Cloud AI vs local AI for confidential documents: an honest comparison
A fair, no-hype comparison of cloud AI and local AI for documents you're personally liable for — what each gives up on privacy, quality, and control, and who each is really for.
The real question for confidential work isn’t which AI is smarter
When people compare cloud AI and local AI, the instinct is to compare answers — which model reasons better, writes cleaner, handles a longer document. On that axis, a large hosted model will often beat one running on your own laptop, and this page won’t pretend otherwise.
But for confidential work, that is the second question, not the first. If the file you are about to feed the model is something you are personally answerable for — an NDA-bound design document, a privileged client file, a patient record, an unpublished manuscript — then the first question is not which answer is better but where does this file physically go, and who can reach it once it is there.
That reframing is the whole comparison. “Cloud AI vs local AI” is really a choice between two trades, and each one costs you something real. What follows is a fair accounting of both — not a case that cloud AI is reckless, because it isn’t.
What cloud AI gives you — and what it asks in return
Start with the appeal, because it is genuine. Cloud AI chatbots and hosted AI assistants run large, frontier-scale models you could never fit on your own hardware. Nothing to install, no GPU to buy, no multi-gigabyte download — you open a browser tab and it works. For a great deal of work, that is simply the right tool, and refusing it on principle would cost you capability for no gain.
The cost appears only when the input itself is confidential. To get an answer, you send your document to the provider’s servers, and from that moment:
- A copy exists somewhere you don’t control. The file now sits on infrastructure you did not build and cannot inspect — the provider, its subprocessors, its hosting layer, and whoever holds access to that chain.
- Retention, logging, and training use are claims you can’t audit. “We don’t train on your inputs” settings and fixed retention windows are commitments about data the vendor already holds. They may be honored in full — but you cannot verify them from outside, and terms can change.
- You’ve added a subpoena target and a breach surface. A third party holding your files can be compelled to produce them, sometimes without notice, and can be breached. Uploaded material also pools — a single intrusion can expose many customers’ documents at once.
- The safeguard is a data-processing agreement you have to trust. Enterprise plans offer real contractual protection. But a contract is a promise about behavior, enforced after the fact — a different kind of thing from the file never leaving your disk in the first place.
None of that makes cloud AI a bad product. It makes it a poor fit for the one case where you would later have to account for exactly where a document went.
What local AI gives up
Fairness cuts both ways, and running the model on your own machine is not a free win. You trade away three real things.
- A local open model, not a frontier hosted one. SealedBrief bundles Qwen3 — 8B and 14B open GGUF models, auto-selected by your hardware. They are capable, but they are not a large hosted system, and you should not expect frontier-scale reasoning. The workflow assumes you read the cited passages and verify answers rather than take them on faith.
- Real hardware, which is yours to provide. Cloud AI runs on someone else’s GPUs; local AI runs on your machine. SealedBrief is comfortable on a 12 GB-or-larger NVIDIA GPU or Apple Silicon, runs CPU-only but slowly, and wants roughly 16 GB of RAM.
- A large one-time download. The models ship inside the app, so the first download is about 15 GB. After that it runs offline with nothing fetched at runtime — but that first step is real, where a cloud tab costs you nothing.
If your work is not sensitive, these are costs with no matching benefit, and cloud AI is the better call. The trade only pays off when the input is something you cannot afford to hand to a third party.
Side by side
| Consideration | Cloud AI chatbots | Local AI (SealedBrief) |
|---|---|---|
| Where your documents go | Uploaded to a third party’s servers | Stay on your own machine |
| Who can technically reach them | The vendor and its subprocessors | You, on your own disk |
| Retention and training use | Governed by terms you can’t audit | Your filesystem is the retention policy |
| Added legal and breach exposure | A subpoena target and a shared breach pool | No third-party custodian to serve |
| Model capability | Large frontier-scale hosted models | A local open model (Qwen3-8B/14B) |
| Hardware you provide | None — it runs on their GPUs | A capable GPU, Apple Silicon, or CPU |
| Getting started | Open a browser tab | A one-time download of roughly 15 GB |
| Document processing offline | No — every query is a network round trip | Yes — the Compute Plane has no network egress |
| The privacy claim | Trust the policy page | Check it yourself with a packet capture |
The pattern in that table is the honest summary: cloud AI wins on capability, convenience, and no install step; local AI wins on data locality and on your ability to prove the privacy claim rather than trust it.
