How to Screen Abstracts for a Systematic Literature Review

01 · Foundations

What abstract screening actually is, and what it is not

You have run your search across Web of Science, Scopus, or EBSCO. You have exported 4,000 results into a spreadsheet or reference manager. Now what? This moment, the transition from database export to paper selection, is where most systematic literature reviews in business and management either succeed or quietly begin to fall apart.

Abstract screening is the second stage of study selection. It follows your initial database search and a deduplication step, and it precedes full-text screening of a much smaller set of papers. Its purpose is to apply your pre-defined inclusion and exclusion criteria to each paper's title and abstract, without yet reading the full text, in order to identify which papers merit closer examination.

This distinction matters. Abstract screening is not:

• Reading papers in full to assess their quality
• Making final decisions about which papers to include in your review
• A subjective filter based on personal familiarity with the literature
• A step you can do once quickly and never revisit

It is a systematic, criteria-driven, documented decision process. Every decision, include, exclude, or uncertain, must be traceable to a criterion you defined before you started screening. This is what makes the process defensible to your supervisor, your viva examiners, and eventually your peer reviewers.

02 · Prerequisites

Before you screen: what must be in place

This is the section most first-time SLR researchers skip, and the omission that causes the most problems. Abstract screening without a properly established protocol produces results that cannot be replicated, defended, or published. Before you open a single abstract, confirm that you have the following four elements documented.

• 01
  A written review protocol

  A protocol documents your research question, theoretical scope, intended databases, search terms, and selection criteria before screening begins. In management research, this is analogous to a pre-registration in experimental science: it commits you to your criteria before you know which papers are in the pool, preventing post-hoc rationalisation of inclusion decisions. Marzi et al. (2025) identify protocol documentation as a foundational pillar of methodologically sound literature reviews, and note that its absence is one of the most frequent quality failures in published management SLRs.

• 02
  A structured research question (PICO or PCC)

  Your inclusion and exclusion criteria must derive from a structured research question framework. In management and business research, the most commonly used frameworks are PICO (Population, Intervention, Comparison, Outcome) for studies examining causal effects, and PCC (Population, Concept, Context) for broader conceptual or theoretical reviews. Your criteria should map directly to these elements, if a criterion cannot be linked to your framework, it should not be a criterion.

• 03
  Written inclusion and exclusion criteria

  Each criterion must be specific, observable, and applied consistently. Szarucki et al. (2025), analysing SLRs published in the International Journal of Management Reviews, identify clear presentation of inclusion and exclusion criteria as one of five essential elements for a replicable search strategy, yet find that many published management SLRs still fail to report them with sufficient specificity. Ambiguous criteria are the primary source of inter-reviewer disagreement and the first thing a peer reviewer will question.

• 04
  A screening tool or form

  You need a structured system for recording decisions, not a mental note, not a colour-coded Excel column without labels. Each record should capture: paper ID, title, abstract, decision (include / exclude / uncertain), reason for exclusion keyed to a specific criterion, and reviewer initials. This log becomes your audit trail, the document that a viva examiner, a peer reviewer, or a supervisor can request and that you can produce without hesitation.

03 · The Process

The step-by-step screening process

With your protocol, framework, criteria, and screening tool in place, abstract screening proceeds in five sequential steps. Each step reduces the pool of papers and produces a documented record that feeds your PRISMA flow diagram.

• 01
  Deduplication

  Before screening begins, remove duplicate records. When searching multiple databases, Web of Science, Scopus, EBSCO, and Google Scholar are common in management research, the same paper will appear in multiple exports. Tools such as Mendeley, Endnote, or Rayyan automate this process. Record how many duplicates were removed: this is the first number in your PRISMA flow diagram and must be reported accurately.

• 02
  Title screening

  The first pass applies your criteria to titles only. This is a rapid triage step, papers whose titles clearly fall outside your scope are excluded here. Be conservative at this stage: if there is any doubt, pass the paper to abstract screening rather than excluding it. The cost of incorrectly excluding a relevant paper at title screening is much higher than the cost of reading an extra abstract.

• 03
  Abstract screening against criteria

  This is the core stage. For each paper, read the title and abstract and apply each inclusion and exclusion criterion in sequence. Record your decision and, critically, the specific criterion that drove each exclusion. Do not use generic exclusion labels such as "not relevant"; record which criterion the paper failed and why. This specificity is what allows your methodology chapter to report exclusion reasons transparently and satisfies PRISMA 2020 reporting requirements.

• 04
  Handling uncertain papers

  Some papers will not yield a clear include or exclude decision from the abstract alone. Record these as "uncertain" and pass them to full-text screening. Do not force a binary decision on ambiguous abstracts, this is a common source of screening errors identified by Azarian et al. (2023) in their assessment of SLR quality across the logistics and management literature.

