AI for legal document work

"Legal AI" is sold as one category and bought as one purchase, but it describes at least four jobs. Drafting produces new wording. Analysis answers questions across a document estate you already hold.

The distinction decides what evidence matters. A research tool lives or dies on the breadth and currency of its source documents. An analysis tool lives or dies on whether it retrieves the right passage from your own material and shows you where it came from. Comparing them on a single feature grid produces a purchase nobody can defend.

The widely reported failure is invented authority: a citation to a case or clause that does not exist. It is the most visible failure and the easiest to test for, which makes it the least dangerous.

The failures that survive review are quieter. A system retrieves a superseded version of a clause and answers confidently from it. It answers from one document when two in the estate disagree, and never mentions the second. It returns nothing relevant and produces a fluent answer anyway. Each of these produces output that reads correct and passes a skim.

A citation is only useful if it can be checked at the level of the individual claim. A footnote pointing at a 200-page document is a gesture. A reference that opens the source, moves to the page and marks the passage is a control.

Test citations claim by claim rather than answer by answer. Split an answer into its material propositions, then ask of each one whether a cited passage supports it, whether the passage is from the document and version the citation names, and whether material claims exist with no citation at all. Correctness and completeness are separate measurements and should not be collapsed into one score.

Legal estates are unusually hostile to document processing. Executed agreements arrive as scans. Obligations sit in tables where a single cell carries the meaning. Amendments change wording without replacing the document that contains it, so the current position is spread across a chain.

Test on the difficult material rather than a clean sample. Include a poor scan, a document whose substance is in a table, an agreement amended twice, a set with a superseded version still present, and a question whose answer is genuinely not in the document set.

A policy that says users should not see other matters is not a control. The question is whether the retrieval layer can return a passage from a document the user has no right to read, and whether a generated answer can carry that content even when the source is not shown.

Ask where the boundary is enforced. Enforcement at the point of retrieval behaves differently from filtering applied to results, and differently again from separate storage per matter. Test it directly: revoke a user's access to a source and ask a question whose answer only that source contains.

Regulatory expectations here are about competence and supervision rather than a prohibition. A solicitor remains accountable for work produced with a tool, which means the firm needs a recorded position on who checks output, against what, and before which decisions.

Under UK GDPR the questions are the ordinary ones asked precisely: what personal data enters the system, which processors handle it, where it is held, how long it stays, and what happens on deletion. Answer them per deployment configuration, because the answers change with the configuration and a generic vendor statement will not survive a DPIA.

Vendor demonstrations run on corpora chosen to succeed. The only evaluation that predicts anything is one on your own files, with questions written before you see any output and expected evidence recorded in advance.

Keep the reviewer's disposition per question rather than a headline accuracy figure. Counts are only meaningful alongside the denominator, the scoring rules and the cases that failed.