What should an AI-assisted social scheduling workflow require before a post can publish automatically?

Separate generation, approval, scheduling, and publishing into explicit states. AI should not move directly from draft creation to a live client account unless the agency intentionally enables that path.

A useful state model is: Draft -> Needs Review -> Approved -> Scheduled -> Publishing -> Published, with Failed and Cancelled as separate outcomes.

Store who or what changed each state. An AI-generated draft should have generation metadata, but approval should record a human or a pre-authorized automation policy.

Before scheduling, validate:
- destination account/page and platform permissions;
- final text and media;
- required link and UTM rules;
- timezone and publish time;
- duplicate protection;
- platform-specific character/media constraints;
- client approval status when required.

For auto-approval, make it a per-client policy with clear boundaries. Examples: only posts from an approved RSS source, only posts based on published CMSPost articles, no regulated topics, no new claims, and a maximum frequency.

At publish time, use an idempotency key so retries cannot create duplicate posts. Log the request, platform response, external post ID, error message, and retry count.

Finally, make failure visible. A scheduled calendar should distinguish Scheduled, Published, and Failed rather than showing everything as if it succeeded. Automation earns trust when it is auditable and reversible, not when it hides complexity.