AI for HR Policy Communication

Abstract

One clause, three readings, and the sentence a model must never write A policy is approved once and read hundreds of times, and almost everything that goes wrong with it happens on the second occasion rather than the first. An approved policy fails as a reading problem rather than a drafting problem. The clause is lawful, signed and filed; three people who stand in different relations to it then read it three different ways, and the divergence stays invisible until somebody has to apply it to a person. That divergence is findable in advance, cheaply, by handing one clause unchanged to the person it constrains, the person who enforces it and the person who answers questions about it, then recording where their readings part. A chat model helps with exactly one part of that work, which is producing the questions a clause raises and a restatement to sit beside it. It must never produce the clause, because a sentence that reads well and means something slightly different is the most expensive artefact in the subject. Audiences: The person who gets the same question every month — Answers the same three questions about the same clause, month after month, and answers them slightly differently each time depending on who is asking and how tired she is. Nothing about the clause is unlawful and nothing about it is clear. The manager who has to apply it to a person — Has to enforce a sentence he understands differently from the person it lands on, in a conversation where the difference surfaces for the first time and cannot be settled from the page. Both readings are honest and only one of them was written down. The person who announced the change — Sent an all-staff message explaining a new clause, watched it get forwarded, paraphrased and shortened, and now finds the version people quote to each other is not the version anybody approved. The approved text and the circulating text have drifted and nothing tracks the gap. Note: written from Indonesian operator context. Frameworks apply broadly to other emerging-market and SME settings.