AI manuscript review: 19 agents check every paper before it leaves.
Every Writing hand-over gets an answer: written correctly, minor issues or mistakes found. The model reads the whole manuscript, every mistake must quote the document, and your reviewers still make the call.
SCI-2026-09-014 · manuscript_v3.docx · version 3 of 3
AI review report
Depth: Deep PDF report
86
Writing verdictWritten correctly
Admin accepted · 10:20
0Major
2Minor
5Notes
ModelMost capable · high effort
ReadWhole manuscript · 42 ¶
Agents19 of 19 finished
Agent resultsWeights editable
MethodologyMethods920
CitationBody, references881
Figure and TableFigures, captions831
StatisticsResults900
Journal-FitAbstract, journal master810
NoveltyIntroduction, discussion780
Findings2 open
Minor · ¶31Figure and Table
Figure 3. Results. The caption does not say what the figure shows.
FixedIgnored
Minor · ¶40Reference
Mixed reference styles: entries 4 and 11 use APA, the rest use Vancouver.
FixedIgnored
Models
Advanced models, one connection, and a basic mode for every feature.
PaperFlow talks to the model providers through one manager. The admin pastes a key once and every AI feature uses it. When no key works, each feature falls back to its rule-based version and says so, so nothing breaks.
Primary provider
Anthropic Claude
Through the official PHP SDK, with structured JSON output, reasoning effort control and prompt caching.
Claude Fable 5.1Most capable · runs Deep analyses
Claude Opus 5.5Strong reasoning
Claude Sonnet 5.5Balanced
Claude Haiku 4.5Fast first looks
PaperFlow also supports the OpenAI GPT-4.1 family and the Google Gemini 2.5 family. Each account picks its provider and uses its own key.
Built in
Guard rails on every call
Encrypts every key and never sends it to the browser.
Validates every answer against a schema and corrects it once if needed.
Caches identical requests for seven days, so re-runs cost nothing.
Logs usage and cost for every call.
Falls back to the provider default if an account cannot use the chosen model.
Nineteen specialised agents
Each agent reads only the sections it needs.
Every agent follows an editable prompt template and returns findings with severity, evidence and a suggested fix. Admins set the pipeline order, time limits and input size for each agent on one page, with no code change.
Structure and writing
Document Intelligence
Academic Writing
Figure and Table
Citation
Reference
Research quality
Research Question
Methodology
Literature Review
Statistics
Results
Discussion
Conclusion
Risk and integrity
AI-Risk
Plagiarism-Risk Indicator
Ethics and Research Integrity
Indicators only. Your team still uploads the formal plagiarism and AI-content reports.
Fit and output
Novelty
Journal-Fit
Scoring editable weights
Final Report PDF
How an analysis runs
From hand-over to findings, without a page ever waiting.
Analyses run on a dedicated queue with a worker that starts itself and stops when idle. Every hand-over creates a new version of the same paper, so you can compare drafts.
01StartThe Writing hand-over starts it, or anyone presses “Analyse with AI agents” on a paper page.
02PreparePaperFlow extracts the text from PDF or Word, numbers the paragraphs, detects sections and screens out hidden instructions.
03ReadEach agent gets its sections, the model that the depth setting picks, and the reviewers’ earlier corrections.
04ScorePaperFlow weights the findings into a score and shows the writer verdict first.
05ReviewPaperFlow produces a PDF report and a findings list, and reviewers mark each finding Fixed or Ignored.
The writing verdict
Three possible answers, and a quote behind every one.
The model reads the whole manuscript, not a sample. Every mistake must quote the document, and PaperFlow verifies each quote against the text before showing it. The admin accepts or overrides the verdict, and that decision feeds the accuracy loop.
Written correctly
Ready to move on once the admin accepts the verdict.
Minor issues
A short list of findings, each with severity, evidence and a suggested fix.
Mistakes found
Quotes every mistake from the manuscript and checks it against the text.
Analysis depth, one switch
Deep is the default
ModelMost capable
ReadsWhole manuscript
ReasoningHigh effort
The default. Most capable model, high reasoning effort, the whole manuscript. Slower and costlier, catches the most.
If an account cannot use the chosen model, the call falls back to the provider default instead of failing.
The accuracy loop
90 days per agent · example
Citation91%
Reference89%
Methodology84%
Statistics80%
Novelty72%
Fixed (a hit)Ignored (a miss)
Before each new paper, PaperFlow hands every agent its recent ignored findings with the reviewer’s note, so it stops raising points the team does not want. PaperFlow does no model training, and no data leaves the system beyond the request itself.
Beyond the manuscript
Six more places the same connection works.
One key, one manager, and the same basic-mode fallback everywhere.
“Catch me up” in team chat
A summary of what happened while you were away.
Team report narrative
A written summary of the team’s numbers for managers.
Journal email reader
Pulls the decision, dates and reviewer points out of an editor’s email.
Workload-aware assignment
The member with capacity and the best quality score, with a reason.
Monitoring classifier
Classifies unknown apps and sites every night and saves editable rules.
Connection check
A cached check that tells every feature whether AI mode is really available.
Questions about the AI
What reviewers and owners ask us.
Want to see it read one of your own manuscripts? Book a walkthrough and bring a sample.
Does the AI replace the reviewers?
No. The agents raise findings with evidence; people decide. The agents learn from those decisions, so they raise fewer points your team does not want.
Does PaperFlow send client data to the model providers?
PaperFlow sends only the manuscript text and the paper details the analysis needs, over an encrypted connection, to the provider you configure. It encrypts your API keys and never sends them to the browser.
Does PaperFlow train models on our data?
No. PaperFlow does no model training. The accuracy loop hands each agent its recent ignored findings and the reviewer’s note before the next paper, inside your own system.
What happens if the AI connection is down or not set up?
Every feature has a basic mode that runs on fixed rules. The platform keeps working and tells the admin that AI mode is off.
Can someone hide instructions inside a manuscript to fool the agents?
PaperFlow screens out instructions hidden inside a manuscript when it extracts the text, before any agent reads it.
How is PaperFlow different from a single AI manuscript review tool?
Most AI review tools read a paper once and return one report. In PaperFlow, 19 specialised agents each read only the sections they need, every mistake must quote the manuscript and PaperFlow checks the quote against the text, and the review runs automatically at every Writing hand-over inside your agency workflow. Reviewers mark each finding Fixed or Ignored, and the agents use those decisions on the next paper.
Does the AI review replace a plagiarism or AI-content check?
No. The AI-Risk and Plagiarism-Risk Indicator agents give indicators only. Your team still uploads the formal plagiarism and AI-content reports to the paper record.
Which AI models does the manuscript review use?
PaperFlow runs on Anthropic Claude and also supports the OpenAI GPT-4.1 family and the Google Gemini 2.5 family. Each account picks its provider and uses its own API key, and the depth switch (Quick, Standard or Deep) picks the model and how much of the manuscript it reads.
15-minute walkthrough
See a paper travel from enquiry to publication in fifteen minutes.
We walk you through a live paper: how the admin assigns it, how the AI review reads it, how PaperFlow keeps the client informed, and how the money closes.