The distinction that matters
Most legal tools today are cloud-first: your documents are uploaded to the vendor's servers, and the software — including any AI — runs there. It's convenient, and for plenty of work it's perfectly sensible. But it means a copy of the material lives somewhere you don't control, governed by the provider's terms, retention settings and security rather than your own.
Local-first inverts that. The material stays on your own machine; the software runs there too. Nothing is uploaded unless you deliberately send it. The difference is not about features — a local-first tool can do everything a cloud one does short of syncing across the internet — it's about where the material sits, and therefore who is responsible for it.
Why it matters for privileged work
For most software, keeping data on the device is a performance choice. For privileged legal work it is a professional one. The moment a brief — or any part of it — is sent to a third-party service, confidentiality and legal professional privilege rest on that provider's arrangements rather than on your own control of the material.
Australian bar associations and courts have published guidance on generative AI in practice. The themes are consistent: preserve client confidentiality, guard against the loss or waiver of privilege when material is exposed to a third party, and remember that the practitioner remains responsible for their own work. That guidance is not limited to public chatbots — it extends to any software that sends matter content to a cloud service, including legal tools with built-in AI. The safest position is structural: if nothing is sent, there is no third party to trust.
This isn't legal advice, and it isn't anti-cloud. It's a way of thinking about the trade-off: convenience and shared infrastructure on one side, control and a smaller trust surface on the other.
"Local-first" is not "no AI"
It's easy to read local-first as a refusal to use AI. It needn't be. The useful question is not whether to use AI but where it runs and what it sees. A model can run on your own machine and never touch the internet. A cloud model can be used deliberately, for a specific task, on material you've chosen — ideally with identifying details masked before anything leaves the device. The point is that it's a decision you make, matter by matter, not a default you have to remember to switch off.
How Lexware approaches it
Lexware is local-first by default. Your documents, the text extracted from them, the chronology, the pleadings analysis and your notes live in a local database on your own machine. There is no Lexware server, no account, and no telemetry.
- Local by default. Date extraction runs on a model on your own machine; nothing leaves your machine.
- Cloud is opt-in, per matter. If you want the accuracy of a cloud model, you can turn it on — Claude or OpenAI's ChatGPT, on your own subscription — but only for a matter you've explicitly enabled it for.
- Masking before it's sent. On the main cloud paths, a built-in pseudonymiser can replace identifying details with placeholders, and Lexware shows you exactly what will be sent before it goes. It's an automated aid rather than guaranteed anonymisation — the text it shows you is always worth the glance.
The result is that AI is something you reach for on purpose, on your terms — not a condition of using the software. For the full detail, see how Lexware handles your data.