Job

Can an AI agent do your literature review?

Researched November 2026

Can an agent do this job today?

Supervised

Agents do parts of this job in production, with people checking the work. Merck researchers ran a PROSPERO-registered review with their ISLaR 2.0 platform and published its results: 0.91 sensitivity in full-text screening and 0.86 accuracy extracting data from study text, but 0.50 accuracy on data in tables. Elicit says Oxford PharmaGenesis, a health-science communications consultancy, used its tool with human-in-the-loop validation to deliver client literature reviews, including one covering 40 research questions across almost 500 papers. Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence state that AI in evidence synthesis should be used with human oversight and that review authors remain responsible. Fabricated and wrong citations are a documented failure: in a Deakin University study, 19.9% of citations GPT-4o wrote into literature reviews were fabricated. Every included study, extracted figure and citation needs a person's check, so this is not ready to run alone.

Abstract illustration of many blank pages passing through three mesh screens, with a few pages left as a glowing green stack

On this page

Can an AI agent do a literature review?

Most of the legwork, under supervision. An AI agent for literature review can search databases, screen papers against your criteria, extract data into a table with supporting quotes and draft a summary. A reviewer still has to verify every citation, decide the borderline inclusions and write the conclusions. Think of it as a fast research assistant whose work you check line by line, not as an author.

A peer-reviewed study reports a prospective review evaluation by Merck researchers using ISLaR 2.0, a platform built by IMO Health with the study's authors: a review on the cost-effectiveness of adult pneumococcal vaccination, registered in advance with PROSPERO, with the platform's results compared against their expert reviewers. Separately, Elicit's customer story says Oxford PharmaGenesis, a health-science communications consultancy, uses Elicit with human-in-the-loop validation for client literature reviews.

The standard-setters agree on supervision. Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence say in a joint 2025 statement that AI in evidence synthesis "should be used with human oversight" and that review authors "are ultimately responsible" for their work.

What the agent does and what stays with you

Agent

0of 8

Agent with review

6of 8

Create search strategies for the databases in scope
Run searches and remove duplicates
Screen titles and abstracts against inclusion criteria
Screen full texts and record exclusion reasons
Extract data from included studies into a table
Assess risk of bias or study quality

Human

2of 8

Verify every citation in the final review
Synthesize findings and write conclusions

8 tasks: 0 agent, 6 agent with review, 2 human

Task	Who does it	Why
Create search strategies for the databases in scope	Agent with review	Elicit's documentation says the research question guides search-strategy suggestions and offers semantic search or Boolean keyword search; the reviewer should check and approve the chosen strategy before running it.
Run searches and remove duplicates	Agent with review	In the Merck study, ISLaR searched PubMed through public APIs and removed the same duplicates the human team did; a reviewer confirms the record counts.
Screen titles and abstracts against inclusion criteria	Agent with review	ISLaR reached 0.90 sensitivity but 0.59 specificity at abstract screening in the Merck study, so a person reviews exclusions and borderline papers.
Screen full texts and record exclusion reasons	Agent with review	Elicit shows the criterion each excluded paper failed with a supporting quote, and a reviewer can override any decision.
Extract data from included studies into a table	Agent with review	ISLaR's extraction accuracy was 0.86 from study text but 0.50 from tables, so extracted values are checked against the paper.
Assess risk of bias or study quality	Agent with review	The otto-SR preprint reports high agreement on risk-of-bias judgements, but the four evidence synthesis organizations say any AI use that makes or suggests judgements must be reported, and authors stay responsible.
Verify every citation in the final review	Human	No source reviewed here shows an agent verifying its own citations reliably, and the Deakin study found 19.9% of GPT-4o's citations fabricated.
Synthesize findings and write conclusions	Human	Elicit generates a draft research report, but no source shows an agent owning a review's conclusions, and the position statement makes authors responsible for the synthesis.

The line runs between finding and judging. The sources here show agents searching, screening and extracting at speed. None shows an agent owning which evidence counts or what it means.

The Merck study gives the clearest numbers from a real review. ISLaR ran without human intervention so it could be compared with the human team. At full-text screening it reached 0.91 sensitivity and 0.87 accuracy. Extracting data from study text it reached 0.86 accuracy, but from tables only 0.50. It screened an abstract in a mean of 2.99 seconds against 66.12 seconds for a person, and the authors report it cut screening time by over 90%. Their conclusion: strong potential to reduce workload, but the platform "requires further development".

A preprint from the developers of otto-SR, a workflow built on OpenAI models, reports higher figures on health reviews: 96.7% sensitivity in screening and 93.1% accuracy in data extraction, against 81.7% and 79.7% for graduate-level human reviewers in the same study. When the authors reran 12 Cochrane reviews, otto-SR found 114 eligible studies against the original authors' 64. The paper had not been peer reviewed in the version read here.

Elicit's documentation shows what a supervised run looks like. Each excluded paper shows the criterion it failed and the reason, with source quotes behind each decision, and you can override any decision. Clicking a cell in the extraction table shows supporting quotes from the paper so you can check the AI-generated answer. Elicit claims its workflow can "save up to 80% of time usually spent on a systematic review"; that is the vendor's figure.

