Every automation pitch eventually meets the same reasonable question: what happens when it is wrong? For social publishing the answer has teeth, because the failure is public, timestamped, and attached to a brand. An approval workflow is the difference between an error someone catches and an error a thousand people screenshot.
This is a concrete list, not a principle. These are the failure modes a review gate exists to catch — all of them ordinary, none of them hypothetical.
Drafts that were never true
Generated content is fluent, which is exactly the problem: a wrong price, a discontinued product, a claim the legal team retired last quarter — all of it reads perfectly. Fluency is not accuracy, and a draft that was never read by a person has never been checked against reality.
The gate does not need to be slow. It needs to exist at the point where a person who knows the answer spends four seconds per item.
Platform rules the draft ignores
Each platform has hard constraints — a 100-character title on Pinterest, no text-only post on YouTube or Instagram, thread shapes on X — and softer ones: disclosure requirements for paid or AI-generated content, per-platform rules about automation. A variant that ignores a hard constraint fails at publish time; one that ignores a soft one can cost more than a failed post.
Tooling should refuse the impossible (and ours does — a text-only Instagram variant is rejected before it can schedule), but the judgment calls — is this disclosed properly, is this claim one we stand behind — are human ones. That is precisely the split an approval gate encodes: machines handle the mechanical checks, people handle the ones with consequences.
The unbranded draft
Voice drift is gradual. One post that is too cute, one that explains nothing, one that sounds like a different company wrote it on a deadline. Individually each is a shrug; in aggregate they are a brand. A review pass per batch is the cheapest possible brand-consistency mechanism, because the reviewer is calibrating against the last batch, not against memory.
Actions that spend money or trust
Publishing is not the only consequential action an automation can take. Replying to a customer in public, boosting a post with budget behind it, sending an email sequence to a list — these deserve gates even more than a feed post does. The rule of thumb we use: anything that spends money or speaks publicly waits for a person. Nothing ships because a model felt confident.
- Publishing to a connected platform — always gated.
- Ad spend and boosts — gated, with the amount visible at the gate.
- Public replies from the inbox — gated by default; you can loosen specific cases once trust is earned.
What the gate should record
An approval that leaves no trace is a rubber stamp with extra steps. The record worth keeping is short: what was approved, exactly — the content, the platform, the spend if any; who approved it, or which human asked which agent to do it; and what changed as a result. When something slips through, the audit log answers "how did this publish" in one query instead of an afternoon of guessing.
That is also the honest answer to "is this autonomous?": not a mood, a label — automated, partially automated, guided, manual, or unsupported — attached to every capability, and an audit entry with the actor's kind on every action.