The 30-Day Switch: Moving From ChatGPT to Local AI

Most people who try local AI do it the same way: they install something on a Tuesday evening, ask it two questions, decide it’s “not quite ChatGPT,” and never open it again. Then they go back to pasting client files into a cloud chatbot and feeling vaguely bad about it.

The problem isn’t the software. It’s that a chatbot is a habit, and habits don’t move because you read an article about why local AI is more private. They move when you actually run your real work through the new thing for long enough to find out where it holds up.

So here’s a plan. Four weeks, one change at a time, using the work you already have on your plate. By the end you’ll know — from evidence, not vibes — how much of your AI use can live entirely on your own computer.

Before you start: one decision, then stop planning

You don’t need to research models, compare quantization formats, or read benchmark charts. That’s the rabbit hole that eats people’s evenings and produces nothing.

Check that your computer has 8 GB of memory or more (here’s how, and why it’s the number that matters), install one thing, and start. If you’re using WorkInPrivate, it looks at your machine and picks a model that fits it, so there’s nothing to choose. If you’re using a developer tool, take whatever it recommends first and move on. You can tune later; you cannot tune your way out of never starting.

One more bit of prep, and it’s the useful one: for one day, keep a note of every single thing you ask an AI to do. Not what you think you use it for — what you actually typed. That list is the real subject of the next four weeks.

Week 1: The easy wins

The first week has one goal: build the reflex of opening the local app instead of the browser tab. So start with the tasks where a local model is already at full strength.

Move these over immediately:

• Rewriting and tightening — “make this email shorter,” “say this more diplomatically,” “fix the tone of this paragraph.”
• Summarizing text you paste in — meeting notes, a long email thread, a document you need the gist of.
• Drafting from a brief — a first pass at an email, a job posting, an outline, a set of talking points.
• Explaining things — “what does this clause mean,” “explain this error message,” “what’s the difference between these two options.”
• Brainstorming and naming — lists of ideas, angles, subject lines.

These are the bread and butter of most people’s AI use, and a modern local model handles them well. If you kept that list of what you actually asked for, there’s a good chance most of it is on this page already.

The one thing to expect: if an answer comes back flatter than you’re used to, it’s usually the prompt, not the model. A big cloud model fills in the blanks when you’re vague; a smaller local one does exactly what you said. Being specific about role, format, and length closes most of the gap, and it takes about ten minutes to learn.

Week 2: The work you were editing yourself around

This is the week the whole exercise pays for itself.

Think about the things you’ve been not typing into a cloud chatbot — the ones where you swapped the real name for “a colleague,” rounded the number, or dropped the detail that actually mattered. The client’s actual situation. The real financials. The medical detail that was the entire point of the question.

Run those, in full, with nothing removed.

This is the category where local AI isn’t a slightly worse ChatGPT — it’s the only option that works at all, because a sanitized question gets you an answer to a question you didn’t ask. Nothing leaves your computer, so there’s no version of the sentence you have to avoid writing. If you handle confidential work for a living, this week is the one that tells you whether the switch is worth making.

Week 3: Documents and disconnection

Two habits to add, both of which push past what a cloud tab does comfortably.

Point it at your own files. Instead of pasting excerpts, give it the documents themselves — the contract, the policy, the stack of research papers, last quarter’s notes — and ask questions against them. This is the point where a lot of people stop thinking of it as “a chatbot” and start thinking of it as a research assistant that has read their filing cabinet.

Use it somewhere with no signal. On a flight, on a train, in the one meeting room where the Wi-Fi dies. Local AI doesn’t need the internet, and the first time you use it offline, the difference stops being theoretical. A cloud chatbot is a service that can be down, rate-limited, or unreachable. This one is a program on your laptop.

Week 4: Be honest about what stays in the cloud

A migration guide that tells you to move everything is selling you something. Here’s the honest boundary.

Keep using a cloud tool for:

• Anything that needs current information. A local model’s knowledge stops at its training date and it can’t browse. For today’s news, live prices, or a just-released API, the cloud wins outright.
• The hardest reasoning you do. Frontier models are genuinely bigger. For a gnarly multi-step proof, a large complex codebase, or analysis where the top few percent of quality really matters, they still have an edge.
• Images, video, and voice, where those matter to you.

Use local for everything else — which, for most people, is the large majority of it, and essentially all of the confidential part.

That split is not a compromise you’re settling for. It’s the actual right answer: use the enormous shared model for public questions where scale helps, and your own private one for anything involving real names, real numbers, and real clients. Most people’s split lands somewhere around 80/20, and the 20 that stays in the cloud is the part they never minded sending anyway.

Now reread your day-one list of what you actually ask AI to do, and mark each item local or cloud. That’s your answer, based on a month of your own work rather than anyone’s marketing.

Making it stick

Three small things separate the people who switch from the people who drift back:

1. Put the local app where the browser tab was — dock, taskbar, keyboard shortcut. Convenience decides more of this than quality does.
2. Don’t run both in parallel forever. Comparing every answer side by side keeps you anchored to the old tool. Use local as the default and reach for the cloud only when you hit one of the exceptions above.
3. Do the math once. A $20/month subscription is $240 a year, every year. A one-time purchase isn’t. The comparison is less close than it looks once you’re past the first year.

The bottom line

You don’t need to decide whether local AI is good enough in the abstract. You need four weeks of your own work to find out — starting with the easy wins, then the things you’d been editing yourself around, then your documents, then an honest look at what’s left.

The free trial runs for seven days on the computer you already own, which covers Week 1 and the start of Week 2 — comfortably enough to know whether the rest is worth it.