ve spent the past few years comparing syllabi, sitting in on trial classes and talking to people who actually finished what they started. Here’s what separates a course worth your evenings from one that leaves you holding a certificate and a vague feeling of having learned something.
What an AI course covers now, and why no two are alike
Underneath the label, artificial intelligence courses split into roughly three job tracks, and they barely overlap.
- Data and analytics. Python, pandas, SQL, statistics, scikit-learn. The output is a model or dashboard that answers a business question. Most entry-level “AI” jobs sit here.
- Machine learning engineering. PyTorch or TensorFlow, training runs, evaluation, deployment, MLOps. You’re the person who keeps a model alive in production at 2am.
- AI application building. APIs, retrieval-augmented generation, vector databases, prompt design, agents. This is currently the fastest route into paid AI work, because you can ship something useful without training a model from scratch.
A “complete AI course” that claims to cover all three in eight weeks is covering none of them properly. The syllabus tells you which track a course really sits on. Count the hours spent on Python and statistics, then count the hours on transformers. That ratio is the honest description of the programme.
Four formats, four very different bargains
Before comparing curricula, work out which constraint you’re actually trying to solve: time, money, or accountability.
- Self-paced online courses on Coursera, edX, DeepLearning.AI and similar platforms. Free to around £40 a month. Flexibility is the selling point and also the risk. Completion rates for these sit somewhere between 5% and 15%, which is not a knock on the teaching. It’s what happens when nobody is waiting for you.
- Cohort bootcamps. Twelve to sixteen weeks full-time, typically £9,000 to £16,000, usually with career support bolted on. Expensive, and effective mainly because a room full of people expects you to show up.
- Mentor-led part-time programmes. Eight to twenty weeks, one or two evenings, with a working practitioner reviewing your code. The feedback loop is the product.
- University certificates. £2,000 to £12,000. Strongest for people who need a credential that clears an HR filter or a visa requirement.
Free courses are better than their reputation
If money is your blocker, don’t treat free as a consolation prize. The free AI courses on Alison cover Python, statistics and machine learning fundamentals at a level that genuinely prepares you for a first project, and the certificates cost a token fee rather than a tuition bill.
The catch isn’t quality. It’s structure. Nobody will notice if you stop in week three, so you need to manufacture your own deadline: a public commitment, a study partner, or a project with a date attached.
How to vet a course in twenty minutes
Most bad courses give themselves away before you hand over a card. Read for these signals:
- Look at the final project, not the welcome video. If the syllabus stops at “introduction to neural networks” and never asks you to deploy anything, it’s an orientation, not training.
- Check the dates. Anything still presenting GPT-3 or TensorFlow 1.x as the current state of the art has been sitting untouched for years.
- Find two graduates on LinkedIn. Look at what they built, not the certificate badge. If you can’t find any graduates, that tells you something too.
- Ask who teaches it. A researcher and a working engineer explain the same concept completely differently, and only one of them will tell you what breaks in production.
- Read the refund and access terms. Some platforms cut your access the day the subscription ends, which matters if you want to revisit material during interviews.
If you’d rather work from a repeatable framework than a gut feeling, this six-step method for choosing an AI programme walks through cost per contact hour, instructor background and outcome data in a sensible order.
A twelve-week shape that works
Course length matters less than sequence. A structure that holds up across most formats looks like this:
- Weeks 1-3: Python, pandas, and just enough statistics to read a confusion matrix without guessing.
- Weeks 4-6: scikit-learn on a dataset you actually care about. Your city’s bike-share data, your local air quality readings, your own Spotify history. Interest is what keeps you debugging at 11pm.
- Weeks 7-9: either a PyTorch model or an API-based build with retrieval and a small vector store. Pick one, not both.
- Weeks 10-12: ship it. Deploy to Hugging Face Spaces or Streamlit, write a README that explains your decisions, and put the link somewhere people will see it.
That final step is where most self-taught learners stall, which is why this walkthrough from signup to shipped project is worth reading before you enrol rather than after.
If you’re a designer, or you want a full-time switch
Two routes come up repeatedly in conversations with people who made the move successfully.
Designers rarely need to train models. They need to understand what AI systems can and can’t do, then design around those limits. Mentor-led programmes built for that audience, like this practical path for designers from Designlab, spend their time on capability, ethics and interface decisions rather than backpropagation.
Career-changers who want intensity usually end up in a bootcamp. BrainStation’s AI courses are a reasonable benchmark for what a full-time programme covers and costs, and just as usefully, for who they don’t suit. Read the fit section carefully. A £12,000 course aimed at people with three years of SQL behind them will not rescue a standing start.
What employers actually look at
Hiring managers spend about ninety seconds on a junior AI application. In that window they’ll open your GitHub, check whether your last commit was three months or three days ago, and click one project link.
What impresses them is unglamorous. A model that solves something small and boring, with a written explanation of where it fails and why. A commit history that shows you kept going after the exciting part ended. Someone who can say “the false negative rate was too high, so I changed the threshold and here’s the trade-off I accepted.”
Titles matter less than you’d expect. Many first roles are advertised as data analyst, junior data scientist or AI product manager rather than AI engineer. Search those terms alongside the obvious one when you’re scanning job boards.
Your first week, in practice
Pick one course. One. Block ninety minutes on four evenings this week, and use the first session purely to get your environment working, because that’s the step that quietly kills momentum for beginners. Aim to finish a tiny project by day ten: load a CSV, train something basic, print the accuracy, and write two sentences about whether the number makes sense.
If two weeks pass and you’ve done nothing, the problem is almost never the subject. It’s the format. Self-paced becomes a cohort programme, evenings become a Saturday morning, solo becomes a study partner. Change the container, keep the goal, and let the shipped project be the thing you’re judged on rather than the certificate hanging on the wall.