What does “learning AI from scratch” actually mean in 2026?
“AI” covers two very different skill sets. One is building models: statistics, Python, machine learning and deep learning. That is the machine learning engineer or data scientist path, and it usually takes a year or more.
The other is building with models: using large language models (LLMs) such as ChatGPT, Gemini or Claude, connecting them to your data and tools, and automating real work. This is what most AI operations, automation and “AI generalist” roles ask for, and it is where a working professional can become useful fastest.
This plan focuses on the second path, with an optional coding track in each month so you can move towards engineering later. If you are unsure which path suits you, read AI vs machine learning vs deep learning first.
How many hours a week do you need?
Plan for 8–10 hours a week: about 1 hour on four weekday evenings and 4–6 hours across the weekend. Over 26 weeks that is 200–250 hours. Less than 5 hours a week usually means you forget last week’s material before you use it.
Split each week roughly 30% learning and 70% building. Watching videos feels productive, but the skill only forms when you try something, it breaks and you fix it.
The 6-month AI learning plan at a glance
| Month | Focus | Free resources to use | What you build |
|---|---|---|---|
| 1 | How LLMs work, tokens, context windows, prompting basics | Elements of AI (Part 1); DeepLearning.AI short courses | A personal prompt library of 20 tested prompts for your own job |
| 2 | Research, summarising and writing with AI; checking facts | NotebookLM or Perplexity free tiers; the prompting guides from OpenAI, Google or Anthropic | A research memo with cited sources on a topic from your industry |
| 3 | No-code automation: triggers, actions, webhooks, APIs | n8n documentation and the self-hosted Community Edition (free) | One automation that saves you at least an hour a week |
| 4 | RAG: embeddings, chunking, vector search, grounded answers | Microsoft’s Generative AI for Beginners (RAG and vector database lessons) | A question-answering bot over 20–50 of your own documents |
| 5 | Agents, tool calling, AI-assisted coding | Generative AI for Beginners (function calling and agents lessons); Anthropic’s “Building effective agents” | A small agent that uses two tools, plus a simple web app built with an AI coding assistant |
| 6 | Evaluation, safety, documentation, job search | OWASP Top 10 for LLM Applications; your own test sets | A capstone with a test set, a write-up and a 3-minute demo video |
Optional coding track, running in parallel: Kaggle Learn’s free Python course in months 1–2, pandas in month 3, and Kaggle’s Intro to Machine Learning or Google’s Machine Learning Crash Course in months 4–6.
Month 1: How do large language models work, and how do you prompt them?
Start with the mental model, not the tools. You need to understand four ideas well enough to explain them to a colleague:
- Tokens: models read and write text in chunks, and pricing and limits are counted in tokens.
- Context window: the model only “knows” what is in its training data plus what you put in the prompt.
- Next-token prediction: the model produces likely text, which is why it can sound confident and still be wrong.
- Temperature and randomness: the same prompt can give different answers.
Elements of AI Part 1 is free and needs no programming. Pair it with our explainer on what an LLM is and the beginner’s guide to prompt engineering.
Project: take five tasks you do every week (a status email, meeting notes, a report summary) and write a prompt for each. Run each prompt three times, note what goes wrong, and improve it. Keep the final versions in a document. By the end of the month you should have about 20 prompts that you trust.
Month 2: How do you use AI for research and writing without getting wrong answers?
This month is about working with sources. Upload documents to a notebook-style tool, ask questions and check every claim against the original. Learn to ask for citations, to spot invented references and to write prompts that say “answer only from the attached file”.
Project: write a two-page research memo on a question from your industry, for example “How are Indian NBFCs using AI for collections?”. Every factual sentence should link to a source you have opened yourself. Note where the AI tool got things wrong. That list of errors is useful material for interviews.
For day-to-day use, how to use ChatGPT for work has more examples, and ChatGPT vs Gemini vs Claude helps you choose a main assistant.
