These Tools Do Need to "Read" Your Files — But Reading and Uploading Are Different Things
To recognize whether a file is an invoice or a contract, what the amount is, who issued it, the tool has to actually understand what's written inside — there's no way around that step, whether processing happens locally or in the cloud. What actually determines the privacy boundary is where that "reading" happens: entirely on your own Mac, or packaged up and sent to a remote API where a cloud model analyzes it and sends the result back.
Local Processing vs. Cloud AI Processing
A number of tools marketed as "smart organizing" are, under the hood, calling a cloud LLM API. The upside is usually better recognition (cloud models are generally stronger than small local ones); the tradeoff is that your file content — even if it's just an extracted text snippet — does leave your device and land on someone else's server. Whether it's retained, for how long, whether it's used for training — that's governed by their privacy policy, not something you control. For files that are inherently sensitive — contracts, invoices, scanned ID documents — that's a real exposure, not a hypothetical one. What makes this harder is that you often can't tell from the product itself which one you're getting — marketing copy like "smart" or "AI-powered" applies equally to a local model and a cloud API, and you usually have to dig into the privacy policy or technical docs to find out which.
What Filewise Actually Does
Filewise defaults to a built-in local Qwen model for reading content and generating names, downloaded and installed automatically on first launch (about 2GB, one time only). If Apple Intelligence is enabled on your Mac, it uses the system model directly and skips that download entirely; if you've installed Ollama, Filewise prefers that instead. Whichever path it takes, the computation happens on this Mac — "upload" isn't a setting that's turned off, it's a network request that was never written in the first place.
Staying Local Isn't Free Either
To be fair about it: a purely local model's understanding genuinely trails the strongest cloud models today. Unusual formats, complex layouts, and ambiguous wording are more likely to trip up a local model, or leave it unable to extract anything at all. That's the real cost of keeping file content on-device — not a "have it both ways" pitch. Filewise's answer is to not fake confidence it doesn't have: keep the original name and queue it for you rather than guessing an answer just to have one.
How to Weigh It
If the files you deal with are already sensitive — contracts, invoices, scans, financial records — the privacy boundary that local processing buys you is worth the tradeoff in recognition accuracy. If what you're dealing with is genuinely low-stakes (public materials, sample documents with nothing sensitive in them), this tradeoff matters a lot less — staying local is more of a cautious default than the one correct answer for every case. For whether this local judgment is actually reliable, and whether it might delete anything, see this post; if you want to start with how finding files works in the first place, see Can't Find a File You Downloaded?
