The Review Bottleneck: 63% of AI-assisted social content never reaches a human decision

The cohort is 104 Bolta workspaces observed between April 1, 2026 and July 27, 2026, covering 11,258 AI-generated posts. Posts written without AI assistance are excluded from the denominator — this study makes a claim about AI-assisted content, so the corpus is restricted to it. A workspace enters the cohort once it has created at least one post inside the window. The window opens in April 2026 because that is when AI agents became generally available; earlier data is pre-release and would not be comparable.

Every post resolves to exactly one final state. The six states below cover 99.5% of the corpus; the remaining 0.5% is content held by an automated quality gate before any human saw it, which we count as neither reviewed nor unreviewed.

Expired unreviewed

Submitted for approval, but the window elapsed with no approve and no reject decision recorded. The post was written, queued, and lost — not judged and turned down. This is the study's headline measure.

Never submitted

Still in draft, needs-revision, or ready-for-scheduling when the window closed. Written, then never sent for review at all — a distinct failure from expiry, at an earlier step.

Rejected

A reviewer looked at it and explicitly turned it down. The only state in which content was assessed on merit and failed.

Published

Reached a live post. An end-to-end yield from creation, not a measure of publishing API success.

Still in flight, failed to publish

Scheduled or awaiting approval inside its window at the window close — not a loss — and content the platform API rejected at the publishing step.

• Internal Bolta and staff-operated workspaces.
• Posts deleted before reaching any terminal state.
• Posts published outside Bolta — we can only observe what our own publisher sent.
• Per-platform breakdowns. Publishing cohorts per platform ranged from 2 to 14 workspaces, far below the minimum cohort size of 25, so no platform-level rate is reported at all rather than reported thinly.
• Month-over-month trend lines. Recent cohorts are right-censored — a post created shortly before the window closed has not had time to expire — so a naive trend would show a decline in expiry that is an artefact of the measurement date, not a real improvement.

• This is a population of Bolta customers, not a random sample of the market. Treat it as a description of this cohort.
• Expiry depends on Bolta's own approval window length. A product with a different window would produce a different expiry rate from identical user behaviour, so the absolute figure is not directly portable to other tools.
• "Never submitted" includes content a user may still intend to send. We cannot distinguish abandoned drafts from deferred ones.
• The cohort spans a single quarter of a product that was changing throughout it. These figures are a snapshot and will move.
• We report shares of created posts, not of workspaces. A small number of high-volume workspaces contribute disproportionately to every share.

All figures are aggregated to cohort level. No cell is reported for a cohort smaller than 25 workspaces. No per-workspace figures, no customer names, and no post content — identifiable or otherwise — are included in this study or in the downloadable dataset.