• 05
  Documenting and resolving disagreements

  If you are screening with a second reviewer, which best practice recommends for at least a sample of your records, disagreements between reviewers must be recorded and resolved through a defined process: discussion, reference to the protocol, or arbitration by a third reviewer. Larasatie et al. (2026) demonstrate that even expert reviewers applying the same criteria disagree most frequently on comprehensiveness of the literature review and methodological rigour criteria, the two most common criteria in management SLRs. A pre-defined resolution process is not optional; it is a methodological requirement.

04 · The Evidence

The reliability problem: why human screening is harder than it looks

A persistent assumption in SLR methodology is that two trained researchers applying the same criteria will reach the same screening decisions. The empirical evidence suggests this assumption requires qualification.

Hanegraaf et al. (2024) conducted the first systematic assessment of inter-reviewer reliability (IRR) in published SLRs, finding an average Cohen's kappa of 0.82 for abstract screening across 45 eligible studies. While this falls within the "strong agreement" range, it also means that in a pool of 3,000 abstracts, two human reviewers applying identical criteria would disagree on approximately 540 papers. The resolution of those disagreements, through discussion or arbitration, is itself a time-intensive process that must be documented.

Larasatie et al. (2026) add nuance to this picture. Using a Many-Facet Rasch Model to analyse expert reviewer behaviour, they find that the specific criteria on which raters diverge most are those related to comprehensiveness of the literature review and methodological rigour, precisely the criteria most common in management SLRs. This has a practical implication: the criteria that matter most for publishability in management journals are also the most difficult to apply consistently.

The same research by Hanegraaf et al. (2024) found that reviewer experience did not correlate with higher IRR scores, suggesting that inconsistency in screening is not primarily a training problem but a criterion clarity and documentation problem. The solution is not more experienced reviewers; it is clearer criteria, better documentation, and a robust disagreement resolution process.

05 · AI-Assisted Screening

AI assistance and the human-in-the-loop imperative

The use of artificial intelligence to support abstract screening has grown substantially since 2020, and the performance evidence is now robust enough to inform methodological decisions. The key finding across multiple independent studies is consistent: AI-assisted screening can dramatically reduce the time burden of abstract screening while maintaining recall rates comparable to manual review, but only when human oversight is built into the workflow.

What the evidence shows

Study	Approach	Key Performance Finding
Cassell et al. (2025)	AI platform (ISLaR 2.0) vs. expert reviewers	Screening time reduced by >90%; sensitivity 0.91; accuracy 0.87
Niraula (2025)	AI-assisted vs. manual SLR, comparative study	2.08× speed-up; total review time cut from 25 to 11 hours; precision improved from 27% to 70%
Li et al. (2025)	LLM + human-in-the-loop design (GPT-4)	Average sensitivity 90%; Cohen's κ 0.71; F1 score 82
Lee et al. (2025)	Agentic AI framework (A4SLR)	Article screening F1 scores 0.917–0.977 across two use cases
Rathi et al. (2023)	LLM comparison (GPT-4, Bison, AI21 Ultra)	Decision match rates 64–67%; sensitivity scores 0.71–0.90
van Mossel et al. (2025)	AI screening in health economics SLRs	Substantial time savings confirmed; semi-autonomous processes more reliable than fully autonomous

The relevance scoring approach, and why it matters for management research

One of the most practically useful developments in AI-assisted screening is the shift from binary include/exclude classification toward relevance scoring, assigning each paper a numerical score indicating its degree of relevance to the review topic. Dennstädt et al. (2024) demonstrated this approach using LLMs to score papers on a structured scale, finding that it allows researchers to set flexible threshold decisions rather than being locked into binary outputs.

This approach has particular value in management research, where the boundary between relevant and irrelevant papers is frequently interpretive rather than categorical. A relevance score allows a researcher to review high-scoring papers first, concentrate attention on borderline papers, and exclude clear non-matches with confidence, all while retaining full control over every final decision.

Why the human must stay in the loop

Niraula (2025) is explicit: AI output contained omission errors in 25% of cases and misinterpretation errors in 10% of instances, making human verification essential rather than optional. Tomczyk et al. (2026), synthesising evidence across 28 AI-SLR studies, identify human oversight as the critical variable determining whether AI assistance enhances or undermines methodological quality. Abogunrin et al. (2025) add an important behavioural dimension: in a discrete choice experiment with experienced SLR practitioners, researchers strongly preferred tools that kept them in control of final inclusion decisions, prioritising transparency and human override capability over speed alone.

06 · Pitfalls

Six common mistakes in abstract screening, and how to avoid them

These are the errors most commonly identified in the literature on SLR quality in management research. Each one is avoidable with proper protocol design and consistent documentation.