Write these down before the model writes for you.

Most teams adopt AI content generation in the same order: first the tool, then the volume, then — after something ships that shouldn't have — the standards. This note is about doing it in the other order. It comes from practice in a regulated environment, where the standards had to exist before the first sentence shipped.

The whole discipline compresses into one principle: the agent suggests, drafts, and surfaces. A human decides, approves, and commits. Every standard worth writing is that line, applied somewhere specific. Four places, in order of when they'll save you.

Standards: written down, not remembered

A model can follow your voice, your vocabulary, and your rules — but only if they exist on a page it can be pointed at. Three things to write down first: a named human owner for every AI-generated surface (a person, not a team alias — if no one owns it, no one catches it), an approved-and-prohibited vocabulary list (every brand has words it never says; the model doesn't know them until told), and voice guidance that lives in a versioned document rather than in the head of whoever's been here longest.

Boundaries: who may do what

Name the judgment calls. Which decisions may AI draft, and which must a human make? The list is shorter than teams expect, but it has to be explicit — assumed boundaries erode under deadline pressure. Two of them are non-negotiable: approval must be structural (nowhere in the workflow should a model be able to mark its own output approved — if it can, it eventually will), and high-stakes surfaces get a list — pricing, legal, financial, medical — where AI drafts always get human review, as a rule rather than a habit.

Review: before it ships

Three conditions. Every AI draft carries provenance — what generated it, when, from which source material — recorded at creation, not reconstructed during the incident review. A binary gate sits before publish: pass or fail, run the same way every time, by a human with the authority to send it back. And accuracy is checked against a source, not a feeling — every factual claim traceable or cut. "Sounds right" is how wrong things ship fluently.

Watch: after it ships

AI content ages like all content, usually faster. Give everything published a review date. Log failures with reasons — when three failures share one reason, that's not bad luck, that's a standard that needs updating. And decide now, calmly, how you'd pull AI-generated content quickly if something were wrong — because deciding it during the something is much worse.

The part most write-ups skip

Standards on a page are necessary and insufficient — the operating layer has to enforce them. If your workflow lives in a database, the boundary can live there too: fields the agent may populate, fields only a human may set, gates that check conditions before anything advances. That's the difference between a policy and a system. Policies get remembered when it's convenient. Systems don't care.

Kyle Simmons builds governed content systems at Fully Inhabit. Fifteen years across financial platforms, universities, and regulated tech; MFA, UNC Wilmington. The systems in these notes are the ones he sells — and runs.