My first real panic was the day I pasted client contracts into ChatGPT.
I got stuck here too: the moment client data leaves your hands
This isn't an "if" problem — it's a "when" problem.
I run Xiaohongshu consulting and pick up legal contract review on the side. Contracts are full of names, ID numbers, bank cards, case details. Before, I'd paste them straight into ChatGPT without thinking and ask, "how risky is this clause?"
Then last month I had drinks with Lao Zhou. He's a divorce lawyer, takes cases solo. Around 11pm he said he'd pasted a divorce agreement involving a publicly listed company's equity into some AI assistant a few days earlier. Next morning, the opposing counsel called to "chat." I didn't verify if it was true — but from that night on, I pulled every client contract out of cloud AI.
What we side-hustlers fear most isn't AI getting the answer wrong. It's "you don't know where your data went."
What is local AI? I figured it out at Lao Zhou's office
Short version: install AI on your own computer. Data stays home, works offline too.
I read a foreigner's article (titled "AI At Home: A Box Of Scraps") who cobbled together a "home AI box" from old hardware to run big models. Sounds like geek toy territory, but the logic fits us solo operators—
Who's already using it?
- Lao Zhou, the divorce lawyer: drops agreements and contracts into local AI, data never leaves his office
- Sister Lin, freelance translator: client source files are confidential, local AI helps her polish them
- Ajie, Xiaohongshu coach: brainstorms topics and drafts copy without burning cloud tokens (those small pay-per-word units)
None of these folks are coders. They just spent one weekend learning it.
Cost to replicate: money, time, barrier, first step
Money: 0 - 3000 yuan
- Bare bones: install a free tool called Ollama on your computer, download the qwen2.5 Chinese model — 0 yuan
- Step up: spend 1500-2000 yuan on a second-hand NVIDIA graphics card for smoother runs
- Full build: assemble a dedicated mini AI machine, 1000-3000 yuan total
Time: 2-4 hours first time, zero after that
My first attempt took about 3 hours, mostly waiting an hour for the model to download. Second time, half an hour.
Barrier: install software, read English error messages — that's it
- No coding needed
- English errors confusing? Install a translation plugin, hover and translate
- Mistake I made: my download cut halfway and I thought it was broken — redownloaded twice. Just wait it out.
First step (the most important one):
- Open ollama.com in your browser and download the installer (Mac users get Mac version, Windows users get Windows version)
- Find your "terminal": on Mac hit Command+Space and type Terminal; on Windows search PowerShell in the Start menu
- Paste
ollama run qwen2.5and hit Enter - Wait for the download (about 3-5 GB, ~10 minutes), then start using it right away
Three types of peers, three different plays
If you're just starting out (no clients yet / under 5000 yuan/month): I'd hold off. Spend that money on getting clients instead. Totally fine to wait — try it once your first contract lands.
If you have 1-2 steady clients: I'd install Ollama today and grab a Chinese model. Just for that one reason: client contracts never touch a third party. Zero cost, half a day — and when clients ask about privacy, you can answer with confidence.
If you're scaling up (3+ people on the team, 10+ clients): I'd run the numbers. Monthly AI subscription × team size × 12 months. If it crosses 5000 yuan/year, buy a second-hand GPU machine — pays for itself in six months. When clients ask "how do you handle data security," you've got a real answer. That's competitive edge.
Last thing: running AI locally isn't a magic fix. It won't teach you to write copy or close clients. It just lets you sleep at night while you use AI.