--- title: "The Review Bottleneck: 63% of AI-assisted social content never reaches a human decision" description: "AI made social content cheap to produce. Approval did not scale with it. Across 104 workspaces and 11,258 AI-generated posts, 46.4% expired waiting for a decision and a further 16.7% was never submitted for one — while just 3.0% was ever reviewed and turned down on merit." canonical_url: "https://bolta.ai/research/ai-social-media-benchmark-2026" markdown_url: "https://bolta.ai/research/ai-social-media-benchmark-2026.md" last_updated: "2026-07-28" content_type: "company" publisher: "Bolta" --- # The Review Bottleneck: 63% of AI-assisted social content never reaches a human decision Source: https://bolta.ai/research/ai-social-media-benchmark-2026 Last updated: 2026-07-28 AI made social content cheap to produce. Approval did not scale with it. Across 104 workspaces and 11,258 AI-generated posts, 46.4% expired waiting for a decision and a further 16.7% was never submitted for one — while just 3.0% was ever reviewed and turned down on merit. ## Summary A measured benchmark of how teams work with AI-generated social posts: what humans do with an AI draft, how far edited posts drift from the draft, how reliably approved posts publish on each of eight networks, and how posting cadence develops. All figures, the charts and the downloadable dataset come from the same cohort-aggregated source, and the canonical HTML page states the cohort size, observation window and minimum cohort size in full. ## Who this is for - Researchers and analysts studying AI-assisted content workflows - Marketing and social media teams evaluating AI drafting - Journalists and operators looking for measured benchmarks ## Problems addressed - Claims about AI content quality are usually vendor assertions or surveys, not measurements. - Nobody publishes what humans actually do with an AI draft once they see it. - Publishing reliability differs sharply by network, and that difference is rarely reported. ## Core capabilities - Draft outcome mix: the share of AI drafts approved as written, approved only after editing, and rejected outright. - Median edit distance: how far the published text drifts from the AI draft, as a normalised character-level distance, across drafts that were edited. - Median time to first approval: wall-clock hours between a draft being generated and its first human review decision. - Draft-to-publish rate by platform: the share of drafts created for each of the eight supported networks that reached a published state. - Publish success rate by platform: the share of publish attempts that succeeded on the first try, with no provider error and no retry-to-failure. - Posting cadence for retained workspaces: median posts published per week by workspaces still active past thirty days, with an optional week-by-week series. ## Limitations and boundaries - Every figure is a cohort aggregate. No per-workspace, per-account or per-user value is published, and no post content, draft text or identifiable customer material is released in any form. - A cell is only reported when at least the declared minimum number of workspaces sits behind it; platforms or periods below that threshold are withheld rather than estimated. - Aggregates are computed from Bolta's own product telemetry. They describe how teams using Bolta behave, not the social media industry as a whole. - This is observational data from a self-selected cohort of Bolta workspaces, not a randomised experiment. Nothing here establishes cause. - Publish success depends on each network's API, account permissions and rate limits at the time of the attempt, so platform-level differences partly reflect provider behaviour rather than content quality. - Approval and edit behaviour depends on how each team configures review; teams that publish without human review are not comparable to approval-first teams. - Edit distance measures how much text changed, not whether the change was an improvement. - Figures cover a fixed observation window and are not updated continuously. The canonical HTML study states the exact window. ## Publication status This study is published. The canonical HTML page carries the figures, the charts and a link to the machine-readable dataset. ## Cohort and observation window Figures describe a cohort of Bolta workspaces over a fixed observation window. The cohort size, the first and last day of the window, and the minimum number of workspaces required behind any single reported cell are all stated on the canonical page and carried in the dataset. Cells below the minimum cohort size are withheld, never estimated or interpolated. ## How the measures are defined Draft outcome is the result of the first human review decision on an AI-generated draft, split three ways into approved as written, approved after editing, and rejected. Edit distance is a normalised character-level distance between the draft and the final published text, computed only over drafts that were edited. Publish success counts a first-attempt success against every publish attempt for that network. Cadence counts published posts per calendar week for workspaces retained past thirty days. ## Data governance Nothing in this study is derived from identifiable content. Values are aggregated across workspaces before anything is written to the published dataset; no per-workspace or per-account figure, and no draft, post or message text, is released. A reported cell always has at least the declared minimum number of workspaces behind it, which is what keeps a single large workspace from being reconstructable from an aggregate. ## Charts and dataset Each chart on the canonical page is a static image with its own permalink, so a single figure can be cited or embedded on its own. When the study publishes, the full dataset is also available as a CSV generated from the same source as the charts. ## Where does this data come from? Bolta's own product telemetry, aggregated across workspaces. It is not a survey, and it is not an industry-wide sample — it describes teams that use Bolta. ## Is any customer content published? No. Only cohort aggregates are published. No per-workspace values, no account identifiers and no post, draft or message text are released. ## Can these figures be reused? Yes, with attribution to Bolta and a link to the canonical study page. The dataset is published as a CSV alongside the study. ## Relevant links - [Research hub](https://bolta.ai/research) - [Guides](https://bolta.ai/guides) - [Pricing](https://bolta.ai/pricing) - [Create an account](https://bolta.ai/register) ## Supporting sources - [The Review Bottleneck: 63% of AI-assisted social content never reaches a human decision (canonical study)](https://bolta.ai/research/ai-social-media-benchmark-2026) - [Bolta Research hub](https://bolta.ai/research)