Month 3: How do you automate work with n8n, Zapier or Make?
Automation is where AI starts saving hours rather than minutes. Learn the building blocks: a trigger (a new email, a form submission, a schedule), actions (create a row, send a message), and an AI step in the middle that classifies, extracts or drafts.
n8n’s Community Edition is free to self-host, according to n8n’s pricing page (checked September 2026). Zapier and Make have free tiers with limits; see each vendor’s pricing page. Our n8n tutorial for beginners walks through a first workflow.
Project: automate one real, boring task. A good example: every enquiry that arrives through a Google Form is classified by an LLM as “sales”, “support” or “spam”, written to a Google Sheet with a one-line summary, and the sales ones are posted to a Slack or Teams channel. Measure the time saved per week.
Would you rather follow this plan with a cohort and mentor feedback?
The ISS AI & Agentic Systems program covers the same ground (prompting, research, automation, RAG, agents and vibe coding) over 16 live weeks. Compare its week-by-week curriculum with this plan, or download the free AI Projects Starter Kit on this page for six project briefs you can start on your own.
View AI & Agentic Systems curriculum →Month 4: What is RAG, and how do you build a knowledge bot?
Retrieval-augmented generation (RAG) lets a model answer from your documents instead of from memory. The pipeline has five steps: split documents into chunks, turn each chunk into an embedding (a list of numbers that captures meaning), store the embeddings in a vector database, retrieve the most similar chunks for each question, and give those chunks to the model with an instruction to answer only from them.
You can build this without code in n8n, or in Python if you are following the coding track. Microsoft’s free Generative AI for Beginners course on GitHub has lessons on search apps with vector databases and on RAG, with Python and TypeScript samples. Read what RAG is in AI for the concepts.
Project: a bot that answers questions about your company’s HR policies, a set of RBI circulars, or a product manual. Write 20 test questions with correct answers before you build it, then measure how many it gets right. A bot with a score (“17 of 20 correct, 2 refused, 1 wrong”) is far more convincing than a demo.
Month 5: How do AI agents and AI-assisted coding work?
An agent is a model that decides which tool to call next, looks at the result, and repeats until the task is done. Learn tool calling (also called function calling), how to limit what an agent can do, and when a fixed workflow is safer than an agent. Anthropic’s article “Building effective agents” is a clear, free explanation of when to use each.
Also spend part of this month on AI-assisted coding (“vibe coding”): describing an app in plain English and letting a tool like Cursor, Claude Code or Lovable write most of the code. You do not need to be a developer, but you do need to test what it builds. See our vibe coding guide and how to build an AI agent.
Project: an agent with two tools, for example one that looks up an order status in a sheet and one that drafts a reply. Add a rule that it must ask a human before sending anything.
Month 6: How do you finish with a portfolio capstone?
Pick one problem from your own industry and combine what you built: an automation that collects inputs, a RAG step that grounds answers, and an agent or app that acts on them. Then do the part most self-learners skip:
- Test it: a set of 20–30 real inputs with expected outputs, and a score.
- Break it: try prompt injection (“ignore your instructions and…”), empty inputs and wrong file types. The OWASP Top 10 for LLM Applications (2025) lists the main risks to check.
- Document it: a one-page write-up with the problem, your design, the test results, the cost per run and what you would do next.
- Show it: a 3-minute screen recording.
For project ideas at different levels, see AI projects for your resume and AI automation projects for beginners.
How do you know you are ready to apply for AI roles?
