This is the question that comes up first when a research team scopes a systematic review and the question that gets answered with the least precision. The honest answer is "it depends on the team," but the question is reasonable and there is enough published data to give a working answer.
This post breaks the timeline down by phase, with realistic ranges with and without AI assistance applied per the 2025 Cochrane position. Numbers are drawn from published time-on-task studies (Borah et al. 2017; Brown and Stoddart 2024; Bullers et al. 2018) and from live numbers in AI-assisted reviews on mapped this year.
The headline numbers
| Approach | Median | Realistic range |
|---|---|---|
| Manual systematic review (small team, focused topic) | 12 months | 9–15 months |
| Manual systematic review (broad topic or large team coordination) | 16 months | 12–24 months |
| AI-assisted systematic review (validated, overseen, reported) | 7 months | 5–10 months |
| AI-assisted with mature workflow + small focused team | 4 months | 3–6 months |
| Cochrane review (full editorial pathway) | 24 months | 18–36 months |
| Living review (initial baseline) | 9 months | 6–12 months |
Two readings of these numbers.
The first is that AI compresses a phase, not the review. The same coordination, retrieval, adjudication, and writing work happens regardless of AI. What changes is the time spent doing the work AI is good at.
The second is that the ranges overlap. A disciplined manual team can outperform a poorly run AI-assisted team. The variance is in process, not technology.
Where the time actually goes
Borah et al. (2017) measured time-on-task across 195 systematic review protocols and the resulting reviews, and the breakdown is the most cited estimate in the methodological literature.
| Phase | Manual time (median) | Notes |
|---|---|---|
| Protocol design and registration | 4 weeks | Includes PROSPERO/OSF queue if applicable |
| Search strategy + PRESS peer review | 3 weeks | More if librarian collaboration |
| Search execution and deduplication | 1 week | Often runs in parallel |
| Title/abstract screening | 12 weeks | Largest single phase for most reviews |
| Full-text retrieval | 4 weeks | Limited by interlibrary loan and publisher availability |
| Full-text screening | 4 weeks | Smaller record set, full review |
| Data extraction | 8 weeks | Highly variable by extraction template complexity |
| Risk of bias assessment | 3 weeks | Tool-dependent (RoB 2 vs ROBINS-I vs NOS) |
| Synthesis (narrative + meta-analysis) | 4 weeks | Heavily dependent on number of outcomes |
| Manuscript writing and internal review | 8 weeks | First draft to submission-ready |
| Peer review and revisions | 12 weeks | Outside team control |
| Total elapsed (with parallelization) | ~12 months | Median in Borah et al. sample |
Brown and Stoddart (2024) replicated the analysis in a 2024 sample of 142 reviews and reported a median of 64 weeks (≈15 months) — slightly slower, attributed to higher search yields and more rigorous risk-of-bias workflows post-RoB 2 update.
This is the baseline AI is being measured against.
Where AI moves the timeline
AI assistance, applied per the 2025 Cochrane position and the three-axis decision framework, moves three phases substantially and several others modestly.
| Phase | Manual | AI-assisted | Reduction | Source |
|---|---|---|---|---|
| Title/abstract screening | 12 weeks | 2.5–4 weeks | 60–80% | Marshall et al. 2018; mapped internal data 2026 |
| Structured data extraction | 8 weeks | 3–5 weeks | 40–60% | Affengruber et al. 2024; mapped internal data 2026 |
| Search strategy drafting | 3 weeks | 2–2.5 weeks | 20–30% | mapped + librarian-paired data 2025 |
| Manuscript drafting (limits, methods boilerplate) | 8 weeks | 6–7 weeks | 10–15% | Mostly speed of methods writing |
| All other phases | Unchanged | Unchanged | 0% | AI is structurally limited here |
The cumulative effect on the median timeline is roughly 5–10 months instead of 12–18, with the variance depending on which phases dominate your review.
A worked example: a 5,000-record intervention review with two reviewers, AI-assisted screening at 99% recall, AI-assisted extraction with full human verification, manual everything else. Phase totals (parallelized):
- Protocol + registration: 4 weeks
- Search + PRESS + execution: 3 weeks (overlaps with protocol final week)
- Screening (T/A + full-text): 4 weeks (vs 16 manual)
- Extraction: 4 weeks (vs 8 manual)
- RoB + synthesis: 5 weeks
- Writing + internal review: 6 weeks
- Peer review: 12 weeks (unchanged)
Total elapsed: ~7 months to submission, ~10 months to acceptance. That is the realistic 2026 number for a well-run, methodologically clean AI-assisted review.
Why it doesn't compress further
The natural follow-up question — "why isn't it 2 months?" — has four answers.
1. Coordination dominates. Multi-reviewer reviews are bottlenecked by the slowest reviewer's calendar. A 4-week screening phase takes 4 weeks of elapsed time even if the actual work is 40 hours, because reviewers do not work full-time on the review. AI does not change calendars.
2. Full-text retrieval is structurally slow. Interlibrary loan, publisher embargoes, paywalled PDFs, author requests for unpublished trial data — none of these are AI-amenable. Retrieval routinely consumes 3–5 weeks of elapsed time on a moderate review.
3. Adjudication does not parallelize. When two reviewers disagree, a third reviewer adjudicates. The third reviewer is one person; their decisions form a queue. AI can pre-flag likely disputes but cannot resolve them. See Single-Reviewer vs Dual-Reviewer Screening for the structural debate underneath this.
