The three-tier oversight model every firm should adopt before deploying AI

Most firm AI policies say some version of 'attorneys must review all AI output.' That sentence is unenforceable at scale, and everyone knows it. Reviewing a scheduling summary with the same rigor as a settlement memo means one of the two reviews is theater.

The productive move is to classify actions by consequence before they run, and to let the classification decide the level of human involvement.

Tier 1 is reversible, internal, and low consequence: research summaries, internal drafts, scheduling. It should run automatically with complete logging.

Tier 2 leaves the building or shapes strategy: client communications, draft pleadings, settlement recommendations below the firm threshold. One-click partner approval is the right friction.

Tier 3 binds the client or the firm: filings, final settlements, new client acceptance. It requires review and a signature, and the signature should be recorded somewhere it cannot be quietly revised.

A tiering model written in a memo is a suggestion. A tiering model enforced by the system that executes the action is a control. The difference matters the first time someone asks who approved a filing.