Local AI in a law practice: what is genuinely good, what is not
Brian Niceley · July 2026
Local AI changes one important fact for a law practice: prompts and matter documents can be processed on hardware the firm controls. It does not turn model output into legal judgment, and it does not make the workstation automatically compliant. Custody is the advantage; disciplined review is still the job.
The workflows that fit
- First drafts from firm-approved templates and supplied facts.
- Search and question-answering over a defined local document set.
- Summaries, timelines, issue lists, and comparison tables for human review.
- Local transcription and dictation where audio should remain on premises.
The workflows I would refuse to oversell
- Unreviewed legal advice or filings.
- Autonomous action on live client systems from day one.
- Research that requires current external authority while the machine is offline.
- Tasks where testing shows the local model misses the practice's quality floor.
The acceptance question
Do not ask only whether the model is impressive. Ask whether a named workflow, on representative documents, meets an agreed quality floor with the network disconnected — and whether every risky action still requires human approval. That is a system a lawyer can evaluate.
The engagement should leave behind a data-flow and data-custody write-up for the firm and its counsel. The document records what stays local, any deliberate exception, who can access the machine, how actions are reviewed, and what the setup does not claim.
For the standing page and printable review material, see Private Local AI for Law Firms.