Use this checklist. If you can tick at least six, start applying.
| Signal | What it looks like |
|---|---|
| You can explain the basics | You can explain tokens, context windows, hallucination and RAG to a non-technical manager in two minutes each |
| You have three shipped projects | An automation, a RAG bot and an agent or app, each with a public write-up |
| You measure quality | Each project has a test set and a score, not just a demo |
| You know the costs | You can estimate cost per run and per month for your projects |
| You think about risk | You can name the failure modes you tested for and how you handled them |
| You can debug | When a workflow fails, you read the error and fix it without starting over |
| You can scope | You can say what should not be automated and why |
| You have used AI at work | At least one project solves a real problem from your current or past job |
Roles to target first are covered in AI jobs for freshers in India and the AI career roadmap for India. Working professionals often have an advantage: an automation that fixes a real problem in your current function (finance, HR, operations, marketing) is strong proof, and you can sometimes move into an AI role inside your own company.
What mistakes slow people down?
- Collecting certificates instead of building. Free certificates are fine, but three tested projects say more.
- Starting with advanced maths. If your goal is building with LLMs, linear algebra can wait. If your goal is ML engineering, it cannot.
- Chasing every new tool. Pick one assistant, one automation tool and one coding assistant, and go deep.
- Building toy demos. A chatbot that answers general questions proves little. A bot over a specific document set with a measured score proves a lot.
- Learning alone with no feedback. Share your work in a community, with a colleague or with a mentor, and ask what is unclear.
Frequently Asked Questions
Can I learn AI from scratch without coding?
Yes, if your goal is to build with AI rather than build AI models. Prompting, research workflows, no-code automation in tools like n8n or Zapier, knowledge bots and AI-assisted app building can all be learned without prior coding. Machine learning engineering roles still need Python and maths.
How long does it take to learn AI from scratch?
About six months at 8 to 10 hours a week is enough to learn to build useful automations, RAG bots and simple agents. Becoming a machine learning engineer, with Python, statistics and deep learning, usually takes a year or more.
What is the best free resource to start learning AI?
Elements of AI Part 1 is a free, no-code introduction to the concepts. For generative AI specifically, Microsoft's free Generative AI for Beginners course on GitHub has 21 lessons covering prompting, RAG, function calling and agents. Kaggle Learn has free Python and machine learning courses if you want to code.
Do I need maths to learn AI?
Not for using and building with large language models. You need basic logic and comfort with spreadsheets. For machine learning engineering you need algebra, statistics and some linear algebra; Google's Machine Learning Crash Course lists these as prerequisites.
Which AI projects should a beginner build first?
Start with a prompt library for your own job, then one automation that saves real time, a question-answering bot over a specific set of documents, and a small agent with two tools. Give each project a test set and a score so you can show it works.
Sources and methodology
- University of Helsinki and MinnaLearn, Elements of AI: free, Part 1 needs no programming, Part 2 recommends basic Python. Checked September 2026.
- Microsoft, Generative AI for Beginners (GitHub): 21 lessons, free and open source, Python and TypeScript samples, lessons on RAG, function calling and agents. Checked September 2026.
- Kaggle, Kaggle Learn: free Python, pandas and Intro to Machine Learning courses. Checked September 2026.
- Google for Developers, Machine Learning Crash Course prerequisites: algebra, basic programming, NumPy and pandas prework. Checked September 2026.
- DeepLearning.AI, short courses. Access terms vary by course; check the course page. Checked September 2026.
- n8n, pricing page: Community Edition is free and self-hosted. Checked September 2026.
- Anthropic, Building effective agents (19 December 2024).
- OWASP GenAI Security Project, Top 10 for LLM Applications 2025. Checked September 2026.
- ISS program details come from our own program page.
Method: the month-by-month plan, weekly hours and readiness checklist are the ISS Editorial Team’s recommendations, not figures from a study. Resource details come from each provider’s official page. Free tiers and course access can change; check the provider’s current page.
Next steps
Start this week: pick the five tasks for your month 1 prompt library and block 8 hours in your calendar. If you would prefer to follow a structured version of this plan with live weekend classes, a mid-week doubt-clearing session and a portfolio capstone, review the AI & Agentic Systems curriculum.
If it fits, you can apply for free. You speak with admissions first and pay only after you accept an offer.