Ask ten people how long it takes to learn AI and you’ll get answers ranging from “a weekend” to “a decade.” Both are honest answers to different questions. The useful version of the question is: learn AI to do what?
A marketing manager who wants to stop burning an hour a day on first drafts is not on the same road as a physics graduate aiming for a research lab. Their timelines differ by roughly a factor of twenty. So here’s the breakdown by destination, with real hour counts, plus the things that reliably shorten or stretch the clock.
Three AI goals, three very different timelines
Goal one: use AI competently at work (2 to 6 weeks)
This is the literacy layer. You learn how large language models behave, why they invent citations, what a context window actually limits, how to write a prompt that produces something usable, and which work data should never be pasted into a chat box. Twenty to thirty hours of deliberate practice gets most people there. That’s three 45-minute sessions a week for six weeks, which fits around a job.
Goal two: build things with AI (3 to 6 months)
Here you’re connecting models to real workflows: calling an API, piping documents into a retrieval pipeline, automating a weekly report, wiring up a small internal tool nobody else wants to maintain. You need basic Python, some comfort with JSON, and the patience to read an error message twice. Budget 120 to 200 hours. People who already script in Excel or write SQL tend to land at the low end of that range.
Goal three: work as an ML engineer or researcher (1 to 3 years)
This path goes through linear algebra, probability and statistics, and the theory behind gradient descent before it reaches anything glamorous. Part-time study at 15 hours a week puts you at roughly 1,500 to 2,500 hours across two years, which is quietly what most bootcamps and master’s programs assume you’ll invest.
Weeks 1 to 6: getting fluent enough to stop feeling behind
The first month is mostly vocabulary and calibration. Tokens. Context windows. Temperature. Hallucination. Retrieval-augmented generation. Fine-tuning versus prompting. Agent. You don’t need to be able to implement any of it, but you do need to know what someone means when they say the model is “grounded” now.
Practical work matters more than reading here. Take one real task you do weekly, such as summarising customer feedback or drafting project updates, and run it through a model five different ways. Compare the outputs. Break the prompt on purpose and watch what happens. That single exercise teaches more about model behaviour in 90 minutes than a three-hour video course on prompt theory. If the jargon is the part that keeps tripping you up, a short AI vocabulary course built around real meeting language can close that gap in an afternoon or two.
Months 2 to 6: the practical builder stage
Once you can get reliable outputs, the next step is making them repeatable. That means a little Python: variables, loops, functions, the requests library, and reading JSON without flinching. Eight hours a week for twelve weeks is about 100 hours, enough to build two or three small projects that work.
Pick projects with a visible finish line. A script that takes a CSV of support tickets and returns a categorised summary is worth more to your learning than a fourth certificate. So is a small tool that drafts replies from your own past emails. The pattern here is unglamorous and effective: automate one annoying task per month, in public, and let colleagues poke holes in it.
What you should not do is spend this phase memorising architecture diagrams. The skills that hold value are the ones that transfer between models and vendors, and an AI essentials training focused on transferable workplace skills is a better use of 20 hours than a deep dive into a tool that will be renamed by spring.
What speeds you up, and what quietly costs you months
- Shipping one small thing every fortnight. Feedback from a real user compresses learning faster than any curriculum.
- Reading documentation over watching tutorials. Vendor docs are dry, but they’re current, and they teach you to find answers independently.
- Tutorial hopping. Restarting a fifth “intro to machine learning” course in month four is the single most common reason people report a two-year timeline for what should take six months.
- Math perfectionism. Most builders need intuition about vectors and probabilities, not a proof-based linear algebra sequence. Save the rigour for the research track.
- Chasing every model release. The underlying concepts barely moved in three years. The version numbers change weekly.
- Learning alone. A study partner, a Slack channel, or a manager who asks what you built this week does more for consistency than motivation ever will.
A schedule that survives a full-time job
Three weekday evenings of 60 to 90 minutes plus one longer Saturday morning block comes to roughly five hours a week, or about 260 hours a year. That’s enough to finish the literacy stage comfortably and get most of the way through the builder stage, provided the hours are focused rather than spent reorganising notes.
Protect the Saturday block. It’s where projects get finished, and unfinished projects are what make people feel like they’ve been “learning AI” for two years with nothing to show. If you’re choosing a structured course, pick one that assumes an adult with a job and a family, which is the reasoning behind this guide to an AI course for adults that fits your real schedule.
How to tell you’re actually making progress
Vague timelines create vague anxiety. Concrete checkpoints fix that. Here’s a realistic sequence for someone averaging five hours a week:
End of week 2: you can explain to a colleague why a chatbot gave a confidently wrong answer, and what to do about it.
End of week 6: you have three prompts saved that reliably do real work, and one task you’ve stopped doing manually.
Month 3: you’ve run code that called a model API and returned something useful, even if a tutorial walked you through most of it.
Month 6: you’ve built and shared one small tool that someone other than you uses.
If month six looks like month one, the problem is almost never aptitude. It’s usually a lack of finished projects.
Where certificates fit, and what they don’t buy you
A six-hour online certificate will not make anyone an AI professional, and the timeline claims on the sales page should be read with suspicion. What a credential genuinely does is signal baseline interest and give a hiring manager something to ask about. That’s worth having, especially if you’re changing fields. Just treat the certificate as a mile marker rather than the destination, and check whether it’s an AI beginner certificate worth putting on your CV before you pay for it.
The same logic applies to learners in different situations. A university student has four months of structured time and should use it on foundations and a portfolio project. Someone learning at 65 has no deadline at all and can take 18 months to reach the same literacy level without stress, which is roughly the approach described in this guide to AI courses for college students choosing what’s worth a semester. The hours differ. The sequence doesn’t.
Pick a finish line, then work backwards
Timelines get real when you attach them to an outcome you could demonstrate to another person. Something like: by the end of March, I can take a messy spreadsheet of customer emails and produce a labelled summary my team actually reads. That’s a 90-day goal requiring maybe 60 hours. Now divide by the weeks you have and check whether the arithmetic survives contact with your calendar. If it doesn’t, extend the deadline rather than pretending you’ll find the time.
Two rules keep this honest. First, count hours, not weeks, because three months of two hours a week is a very different thing from three months of eight. Second, build in a checkpoint every four weeks where you show something to someone. The people who learn AI fastest aren’t the ones with the best course or the strongest maths background. They’re the ones who keep producing small, imperfect, finished things and letting other people see them. Start with one task this week, keep it small enough to finish by Sunday, and the timeline stops being an abstract worry and becomes a schedule you can actually run.