A worked example: one paper

This is an illustration, not a customer story. Assume, for this example, that your criteria include only studies of adults and the agent screens a cost study of a children's vaccination program.

The agent excludes it: the population is children. Its log shows the quote it relied on. Your reviewer, sampling exclusions, opens the paper and sees that it measures the effect on adults through herd protection. The paper belongs in the review. The reviewer overrides the decision, adds a line to the criteria saying indirect effects on adults count, and reruns screening on the remaining records.

This is not hypothetical as a failure type. The Merck team reports the same confusion between intervention and population in two of ISLaR's wrong full-text exclusions.

What you need to provide

Write the protocol first. The agent needs the question, inclusion and exclusion criteria, date range, languages and study designs. In the Merck study, a vague criterion produced errors until the team reworded it; the first prompt failed to exclude studies with any participants under 18.

Give it access to the databases you would search yourself and to full texts your license allows you to process with third-party tools. The agent touches published papers and your protocol. It should not touch unpublished data, patient-level data or confidential client material unless your data owner approves the tool.

Give it 5–10 papers you know belong in the review. If the search misses them, the search is wrong. Define each data field in plain words with one example value.

If you publish with Cochrane, Campbell, JBI or CEE, their statement says AI may be used as long as you can show it will not compromise the review's rigour, and any AI that "makes or suggests judgements" must be fully reported. The statement includes a reporting template for your protocol.

Job brief template

Use this as the brief for a vendor, an operator or your own team. Replace the bracketed values with your own.

Copy-ready brief

Goal: Every paper relevant to [research question] published in [date range] is found, screened against the agreed criteria and summarized in a cited evidence table within [10] business days.
Deliverable: A shared folder containing the search strings and record counts per database, a screening log with the reason for every exclusion, an evidence table with a source quote and page for each extracted value, a draft narrative summary, and a reference file with a DOI or stable link for every citation.
Inputs: The research question and a written protocol: population or scope, inclusion and exclusion criteria, date range, languages and study designs
Access to the databases in scope (for example PubMed, Embase or your licensed indexes) and your institution's full-text access
A list of 5–10 papers you already know belong in the review, used to check the search
The data fields to extract, with a plain-language definition and one example value for each
Your citation style and reference manager export format
Acceptance criteria: All of the 5–10 known relevant papers are found and included, or each exclusion is explained.
Every citation resolves to a real paper with matching authors, title, year and DOI; zero fabricated references.
On a sample of [20] extracted values, a reviewer finds at most [1] that differs from the paper.
On a random sample of [50] excluded records, a reviewer disagrees with at most [2] exclusions.
Papers the agent could not access in full text are listed, not silently dropped.
Systems: bibliographic databases (for example PubMed or Embase)
literature review or screening tool (for example Elicit)
reference manager (for example Zotero)
shared spreadsheet or document store for the evidence table
Constraints: The agent reads published papers and your protocol; it should not receive unpublished data, patient-level data or confidential client material unless your data owner approves the tool.
Respect publisher licenses: upload only full texts your institution is allowed to process with third-party tools.
If the review will be published with Cochrane, Campbell, JBI or CEE, their position statement says AI use must not compromise methodological rigour, and any use that makes or suggests judgements must be reported.
Check the vendor's data retention and model-training terms before uploading documents.
Volume: Example: 500–5,000 records screened and 20–100 papers included per review
Cadence: Example: one review per project, or a monthly update for a living review
Human checkpoints: Protocol and search strings: the reviewer approves them before searches run.
Screening: a reviewer checks a sample of exclusions and every borderline paper.
Extraction: a reviewer checks sampled values against the source page, with extra checks on values from tables.
References: a reviewer opens every citation before the review leaves the team.
Conclusions: the author writes and signs off the synthesis.

The known-papers criterion does the most work for the least effort. If the agent finds every paper you already know about, the search is probably sound. If it misses one, you learn that before you trust the rest.

Risks and human checkpoints

The review cites papers that do not exist, or real papers with the wrong authors, year or DOI.
How you notice it
Resolve every reference through its DOI or a database search before sign-off; the Deakin study found 19.9% of GPT-4o's citations fabricated and errors in 45.4% of the real ones.
The agent wrongly excludes relevant studies, so the review misses evidence.
How you notice it
Seed the review with known relevant papers and check every one is included; sample the exclusion log.
Values are extracted from the wrong place, such as a year or a figure from a table that belongs to another subgroup.
How you notice it
Sample extracted values against the paper weekly; the Merck study found most text-extraction false positives came from misattributed cost years, and accuracy on table data was 0.50.
Paywalled or supplementary material is never read, and the agent decides without it.
How you notice it
Require a list of papers read from abstract only; otto-SR's authors attribute some wrong exclusions to supplementary data the model could not access.
The evidence goes stale because the search date is not recorded or rerun.
How you notice it
Put the search date and database list in the deliverable and set a rerun date for any review that informs a live decision.
Reviewers stop checking because the evidence table usually looks right.
How you notice it
Track the error rate found in sampling; plant a known-wrong value or reference now and then to confirm reviewers still catch it.

