Article
Systematic Review Traditional vs AI-Assisted Methods

Systematic Reviews with AI: Decisions, Checks and Reporting

INRA Logo

INRA.AI Team

AI Research Platform

AI assistance does not change the question a systematic review must answer: how were studies found, selected, assessed and used to reach the conclusion? Start with those decisions, then choose where software can help.

This guide is for researchers planning a systematic review of interventions. It outlines work to discuss with your review team, rather than a complete protocol or a tested INRA performance comparison.

Define the review before configuring a tool

Write the objective, eligibility criteria and planned analyses before making review decisions. For an intervention question, specify the population, intervention, comparator and outcomes; explain important settings and follow-up periods. Cochrane chapter 2 explains how the question shapes the review. Record and explain later changes rather than quietly adapting criteria to the papers you find.

If you are still deciding whether you need an explanation of a topic, an evidence map or an intervention assessment, start with choosing a review approach. A systematic review is not automatically better suited to every research purpose.

Keep decisions inspectable at each stage

Search: preserve the strategy, not just the results

Work with an information specialist to choose sources and test the search. Save the database and platform, exact strategy, date, limits and results. A search engine's index size does not tell you whether your strategy retrieved the eligible evidence. See Cochrane chapter 4 and the search-planning example.

Selection: distinguish records, reports and studies

Different papers can report the same study. Removing duplicate database records is not enough to identify those relationships. Cochrane chapter 4 calls for independent assessment by at least two people at the final full-text eligibility stage, with a process for resolving disagreements. Preserve the final decision and reasons. An AI relevance score is not a reason to replace that process with review of only borderline cases.

Extraction: check the value and what it describes

A number needs its outcome definition, units, group, denominator and time point. Cochrane chapter 5 recommends piloting collection forms and independent duplicate extraction of outcome data. Resolve discrepancies against the reports; do not infer an unreported value because a field expects one.

Assessment: use a tool suited to the study and result

Cochrane chapter 8 describes RoB 2 for results from randomized trials. It is not a general score for every study design. A relevant paper or an accessible PDF is not automatically reliable evidence, and a risk-of-bias judgment is not the same as certainty across the body of evidence.

Evaluate a proposed AI task before relying on it

Use the following questions to discuss an evaluation with your methods team. They are editorial planning prompts, not a validated quality-control instrument.

  • Define the task: suggesting search terms, prioritizing records and extracting outcome values have different failure costs.
  • Specify the reference: decide who will establish the comparison decisions and how disagreements will be resolved.
  • Inspect errors: record missed eligible studies, incorrect values and unsupported suggestions, including difficult cases.
  • Count correction work: include checking and rework when evaluating time savings.
  • Set stopping rules: agree when the tool must be corrected, restricted or removed from the task.

This article does not establish a safe sampling percentage or a productivity gain. Checking a convenient sample cannot, by itself, establish that the remaining decisions are accurate.

PRISMA supports reporting; it does not certify the review

PRISMA 2020 guides reporting of systematic reviews. Its authors distinguish that purpose from guidance on conduct or assessment of methodological quality. Reconcile a flow diagram with the underlying records and decisions, and check the rest of the applicable reporting requirements. A software export is not evidence that all requirements have been met.

Describe any automation actually used, its role, and how outputs were checked. Keep tool and version details where available. Do not report human verification, a validation result or a reviewer who did not participate.

Where INRA fits

SLR Copilot supports reviewer-led setup, import, screening, full-text review, extraction, RoB 2 assessment and export. Its current in-app assessment form is RoB 2. Choose methods for your research question before deciding whether those features fit. NLR drafting is a separate workflow; these SLR features do not establish the quality of a narrative draft.

Editorial update: October 19, 2026. This replaces unsupported speed and quality guarantees. Qualified methods review of this revision is pending; no clinical results or measured INRA advantage are claimed.