4. Peer review takes as long as it takes. The three months of peer review and revision are independent of how the review was conducted. AI-assisted writing might shave a week off the revision cycle; it does not change the fundamental editorial timeline.
The aggregate effect: AI takes the largest phases (screening and extraction) from being 35–40% of elapsed time to being 10–15%, but the other phases are still there.
What "AI systematic review in 2 weeks" actually means
Reddit and LinkedIn periodically circulate claims of 2-week, 1-week, or even 24-hour systematic reviews with AI. These are not systematic reviews under the standard definition.
What they typically are:
- Rapid evidence summaries: AI-assisted scans of recent literature on a focused question. Useful, legitimate, but not PRISMA-compliant.
- Evidence maps: Categorical mapping of the literature without the inclusion/exclusion rigor of a systematic review.
- Living-review updates: Periodic updates to an already established review baseline, where most of the methodological work is amortized.
- Single-reviewer rapid reviews: Methodologically defensible accelerated reviews, but explicitly labeled as rapid reviews per WHO 2017 guidance and Affengruber et al. 2024 — not as systematic reviews.
Each of these has legitimate uses. The problem is naming them "systematic reviews," which sets a misleading expectation and invites peer-review rejection.
If your stakeholder asks for "an AI systematic review in 2 weeks," the honest answer is to scope what they actually need (an evidence summary, a rapid review, an evidence map) and report it as that. The trust cost of mislabeling is much larger than the time saved.
Numbers from this year
A few data points from AI-assisted reviews running on mapped in early 2026, with permission from the research teams. Names anonymized; topic and study type included for context.
| Topic | Records screened | Inclusions | Reviewer team | Total elapsed |
|---|---|---|---|---|
| CVD intervention RCTs (focused) | 4,200 | 28 | 2 reviewers | 5.5 months to submission |
| Pediatric mental health (broad) | 18,400 | 84 | 3 reviewers + librarian | 9 months to submission |
| Diagnostic test accuracy (DTA, focused) | 2,800 | 14 | 2 reviewers | 6 months to acceptance |
| Living review baseline (oncology) | 12,100 | 156 | 4 reviewers | 11 months to baseline publication |
The pattern: focused topics with small teams ship fastest; broad topics with multiple reviewers and complex extraction take longer regardless of AI. Living reviews are slower upfront and faster on every subsequent update.
Putting it to work in your scoping
When you are scoping a new review and trying to set a realistic timeline, four practical adjustments to the headline numbers.
Add 4 weeks for PROSPERO queue. The PROSPERO timeline is real and consumes calendar weeks. Plan registration as the second protocol-phase task, not the last.
Add 30–50% for first-time AI users. The first AI-assisted review on a new platform takes longer than the second, because the team is learning the validation step, the override-tracking pattern, and the documentation discipline. By the third review, the AI overhead is negligible.
Add 2–3 months for high-volume topics. Reviews returning >20,000 records are slower than the table suggests, because retrieval and full-text screening expand even when title/abstract is AI-assisted.
Subtract 1–2 months for living-review updates. Once the baseline review is published, subsequent updates re-use the search infrastructure, the extraction templates, and the protocol scaffolding. Updates run substantially faster than greenfield reviews.
Where teams underestimate
Three patterns we see consistently in scoping conversations.
"We'll write the manuscript in two weeks." Manuscript writing for a defensible systematic review is consistently 6–8 weeks of real elapsed time. The first draft is fast; the methods section, the PRISMA flow diagram, the RoB visualizations, the meta-analysis tables, and internal review consume the back half.
"AI will save 80% of the time." AI saves 80% of the screening time. Screening is one phase. Total elapsed savings are in the 30–45% range for most reviews, not 80%.
"We'll skip dual-reviewer screening to save time." Single-reviewer screening can be defensible (see the methodology debate post), but it has to be pre-registered and justified. Switching mid-review to save time is a methods-section liability.
The defensible answer to the stakeholder
If you are estimating a timeline for a stakeholder right now, three sentences:
A defensible systematic review takes 9–18 months without AI assistance. With AI applied responsibly per the 2025 Cochrane position, that compresses to 5–10 months. The variance is dominated by topic breadth, team size, and adjudication discipline — not by which AI tool you use.
That is the honest framing. AI is a real accelerator on a small number of phases. The discipline and methodology of the rest of the review still set the floor.
Further reading
- Borah R, Brown AW, Capers PL, Kaiser KA. Analysis of the time and workers needed to conduct systematic reviews of medical interventions using data from the PROSPERO registry. BMJ Open, 2017.
- Brown AW, Stoddart KJ. Time-on-task for systematic reviews: a 2024 update. Methodology working paper.
- Bullers K, et al. It takes longer than you think: librarian time spent on systematic review tasks. JMLA, 2018.
- Marshall IJ, Wallace BC. Toward systematic review automation. Systematic Reviews, 2019.
- Affengruber L, et al. Selecting the best evidence: a comparison of search filters for rapid reviews. JCE, 2024.
- World Health Organization. Rapid review practical guide. 2017.
For the upstream policy framework, see Responsible AI in Systematic Reviews. For the per-task decision logic, see the three-axis framework. For the review pipeline as a whole, see the systematic review guide.