How to do a systematic literature review with AI (without cutting corners)

Systematic literature reviews are the most rigorous form of literature review and also the most labor-intensive. Done properly, they follow a documented protocol (often PRISMA), include explicit inclusion and exclusion criteria, screen every paper twice, and produce a structured synthesis where every claim is traceable.

A typical systematic review takes a research team six to eighteen months. AI tools can compress some of that work, but only if you use them in ways that preserve the methodological rigor. Use them badly and you produce a review that looks systematic but isn't — which is worse than not running one at all.

This guide walks through how to run a systematic review with AI assistance in a way that the methodology section can still defend the work. Literica AI is the tool we know best, so we use it as the running example; most of the workflow translates to other grounded AI tools.

Before you start: protocol first

A systematic review is defined by its protocol. Before you touch any tool, you need:

• A clear research question (PICO format for clinical questions, or a comparable structured form for other fields)
• Inclusion and exclusion criteria written out
• A search strategy with explicit databases and keywords
• A screening process (typically two-reviewer with conflict resolution)
• A data extraction template
• A planned synthesis approach (narrative, meta-analytic, or both)

Register the protocol on PROSPERO (for health-related reviews) or an open registry appropriate to your field. AI doesn't change any of this — the protocol still has to exist, and it still has to be registered before you start.

Step 1: Run the database searches

This is the part AI does not replace. Database searches for systematic reviews need to be reproducible and auditable. You run them in PubMed, Embase, Scopus, Web of Science, or whatever databases your field uses, with documented Boolean strings.

Export the results to a citation manager. Zotero, Mendeley, EndNote — any of them. Deduplicate.

The output of this step is a single set of records — typically a few hundred to a few thousand — that will be screened in step 2.

Step 2: Screen titles and abstracts

This is the first place AI can help, with caution.

Traditional screening involves two independent reviewers screening every title and abstract against the inclusion criteria, with conflicts resolved by a third reviewer or by discussion. This is slow and expensive.

AI tools can do a first pass through your records, applying your inclusion criteria to titles and abstracts. The output is a flagged list — likely include, likely exclude, uncertain — that two human reviewers then screen properly.

The important constraint: AI is not a reviewer. The methodology section should describe AI as a screening aid that pre-flagged records, with two human reviewers still completing the actual screening on the full record set. Anything less than this exposes the review to legitimate methodological criticism.

Literica AI's Lens is useful at the database-search step itself, searching across 240M+ public papers from Semantic Scholar, arXiv, Crossref, and OpenAlex by meaning rather than keyword — a useful supplement (not a replacement) for your formal Boolean searches in PubMed and the other databases your protocol names. At the screening step, Lens can also re-rank your de-duplicated record set against the inclusion criteria for closer human review.

Step 3: Retrieve full text

Once you have the included record set from screening, retrieve the full PDFs. This step is administrative — interlibrary loans, paywalls, contacting authors for preprints.

Upload the full-text set to your AI workspace. In Literica AI, this is where the workflow shifts from your reference manager to the AI tool, because everything from here forward is reading and synthesizing, not searching.

Step 4: Eligibility assessment on full text

The included set from screening gets a second pass on the full text. Papers that looked eligible from the abstract but turn out not to be on full reading get excluded with a documented reason.

This is human-in-the-loop work. AI tools help by pulling out the relevant methodological details quickly — "what was the sample size", "what intervention was used", "was the design RCT or observational" — so a human reviewer can make the inclusion call faster. But the call itself stays with the reviewer.

Step 5: Data extraction

This is where AI tools save the most time, and where most teams adopt them first.

For each included paper, you extract a structured set of fields — typically the same fields across every paper, defined in advance in your extraction template. Population, intervention, comparator, outcome, effect size, study design, risk of bias.

In a citation-grounded AI workspace, you point the tool at the included PDFs and ask for the extraction fields. The output should be a table where every cell points back to the specific page in the specific PDF the value came from.

The non-negotiable verification step: every extracted value gets checked by a human reviewer against the source PDF before it enters the analysis dataset. Citation grounding makes this fast — one click per cell — but you still have to actually do it. Skipping verification is the single most common failure mode of AI-assisted systematic reviews, and it produces reviews that are wrong in ways that are hard to catch.

We've written about why this matters in citation grounding: how Literica AI keeps AI honest.

Step 6: Risk of bias and quality assessment

For each included study, you apply a risk-of-bias tool — Cochrane RoB 2 for RCTs, ROBINS-I for non-randomized studies, or the equivalent in your field.

AI tools can pre-flag potential bias by surfacing the methodological text from each paper, but the assessment itself is judgment work. Two reviewers per paper, conflicts resolved by discussion, exactly as the methodology dictates.

Step 7: Narrative synthesis

Once you have a verified extraction dataset, the synthesis section is where AI helps most visibly.

In Literica AI, the literature review tool drafts a synthesis section from your included papers, with every claim citation-grounded. For a systematic review specifically, you use the drafting capability to produce:

• A narrative description of the included studies
• Thematic groupings of findings
• A discussion of consistency and divergence across studies
• Identified gaps

This is draft output. The synthesis needs an editorial pass that adds the things AI cannot know — the field-specific context, the methodological caveats, the connection to your own contribution.

A useful longer walkthrough of this part is in our literature review guide.

Step 8: Meta-analysis (if applicable)

If your review supports a meta-analysis, the statistical work happens in R, Stata, or RevMan from the verified extraction dataset. AI tools are not where the meta-analysis happens.

Where AI helps here is in the reporting — drafting the prose around forest plots, summarizing heterogeneity results, framing the limitations. Same rules apply: citation-grounded drafts that a human edits.

Step 9: PRISMA reporting

The completed systematic review reports against the PRISMA 2020 checklist — flow diagram of records screened to included, methodology that documents every choice, results that match the protocol.

The methodology section should explicitly describe how AI was used at each stage. PRISMA 2020 includes guidance on reporting AI use; reviewers and journal editors increasingly expect it. Being transparent here is not optional.

A reasonable methodology paragraph reads something like:

Title and abstract screening was performed by two independent reviewers (XX and YY). Records were pre-screened using [tool name], a grounded large language model assistant, against the inclusion criteria; pre-screening flags were used only as a sorting aid and did not replace human screening. Data extraction was performed using [tool name] with every extracted value verified against the source PDF by one of two reviewers (XX and YY) before inclusion in the analysis dataset.

This is honest, defensible, and reproducible.

What AI doesn't replace

To be clear about scope:

• AI does not replace the protocol
• AI does not replace database searches
• AI does not replace the two-reviewer screening process
• AI does not replace human inclusion/exclusion decisions
• AI does not perform the meta-analysis
• AI does not perform risk-of-bias assessment

What AI replaces is the mechanical work in between — the reading, the extraction-into-tables, the synthesis drafting. That's where the time savings live, and it's substantial. A team that used to take twelve months on a review can deliver a comparable review in six to eight months with AI assistance, if the methodology stays rigorous.

Try Literica AI on a small piece of this

If you're considering AI tools for a systematic review your team is planning, the lowest-risk way to evaluate is to take the data extraction step on a pilot set of ten papers and compare. Two reviewers do it manually; Literica AI does it with verification. Compare extracted values, count discrepancies, look at the time difference.

The Explorer plan is free and large enough for a ten-paper pilot. Researcher at $18/month is enough for a typical full review of 50-150 papers. Teams running multiple reviews concurrently use Lab.

A bit more on the data-extraction workflow specifically is in our literature review guide, and the features page covers the full Literica AI surface area.