Arakis searches four databases, dual-screens every record, retrieves and extracts the full texts, assesses risk of bias, pools with REML + Hartung-Knapp, grades certainty, and drafts the manuscript. You adjudicate only the calls it pauses on.
$100 credit when you add a card · $9 per completed review · 30-day money-back
Arakis vs a published Cochrane review · 8 dimensions × 5 points
One question, one reviewer. Scores are not comparable across rounds.
How we measuredThe problem
67 weeks
Mean time from registration to publication across 195 PROSPERO-registered medical-intervention reviews — and the authors note this understates the real elapsed time.1
What Arakis does
Runs the pipeline to completion and pages you only for adjudication.
~$141,000
Estimated cost of a single systematic review at academic institutions, based on 1.72 scientist-years of labour.2
What Arakis does
$9 per completed review; you keep the judgment, not the labour.
3,585 records
Median title/abstract screening set across 259 systematic reviews — the largest single workload in a review.3
What Arakis does
Two independent AI passes on every record; disagreement and malformed output both become MAYBE for you to decide.
67%
Of meta-analyses contain at least one trial with a data-extraction error.4
What Arakis does
Every extracted number traces to a span in the source; extraction review is a stage, not an afterthought.
42%
Median PRISMA adherence in systematic review abstracts; 0% of PROSPERO-uploaded protocols adhered completely to PRISMA-P.5
What Arakis does
PRISMA 2020 flow and a PRISMA-trAIce transparency report are generated from the run, not written from memory.
23%
Of systematic reviews are signalled as out of date within two years of publication.6
What Arakis does
Every stage checkpoints; rerun the search and the downstream stages rebuild with a before/after diff.
How it works · 15 stages, 4 phases
Boolean strategies generated from your question and executed against PubMed, ClinicalTrials.gov, OpenAlex and Semantic Scholar. Embase, Scopus and Web of Science with your institution's API key. CENTRAL and CINAHL by import.
Two independent AI passes per record. MAYBEs and conflicts stop the run until you rule. Full texts retrieved through your own library access; abstract-only records are labelled as such, never counted as success.
Random-effects pooling (REML, Hartung-Knapp), prediction intervals, Egger, leave-one-out, subgroups, GRADE per domain. Fewer than two compatible studies and Arakis says so instead of inventing a pool.
A full draft with figures and citations, every numeric claim reconciled against checkpoint data or flagged as unverified.
Evidence
Competitors quote hours saved and testimonials. We put an Arakis manuscript and a published human review in front of a reviewer who did not know which was which, on an 8-dimension methodological rubric. Here is the whole result, including the dimension we lost.
| Dimension | Arakis | Published review | Result |
|---|---|---|---|
| Question & eligibility | 4 | 3 | won |
| Search | 2 | 1 | won |
| Selection & flow | 3 | 2 | won |
| Risk of bias | 4 | 1 | won |
| Synthesis | 4 | 3 | won |
| Certainty of evidence | 4 | 1 | won |
| Interpretation discipline | 3 | 4 | lost |
| Reproducibility | 3 | 1 | won |
| Total | 27 / 40 | 16 / 40 | 7 of 8 dimensions |
Comparator
Cochrane 2024 review of zinc for the common cold (Nault et al.)
Round history
Within-round gap only — totals are not comparable across rounds
Full protocol, rubric and per-round scores are in the repository's benchmarking record. Ask for it in the form below.
Guardrails
Never downgraded to a single pass for cost.
Zero silent exclusions.
Screening cannot finalise with an unreviewed MAYBE.
meta_analysis_feasible=false with a machine-readable reason, never a placeholder estimate.
No robustness claim without a sensitivity analysis, no safety claim without adverse-event synthesis.
Institutional access is user-driven in a managed browser session; no cookies, tokens or CAPTCHA bypass.
Compare
Vendor pages, September 2026. Tell us if we got one wrong.
| Capability | Arakis | Covidence | Rayyan | Nested Knowledge | Elicit | RevMan |
|---|---|---|---|---|---|---|
| Executes the database search | Yes | No(import) | No(import) | partial | Yes | No |
| Dual AI screening with human adjudication | Yes | partial(human dual) | partial | Yes | partial | No |
| Full-text retrieval | Yes | No | No | No | partial | No |
| Extraction with source-span traceability | Yes | manual | manual | manual | Yes | No |
| Risk of bias (RoB 2, ROBINS-I, QUADAS-2 …) | Yes | manual | manual | manual | No | No |
| Pooled meta-analysis (random effects, HKSJ) | Yes | No(export) | No | Yes | No | Yes |
| GRADE certainty per domain | Yes | No | No | No | No | Yes |
| Manuscript draft with reconciled numbers | Yes | No | No | Yes | partial(report) | partial |
| Published blinded head-to-head vs a human review | Yes(one) | No | No | No | No(self-run eval) | No |
| Price | $9 / completed review | $339 / yr | free–$40 / mo | $19.95–$695 / user / mo | $12–$89 / mo | free–£560+ |
Sources (vendor pages, September 2026): Covidence · Rayyan · Nested Knowledge · Elicit · RevMan. Tell us if we got one wrong.
Try it
Pricing
Save a card to start; each completed review is $9, charged only after delivery, with promotional credit applied first.
Credit when you add a card
Charged only after completion
Enterprise or institutional pricing? Get in touch
Tell us what you're reviewing. We'll reply within one business day, and we'll only email you about the thing you asked for.
FAQ
Drop in a research question and walk away. We'll page you for the calls only a human should make.