Human checkpoints

Protocol and search strings: the reviewer approves them before searches run.
Screening: a reviewer checks a sample of exclusions and every borderline paper.
Extraction: a reviewer checks sampled values against the source page, with extra checks on values from tables.
References: a reviewer opens every citation before the review leaves the team.
Conclusions: the author writes and signs off the synthesis.

Citations are where literature reviews fail. In a Deakin University study published in JMIR Mental Health, GPT-4o was asked in June 2025 to write six literature reviews. Of its 176 citations, 35 (19.9%) were fabricated, and 64 of the 141 real ones (45.4%) contained errors, most often a wrong or invalid DOI. Fabrication was higher for less-studied disorders: 28% and 29%, against 6% for major depressive disorder. That study tested a general chatbot, not a review tool that searches databases, but it shows why no reference should leave your team unopened.

Tools that work from a database search reduce this risk, because each paper comes from a real record. They do not remove the other failures: wrong exclusions, values taken from the wrong row, and papers judged from the abstract because the full text was out of reach.

How to run a trial

Test the agent on a review you have already done. You know the right answer, so every difference is visible.

Pick a completed review with a known set of included papers and a data table.
Give the agent the original protocol, the same databases and date range, and your field definitions.
Compare its included papers with yours. Count the ones it missed and read each exclusion reason.
Sample 20 extracted values against the papers. Count errors separately for values from text and from tables.
Open every citation in its draft summary. Any fabricated reference fails the trial.
Time your reviewers' checking. That time is what you trade against screening and extraction hours.

Pick a completed review
With a known set of included papers and a data table
Give the agent the original protocol
The same databases and date range, and your field definitions
Compare its included papers with yours
Count the ones it missed and read each exclusion reason
Sample 20 extracted values
Count errors separately for values from text and from tables
Open every citation in its draft summary
Any fabricated reference fails the trial
Time your reviewers' checking
That time is what you trade against screening and extraction hours

If it passes, run it on a new review beside one human reviewer, as the second screener. The four organizations' statement points out that single-reviewer abstract screening has an estimated 13% risk of falsely excluding a relevant study, and that AI as a second reviewer could help reduce it. Widen its role one stage at a time.

For the brief, see how to write a job description for an AI agent, and for structuring the test, how to run a paid trial with an AI agent. Related pages: deep research reports, market research, competitive intelligence, the research assistant role and the research overview.

Questions buyers ask

Can I just ask a chatbot for a literature review? You can, but check every reference. The Deakin study above found nearly a fifth of GPT-4o's citations fabricated. Use a tool that searches real databases and links each claim to a paper.

What does a tool cost? Prices vary by vendor and plan. As one published example, Elicit lists its Pro plan, which includes the systematic review workflow and screens up to 5,000 papers per review, at $49 per user per month billed annually ($588 a year). Its Enterprise plan is priced on request.

Is this accepted for published reviews? Cochrane, Campbell, JBI and CEE allow it with conditions: human oversight, a justification that rigour is not compromised, and full reporting. Ask the journal or body you publish with what it requires.

Sources

Position statement on artificial intelligence (AI) use in evidence synthesis across Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence 2025
Environmental Evidence (BioMed Central), Flemyng et al.PrimaryAccessed Nov 14, 2026
Analysis of article screening and data extraction performance by an AI systematic literature review platform
Frontiers in Artificial Intelligence (Cassell et al., Merck & Co.)PrimaryAccessed Nov 14, 2026
Automation of Systematic Reviews with Large Language Models (preprint, v4)
medRxiv (Cao et al.)VendorAccessed Nov 14, 2026
Influence of Topic Familiarity and Prompt Specificity on Citation Fabrication in Mental Health Research Using Large Language Models: Experimental Study
JMIR Mental Health (Linardon et al.), via PubMedPrimaryAccessed Nov 14, 2026
Systematic Reviews in Elicit
Elicit Help CenterPrimaryAccessed Nov 14, 2026
How Oxford PharmaGenesis uses Elicit to deliver literature reviews at an unprecedented scale
ElicitVendorAccessed Nov 14, 2026
Elicit pricing
ElicitVendorAccessed Nov 14, 2026

What we could not verify

The Merck/IMO Health evidence is a prospective evaluation reported by the researchers involved; the Oxford PharmaGenesis account is Elicit's vendor-published customer story. No independent audit of an agent-produced literature review in commercial use was found.
Elicit's claim that its workflow saves up to 80% of systematic review time is the vendor's own figure; the method behind it is not published on the page cited.
The otto-SR results are from a preprint by the tool's developers and had not been peer reviewed in the version read.
The accuracy figures come from health and medical reviews. No comparable study was found for literature reviews in business, engineering or the social sciences.
Agent pricing is shown for one vendor only, and no human wage benchmark was researched for this job, so the page has no cost comparison.
A peer-reviewed study comparing Elicit and ChatGPT with human reviewers for data extraction (Cochrane Evidence Synthesis and Methods, 2025) could not be opened from the publisher's site during research, so it is not cited.

Hiring AI agents, opening soon

One email when the job board opens. Optionally tell us the job you want done.

I build or operate AI agents