
Verification in Legal AI Is a Design Problem
Lawyers remain responsible. Legal AI should make its work easy to check before they sign.

A litigation matter can have every document safely stored and still have no shared account of what happened.
The facts may be spread across a chronology, an associate's notes, witness folders, draft pleadings, emails and comments in a review platform. Several people can know different parts of the matter without the firm having one reviewed account that carries across the work.
Mary uses the factual record to describe that account.
The factual record of a litigation matter is the reviewed, source-linked account of what the matter materials establish, dispute and leave unresolved.
It is more than a list of extracted statements. A useful factual record connects each factual proposition to the page or source from which it came, records how confident the team is in it, and preserves later corrections. It also keeps conflicting accounts visible rather than forcing them into one clean narrative.
That definition is deliberately narrower than "everything known about the matter". Legal advice, strategy, witness credibility and the weight to be given to evidence remain matters for lawyers. The factual record supplies the material over which those judgments are made.
A factual record should preserve at least six things:
A chronology may present many of those facts by date. The chronology is one view of the record, rather than the record itself.
A document management system has a different primary job. Platforms such as iManage and NetDocuments are designed to store and govern documents, preserve versions, enforce permissions and keep work connected to a matter. Modern DMS products increasingly add search, AI and context capabilities. That makes them more useful, but it does not automatically create a reviewed factual account of a dispute.
eDiscovery has another role. The EDRM model separates identification, preservation, collection, processing, review, analysis, production and presentation. An eDiscovery platform can be the right place to manage very large collections, code documents, run analytics and prepare productions. A responsiveness code or review tag still does not necessarily preserve the legal team's final understanding of the event described in the document.
A chronology arranges events over time. A matter summary compresses a current understanding into prose. A general AI chat holds the context supplied to that conversation and generates an answer from it. Each can be useful. None, by itself, guarantees that the firm's reviewed factual work will persist when the matter moves to another task or tool.
The boundary is not absolute. DMS and eDiscovery providers continue to add more analytical capability, and specialist products can store documents of their own. The practical question is which system is responsible for the reviewed factual state of the matter.
Litigation does not arrive as a complete, settled dataset.
New documents are produced. An attachment appears months after the email that referred to it. A witness adds detail or changes an earlier account. Expert evidence alters the apparent significance of a transaction. A lawyer discovers that two names in the document set refer to the same person.
A static summary becomes stale. A working factual record needs to absorb new material without silently replacing what the team had previously accepted. Proposed changes should remain visible until an authorised person approves them. Earlier conclusions should be traceable so the team can understand why its view changed.
This is also why correction persistence matters. When a lawyer fixes a date, resolves an entity or changes an allegation from "established" to "disputed", the correction should affect later work. Repeating the correction in every prompt is not a durable workflow.
Legal AI makes it cheaper to read, compare and draft from large document sets. It also increases the number of tools that may form their own version of the matter.
A dispute may pass through the DMS, an eDiscovery platform, email, spreadsheets, a factual-analysis product and several drafting tools. If each system reconstructs the facts independently, the firm gains speed at the beginning of each task and loses some of it reconciling the results.
A factual record gives those tools a reviewed state to work from. The record can support a chronology, a witness brief, a gap analysis, deposition preparation or a draft, while preserving the route back to the evidence.
The model still has work to do. It may need to interpret messy language, connect people across documents and propose an inferred date. The system around the model should preserve those proposals, the underlying sources and the lawyer's decision about them.
A firm can test whether it has a factual record by asking a few direct questions.
Can a lawyer open any material fact to the exact source and surrounding passage? Can the team see conflicting accounts without searching for them again? Is missing or unread material visible? Does a correction survive into the next output? Can another authorised lawyer understand how the current view was reached?
If the answers live only in one person's notes or one generated response, the factual work has not yet become a shared record.
Mary is built around maintaining that record as the matter develops. The Complete Guide to Legal Fact Management explains how documents are turned into structured, reviewable facts. The factual record is becoming infrastructure explains why that persistence becomes more valuable as firms use more AI tools.