Systematic literature review with AI: accelerate without sacrificing rigor

A systematic literature review (SLR) is the gold standard of evidence synthesis—and one of the most time-consuming exercises in research. AI can now automate its heaviest stages. Here is what the science really shows, the pitfalls to avoid, and where Charlie fits in.

Emerit Science Team

July 2026

Systematic literature review accelerated by Charlie

A systematic literature review is not an ad hoc review. It is a reproducible protocol: you define a question, an exhaustive search strategy and inclusion criteria, then you screen every reference, extract the data from the selected studies, and synthesize them—often as a meta-analysis. All of this follows standards such as PRISMA 2020 or the Cochrane guidelines. That is what makes it the most reliable source of evidence in science and medicine.

The problem: a rigorous SLR takes months. Screening 5,000 to 20,000 titles and abstracts by hand, in duplicate, then methodically extracting data from hundreds of articles ties up entire teams. And scientific output doubles every few years, so the workload only keeps growing. This is precisely where the 2026 generation of AI agents changes the game—provided you know what they do well, and what they do badly.

Where the time goes: the five stages of an SLR

Every systematic review goes through the same phases. Each has a different automation profile—some lend themselves very well to AI, others require strict human oversight.

Stage	What needs to be done	Automatable?
1. Search	Translate the question into Boolean search strings, query PubMed, Embase, etc.	Partly—AI drafts the queries, the expert validates them
2. Screening	Sort thousands of titles/abstracts, then full texts (in duplicate)	Highly—the most mature stage for AI
3. Extraction	Extract PICO, sample sizes, results and methods from each study	Assisted—good gains, but human verification is essential
4. Synthesis	Pool the results: meta-analysis, qualitative narrative	Assisted—AI drafts and structures, the human draws the conclusions
5. Risk-of-bias assessment	Judge the methodological quality of each included study	Barely—expert judgment required; AI alone is unreliable

This gradient is essential: the value of AI is not to “do the review for you,” but to absorb the mechanical workload (screening, structured extraction) where it excels, leaving scientific judgment to humans where it matters.

What the research really shows (2025 figures)

Rather than marketing promises, let's look at the results published in 2025 by teams that measured real systems on real reviews.

Screening—the most mature stage. A review of more than 20 tools shows that assisted screening can cut manual screening by more than 60% while keeping recall above 90% (Mikriukov et al., 2025). On AutoLit, supervised learning models reach 82 to 97% recall, on par with a human reviewer (Kallmes et al., 2025).
Search—from question to query. A reasoning agent turns a research question into complete Boolean search strings with 76.8% and 79.6% recall on two validation sets—compared with queries designed by Cochrane experts (Kallmes et al., 2025).
Extraction—good on structured data. Extraction of PICO elements (Population, Intervention, Comparator, Outcome) exceeds 85% accuracy (Mikriukov et al., 2025); metadata such as study type, location and size are extracted with 74%, 78% and 91% accuracy (Kallmes et al., 2025).
Concrete time savings. On real reviews, the authors report 50% time saved on abstract screening and 70 to 80% on qualitative extraction (Kallmes et al., 2025).

The underlying trend is broader: scientific AI agents—able to reason, plan and chain tools autonomously—are identified as a new frontier for literature review, hypothesis generation and results analysis (Gridach et al., ICLR 2025), particularly in bioinformatics and biomedicine (Yang et al., 2025).

The pitfall you cannot ignore: unsupervised AI gets it wrong

Automating an SLR as a “black box” is dangerous. The same studies that measure the gains also measure the failures, and they are severe precisely where judgment matters most:

! Without human verification, the error rate reaches 31% for data extraction and exceeds 50% for automated risk-of-bias assessment (Mikriukov et al., 2025).
! Hallucinations: a model can invent a plausible but false value or citation—unacceptable in an evidence synthesis.
! Reproducibility: without traceability of queries, prompts and decisions, the review is no longer reproducible—yet reproducibility is the very definition of an SLR.

The authors' conclusion is unanimous: the right model is not full automation but “human-in-the-loop”—AI proposes, the expert decides. The agent becomes a tireless second reviewer whose every decision remains auditable, never a substitute for scientific judgment (Kallmes et al., 2025).

When the evidence falls short: formal consensus (RAND/UCLA)

A systematic review does not always settle the question: sometimes the evidence is scarce, heterogeneous or contradictory. Structured consensus methods then take over—and the most widely recognized is the RAND/UCLA Appropriateness Method (RAM), developed in the 1980s by the RAND Corporation and UCLA. It formalizes exactly what this article argues for: combining the evidence from a systematic review with the judgment of an expert panel.

