Why Legal AI Should Run on Your Machine, Not Someone Else's Cloud
Sensitive testimony shouldn't be uploaded to a vendor just to be useful. The case for on-device legal AI — privacy, continuity, and cost — and where the line actually is.
There’s a default in modern software that the legal field should push back on harder than it does: the assumption that to be useful, your data has to leave the building. Upload the transcript. Send the recording to the API. Trust the dashboard.
For most software, fine. For legal work — depositions, medical records, privileged material — that default is backwards.
Clarification — September 4, 2026: Local processing reduces particular dependencies; it does not automatically make a tool private or independent of its vendor. The privacy and continuity paragraphs below now distinguish contractual protections, installed capabilities, and network requirements.
The case for on-device
Privacy is the obvious one. Sending testimony to a hosted tool puts it on infrastructure you don’t operate. A no-training promise answers only one question: review the applicable agreement, retention, access, deletion, change terms, and actual settings too. ABA Formal Opinion 512 calls for understanding the tool’s terms and confidentiality risks; it’s Model Rules guidance, and the applicable jurisdiction’s rules still matter. Local files need device security and backup controls as well.
Continuity is the underrated one. A workflow with its required software and models installed locally can keep working through a hosted-service outage. (I wrote about one interruption separately.) But check for remote sign-in, license validation, model downloads, and cloud storage before calling it independent. Test the work with the network unavailable; a desktop window alone doesn’t establish that the workflow can finish offline.
Cost is the quiet one. Per-seat, per-document, per-minute pricing adds up fast across a firm. A lot of legal AI work — format conversion, search, first-pass extraction — doesn’t need a frontier model in a data center. It needs to run where the file already is.
Where the line actually is
On-device isn’t a religion. Some things genuinely need a big model you can’t run locally yet, and that’s fine — use it, with the safeguards from the reliability checklist. The point isn’t “never use the cloud.” It’s “don’t send your most sensitive material across the wire by default, for tasks that never required it.”
It’s why DepoAudio keeps core recording conversion on your machine — the audio is testimony, and testimony shouldn’t need a round trip to be playable. Its reviewed v1.0.2 privacy statement distinguishes local processing from optional model downloads and users’ own backup controls. The more of your pipeline runs on infrastructure you control, the less any single vendor’s bad week is your problem.
Ask of any legal AI tool: where does my data go, and what happens to my workflow if that destination disappears? If the honest answer makes you uneasy, that’s the tool telling you something.
Curious how far on-device can actually go for legal work in 2026 — I’m building toward an answer.