How to learn AI in 90 days, one hour a day

Before you start: decide which outcome you want

"Learning AI" hides three very different goals, and mixing them is the main reason people quit.

• Applied use. You want to do your existing job faster and better. No maths required. This plan is written for you.
• Building with AI. You want to ship tools that call models — apps, agents, integrations. Add basic Python and API work from week seven onward.
• Machine learning as a field. You want to train and evaluate models. That is a linear-algebra-and-statistics path measured in years, not weeks. Do not start there to become productive this quarter.

Write your goal down. Every time you are tempted by a tutorial, check it against that line.

Weeks 1–3: fluency with one general model

Pick a single assistant and use it for everything, badly, until it stops feeling novel. Breadth now is a trap; you want reps.

What to actually do

Take three tasks you repeat every week — a status update, a summary of a long document, a first draft of something you dread — and do all three with the model daily. Keep a running file of what worked. By the end of week three you should be writing briefs (audience, source, constraints, format) instead of one-line questions.

Milestone: a personal prompt file with at least ten reusable briefs, each one tested more than once.

Weeks 4–5: verification and limits

This is the block most self-taught learners skip, and it is the one that makes you employable. Deliberately catch the model being wrong: ask for citations and check every one, give it arithmetic and verify it in a spreadsheet, ask about a topic you know deeply and mark the errors.

Learn the shape of the failure modes: fabricated sources, confident outdated facts, silent unit and currency mistakes, flattened nuance in summaries, and losing the thread in very long context. Build a short checklist you run before anything leaves your hands.

Milestone: a written verification checklist and at least five real errors you caught and documented.

Weeks 6–7: your own information, not the public internet

Move from generic answers to work grounded in your material: upload the actual PDFs, transcripts, spreadsheets and policies you deal with, and practise extraction and comparison rather than open-ended generation — pull every deadline out of a contract, compare two vendor quotes into a table, turn a meeting transcript into owner-and-date actions.

Also learn where the line is: what you may and may not upload. If you handle customer data, medical records or anything under an NDA, find out your organisation's rule before you paste, and prefer tools it has approved.

Milestone: one document-grounded workflow you now use weekly instead of doing manually.

Weeks 8–9: automation of one real handoff

Pick the dullest handoff in your week — form to spreadsheet, email to ticket, invoice to ledger — and build it in Zapier, Make or the automation features of a tool you already pay for. Then break it on purpose: empty field, wrong format, service down. Add the failure branch.

Milestone: one automation running unattended, with a documented failure path and a measured time saving.

Weeks 10–11: depth in your own domain

Now specialise. A recruiter should be automating screening summaries and outreach personalisation; an accountant, reconciliation and anomaly flagging; a teacher, differentiated materials and feedback drafts; a developer, review, test generation and refactoring. Search for how people in your exact role are using these tools, copy their workflows, and adapt.

This is also the point to compare models properly. Run the same three real tasks through two or three assistants and judge on output you care about, not benchmark charts.

Milestone: three domain-specific workflows and a one-paragraph note on which tool you chose for what, and why.

Week 12: package the evidence

Skills nobody can see do not convert into money. Spend the final week writing one public artefact: the task before, the workflow now, the artefacts, the failure handling, the measured saving. Publish it, link it from your CV or profile, and rewrite your professional summary around that single example.

Milestone: a link you can send a stranger that proves you finish work with these tools rather than merely following news about them.

What free resources are worth the hours

Prefer primary documentation over influencer roundups: the official docs and prompting guides published by the model providers themselves are short, accurate and updated when behaviour changes. Add one structured free course for vocabulary — several major providers and universities publish beginner courses at no cost — and one active community in your domain so you see real workflows rather than demos. Skip anything selling a "secret prompt list"; the prompts are the easy part and you will write better ones by week three.

How to tell you are actually improving

Not by tools tried or videos watched. Three honest signals: the time a recurring task takes you has dropped and you can prove the number; you can predict when the model will fail before it does; and someone else has adopted a workflow you built. If none of those has moved in a month, you are consuming content rather than practising.

Where to go next

This plan is deliberately tool-agnostic. If you want the filled-in version — the specific tools grouped by job, the free courses worth the hours, and the 90-day sequence laid out day by day — that is exactly what the AI Survival Kit in the store is. Everything above stays free and works on its own.