1 Evidence synthesis—a systematic literature review is carried out and summarized for the panel, which thus shares a common factual basis.
2 Individual rating—each expert rates the appropriateness of each indication on a scale of 1 to 9, in light of the evidence provided.
3 Discussion, then a second round—after a round of discussion (modified Delphi), the experts re-rate; each indication is classified as appropriate, uncertain or inappropriate according to the median and the level of disagreement.

Charlie fits naturally upstream of this process: it speeds up the systematic review that feeds the panel—search, screening, extraction, sourced synthesis—and gives the experts a clean, traceable and up-to-date evidence base before the vote. The quality of the consensus depends directly on the quality of the review behind it: that is precisely where Charlie saves time, without taking anything away from collective judgment.

“AI does not replace the reviewer of a systematic review. It takes away the 80% of mechanical work—sorting, extracting, formatting—so that they can devote their time to the 20% that truly requires their judgment. The gain is not just speed: it is rigor regained.” — Emerit Science Team

How Charlie supports every stage—while you stay in control

Charlie is not a magic “systematic review button.” It is a scientific agent designed to support researchers at every phase, keeping track of the source of every statement:

1 Scope the search—Charlie turns your question into a search strategy, suggests synonyms and search strings, and natively queries PubMed, PMC, bioRxiv, HAL and the web. You validate and refine before launching.
2 Pre-screen—based on your inclusion/exclusion criteria, Charlie ranks and prioritizes the references and justifies each suggestion. It acts as a second reviewer: the final inclusion decision remains yours.
3 Extract in a structured way—PICO, population, sample sizes, main results: Charlie fills in an extraction grid and links each data point to its clickable citation, so you can check the source in one click.
4 Synthesize—Reflection mode cross-analyzes the included studies, identifies convergences and contradictions, and drafts a first sourced narrative synthesis that you correct and complete.
5 Track and build on your work—persistent project memory: criteria, decisions and notes carry over from one session to the next. The review stays reproducible and can be handed over to the whole team.

The guiding principle: Charlie handles the stages that research identifies as the most automatable (search, screening, structured extraction) and systematically leaves judgment—final inclusion, risk of bias, conclusions—to the researcher, with sources to back it up.

An end-to-end example

Day 1—scope and search

“I'm preparing a systematic review on the efficacy of PARP inhibitors in BRCA-mutated ovarian cancer. Help me build the PubMed search string and suggest PICO inclusion criteria.” Charlie proposes the search string, the MeSH synonyms and a PICO grid—you adjust them, then launch.

Day 2—screen

“Here are my 3,800 abstracts. Rank them according to my criteria and explain why each borderline candidate should be included or excluded.” Charlie prioritizes and justifies, and you mostly go through the gray zone—not the thousands of obvious cases.

Days 3–4—extract and synthesize

“For these 42 included studies, extract population, design, sample sizes and progression-free survival HRs, and flag any heterogeneity.” Charlie fills in the grid with clickable citations, you check doubtful values at the source, and it drafts the first synthesis—which you finalize.

Reproducibility and sovereignty: two non-negotiable requirements

A systematic review is only valuable if it is reproducible and if the data stays under control. Two structural strengths of Charlie directly meet these requirements:

Built-in traceability—every statement is linked to a clickable citation with a unified identifier, and the project memory keeps criteria and decisions. The “query → decision → source” chain remains auditable, a prerequisite for a compliant SLR.
French sovereignty—Scaleway hosting, GDPR and AI Act compliance, ProConnect for the public sector. Your review protocols and your data stay in Europe, a key point for reviews based on sensitive or unpublished data.
Native scientific sources—PubMed, PMC, bioRxiv, HAL, the web: the literature flows straight into the workspace, with no copy-pasting between tools.

Conclusion: the SLR is still your review—Charlie takes away the drudgery

The 2025 figures are clear: AI saves 50 to 80% of the time spent on the mechanical stages of a systematic review, while reaching recall levels on par with a human reviewer for screening. But the same studies point out that blind automation multiplies errors where judgment matters. The right architecture is human-in-the-loop.

That is exactly Charlie's philosophy: an agent that absorbs the workload, sources every data point, keeps track—and leaves you in charge of the science. For European teams, it is the way to speed up their reviews without compromising on rigor or on control of their data.

Your next systematic review, in a fraction of the time

Starting an SLR or a scoping review? We will show your team how to scope, screen and extract with Charlie—while keeping methodological control. A 30-minute demo on your real case.

Book a meeting “For teams” page

Share this article: