Monday morning, I nearly sent a client contract to the wrong group

Last Monday morning, I was organizing documents for freelance photographer Chen Ke in Hangzhou. I had a spreadsheet containing home addresses and ID information. The client said, “Don’t put it online,” but I nearly dragged it straight into a cloud AI tool. Then I stopped: being able to process the material doesn’t mean I have the right to decide where it goes. That was when I realized privacy isn’t something a prompt can solve.

What local AI actually is, and who’s using it

Pete Warden has argued that local models, open models, and voice interfaces will reshape how we use our devices and organize our work. In simple terms, local AI means running the AI model on your own computer without sending the material to an external server. A voice interface lets you state what you need without typing everything.

I know Zhou Yuan, an indie developer who took on an image-archive project in Chengdu last year. The client didn’t want to upload the original photos, so he tested local models on an old laptop, extracted text folder by folder, then checked the results manually. Zhou Yuan said installation wasn’t the slowest part; figuring out which files could leave the original device was. It isn’t “install it and become invisible.” It adds another decision point.

What it costs me to recreate this workflow now

If you use an existing local AI desktop app, the hardware cost is roughly ¥0–¥500: if your computer already runs normally, you can try it as-is; a machine with more memory might cost a few hundred yuan more. My first attempt took about 90 minutes and produced a small-scale workflow. If you have a lot of material, I’d block out an afternoon. The technical bar is willingness to check where files are stored, where the model comes from, and what it outputs—not coding ability.

The first step isn’t downloading a model. It’s opening your computer’s “file storage location” or “privacy” settings and checking whether client files sync to cloud storage. The second step is finding a clear interface that supports local processing, then starting with a photo the client has authorized that contains no identity document. When I’m unsure, I’d rather process a fictional copy.

What I’d recommend at each stage

If I’m just starting out and still have little material, I’d keep using a trusted cloud service, remove sensitive fields first, and avoid adding hardware costs just to be “local.” If I already have one or two clients, I’d classify the data: what can be uploaded, and what must stay local, with a separate workflow for the latter. If I’m scaling, I’d bring local AI into the formal handoff process, recording who gave authorization, where the files are, and how they’re deleted afterward. That often matters more than buying another tool.

Not everyone needs this setup; cloud services are already enough for most small teams. There’s no rush to try it. Start with client authorization and data classification, then build a local workflow when a task genuinely can’t be uploaded.