Peer review triage time cut 67% on a manuscript pipeline serving 14 journals.

Context

A long-running open-access scholarly publisher running its own peer-review platform. Manuscript submissions across 14 journals, growing 30% year-on-year. Editorial board members are unpaid volunteer academics — their time is the bottleneck. The platform handles submission, reviewer matching, anonymous review cycles, copy-editing, and DOI assignment. Built and maintained by UES for over a decade. The platform supports thousands of submissions per year across biomedical, engineering, and social-sciences journals.

The problem

Each new manuscript required an editor to read the abstract, classify it against the journal’s scope, check for obvious plagiarism flags, identify candidate peer reviewers from a database of 8,000+ academics, and write the initial editor’s note. Editors spent 11–14 hours per week each on this triage layer alone — work that was necessary but not the part of their role they signed up for. Reviewer matching was particularly painful: scope-relevant plus not-conflicted plus likely-to-respond is a three-way constraint problem most editors solved by going to the same 30 names every time, burning out their best reviewers.

What we shipped

A triage layer built on top of the existing platform. Inputs: manuscript abstract, declared field, author affiliations. Outputs:

• A 3-line auto-classification against the journal’s scope with confidence score
• A plagiarism pre-screen against open scholarly indexes
• A ranked list of 12 candidate reviewers with reasons (topical match score, recent activity, declined-rate history, conflict checks against author affiliations)
• A draft editor’s note for the editor to edit and approve

Built with a retrieval-augmented approach over the platform’s existing manuscript and reviewer databases — no training on client data, no leakage to public models. Hard cost cap of $0.09 per manuscript. Editors retain final approval on every output — the AI never sends a reviewer invite directly. Kill switch and manual fallback documented before launch.

Results

The 67% reduction in editor triage time is a weekly average measured across all 14 journals over the eight months following deployment. The 38% improvement in reviewer response rate is measured against the prior six-month baseline and attributed primarily to broader reviewer sourcing — the AI surfaced underused reviewers the editors weren’t reaching for. Zero rollback events in 8 months of production.

What we’d do differently

We’d build the reviewer-conflict ontology in week one, not week four. We rebuilt it twice because the first version under-modeled cross-institutional collaboration networks. A week of upfront work would have saved three weeks of rework.