Is cloud AI safe for confidential work?
“Safe” quietly bundles two different questions. The first: is the provider likely to mishandle your file? Probably not — major providers invest heavily in security. The second: can you account for where the file went, and defend that account afterward, to a client, a regulator, a court, or a source? There, the answer is no — you are relying on infrastructure and policies you cannot inspect.
For non-sensitive work, only the first question matters, and cloud AI clears it easily. For material you are personally liable for, the second question is the one that counts, and “the vendor is probably fine” is not a defense you can offer when someone asks where a specific document went. That is the distinction that should drive the choice — not a blanket verdict on whether cloud AI is “safe,” but whether you can answer for the file.
Who each is actually for
- Cloud AI is the right default for work you are not personally liable for. Public research, brainstorming, drafting from non-confidential material, summarizing an article, general coding help — anywhere the input isn’t sensitive and frontier reasoning is worth having, the hosted model is the better tool. Reaching for a local model there just costs you quality for a benefit you don’t need.
- Local AI is for the material you would have to answer for. NDA-bound R&D, privileged files, patient or client records, source documents, unpublished IP — cases where the input itself is the sensitive thing, and “it probably won’t leak” is not a standard you are allowed to work to. There, keeping the file on your machine isn’t caution; it’s the requirement.
Most people need both. The useful question is not which one to adopt forever, but which one a given document belongs in.
Where SealedBrief fits — and how to check it rather than trust it
SealedBrief is the local option in this comparison: an air-gapped, local-first workspace that indexes a folder of your documents and answers questions across them, with the model running on your own hardware. The deeper mechanics live in the pillar guide on local AI for confidential documents; the short version is what makes the privacy claim checkable.
It runs as two operating-system processes that cannot swap roles. The Compute Plane — the part that reads your documents, runs the model, retrieves passages, does OCR, and holds the stores — has no network egress by construction. The Presentation Plane, the interface, reaches the network for exactly two narrow things: a licence check and an update check. Neither of those carries your documents or your questions. So the honest claim is not that the app never touches the network — it is that the process handling your files never does, and the process that does talk to the network never has your content to send.
Two more honest notes, because precise beats impressive. On disk, your document text, source references, and metadata are encrypted with AES-256; the embedding vectors used for similarity search are stored unencrypted, because approximate-nearest-neighbour math needs plaintext to run. And the answers come from that local open model, so verifying citations against the cited passages is the workflow, not an afterthought.
The point of all this is that you don’t have to believe it. Run a packet capture on your own machine while you index files and ask questions, and watch the document-handling process stay silent. It is a five-minute check your own IT can reproduce — the exact commands are on the verify-it-yourself page. That is the real difference between the two categories: with cloud AI you audit a policy; with a local tool built this way, you can audit the traffic.
What it costs, and how to try it
The way to judge a local tool is against your own files, on your own hardware, watching your own network — so the download is the full 14-day trial, not a limited demo. On first launch you choose “Start 14-day trial”: no account, no credit card, and starting the trial makes no network call, which you can confirm with the same packet capture. After 14 days the app opens read-only — nothing is deleted, and you keep full read, search, and export of everything already indexed; only new indexing and new questions pause until you enter a licence.
SealedBrief is a 3-year term licence, not a subscription and not perpetual: Personal at $199 for one device, Professional at $499 for up to three devices plus a commercial-use grant and an exportable audit-evidence surface. Every purchase carries a 30-day, no-questions refund that is separate from the free trial. Linux (a portable AppImage) and macOS on Apple Silicon ship today; Windows is a waitlist, and you are not charged until it ships.
If the document in front of you is one you would have to answer for, that is the one to test it on. Download the free 14-day trial, or compare both tiers on the pricing page.