AI Policy Must Become Workflow

Many institutions now have an AI policy. Staff attend a webinar, acknowledge an acceptable-use statement and receive a warning to verify generated output. Leaders may then assume that the risk is governed.

That assumption confuses a written expectation with an operating control. A policy can say that every citation must be checked while the workflow hides whether AI was used, compresses deadlines, leaves one person to draft and file, and gives the supervisor no reliable verification point. The rule exists, but the path to completion makes compliance fragile.

Responsible governance must therefore move from document to design. The practical chain is: legitimate purpose → competent person → declared AI role → protected information → primary-source verification → supervisory checkpoint → independent performance → correction and remedy → incident learning → renewed approval.

The objective is not to eliminate professional judgement through more bureaucracy. It is to make responsibility visible at the moments when an error can still be prevented, explained or corrected.

Reuters reported on 2 October 2026 that United States District Judge Arun Subramanian declined to sanction lawyers after AI-generated citation errors appeared in a filing, but warned that increasing reliance on AI could harm clients and hinder the training of junior lawyers. He urged careful verification and suggested that firms consider requiring initial drafts of some briefs to be prepared without AI assistance.…