Three professional jobs
Legal AI can help collect facts from documents, organise those facts into an issue map and prepare a first draft. Those are separate jobs. Combining them without intermediate checks makes it harder to discover whether an error began in OCR, extraction, context compression or generation.
A verification loop
- Compare extracted names, dates and amounts with the source page.
- Distinguish the client or court record from general model knowledge.
- Open every cited authority and verify the proposition it supports.
- Mark assumptions and missing evidence instead of filling gaps.
- Keep the final reviewer identifiable and accountable.
The sanctions order in Mata v. Avianca is a durable warning about plausible citations that were not verified. Good legal AI design should make verification easier, but professional duty cannot be outsourced to interface design.
Document-first rather than prompt-first
A prompt-first tool rewards users who can compress a complex file into a few sentences. A document-first tool lets the record do more of that work, then asks the user for the desired output and jurisdiction. This supports judges, litigators and in-house teams without assuming they want the same document.
Choose by risk, not by demo
Evaluate a tool on representative confidential workflows, retention terms, source traceability, correction cost and failure reporting. A polished answer from a sample contract is not evidence that the product can handle hundreds of poorly scanned pages. Explore English legal document templates or the AI lawsuits tracker.