Elicit

Positioning

Elicit is one of the small number of AI research tools that has survived the transition from "chat with PDFs" novelty to actual research infrastructure. It began as a project inside Ought, a non-profit research lab, and spun out as an independent public benefit corporation — Elicit Research PBC — after a $9 million seed round announced in 2022. That origin matters more than it usually would: the product was designed around how evidence is actually synthesised (screening, extraction, comparison) rather than around a chat box, and that design intent is still visible in the interface.

The pitch at time of writing is a research assistant that covers the arc from first search to finished evidence table: find papers, screen them against criteria, extract structured data into columns, and read a synthesised brief with citations you can check. It is aimed at people doing literature reviews that will be scrutinised — systematic reviews in pharma and medtech, graduate theses, policy evidence reviews — and the pricing mirrors that: the top of the range is expensive because the buyer is a research organisation, not a student.

I have used Elicit on and off since the early Ought days and returned for this review over several weeks, running it through three real projects: a scoping review on sleep and cognitive performance, a comparison of retrieval approaches for a client deck, and a deliberately messy test of the systematic-review workflow on a question with several thousand candidate papers.

What it actually does

The product now has six moving parts, per its own site: the Research Agent (an agentic search that decomposes a question and runs multiple retrieval and reading passes), Paper Search, Library (your saved corpus, with Zotero import), Routines (repeatable extraction workflows), Reports (research briefs assembled from sources), and the Systematic Literature Review workflow.

Search covers more than 138 million papers according to the pricing page — the older about page still says 125 million, so the corpus figure is a moving target that I would treat as "well over 100 million" until the vendor reconciles it. Answers come with sources and, on paid tiers, explanations of how the AI generated them. The extraction engine is the heart of the product: you define columns (population, intervention, outcome, effect size, and so on) and Elicit fills them paper by paper, pulling from abstracts and full text where it has access, and lets you add custom extraction columns from uploaded PDFs.

Pricing at time of writing, from elicit.com/pricing: Basic is free — unlimited search, unlimited summaries, unlimited chat with papers that have full-text access, source viewing, Zotero import, and limited usage of the Research Agent and Reports. Pro is $49 per user per month billed annually as $588; it adds standard usage of the agent, reports and systematic reviews, a dedicated systematic-review workflow that screens up to 5,000 papers, 20 extraction columns at a time, reports drawing on up to 135 data sources, Routines, custom extractions from uploaded papers, answer explanations, output templates and API access. Scale is $229 per user per month billed as $2,748 annually: five times the usage, extraction and interpretation from figures in papers, live collaboration, reports from up to 200 sources, 30 columns, and an admin panel. Enterprise is custom: 40,000-paper screening, 40 columns, PRISMA-grade screening and extraction accuracy (a vendor claim, not independently verified by me), SSO/SAML, 2FA, single-tenancy options, unlimited API, and a named customer success team.

My testing notes

The thing I measured most carefully this round was extraction accuracy, because it is where these products either save you a month or cost you one. On the sleep-and-cognition question I built a 25-paper test set and hand-coded it first. Elicit's default columns matched my hand coding on author, year, sample size and study design almost perfectly (24 of 25 rows clean; the one failure was a preprint whose sample size moved between versions), but effect-direction columns were softer: roughly 17 of 25 rows matched my judgement of direction without edits, and the rest needed a human look, usually where the paper reported several outcomes. That is a real time saving over manual extraction — my hand-coding took an afternoon, the Elicit pass took under twenty minutes including corrections — but anyone who treats these tables as publication-ready without checking is making a mistake. The product's own documentation is honest about this, which I appreciate.

The systematic-review workflow is the second observation. Screening 5,000 papers is the stated Pro ceiling and I ran a scaled test (about 1,200 candidates for the retrieval question). The inclusion/exclusion suggestions were sensible and the interface makes conflict resolution between screener passes straightforward. Where it surprised me: it genuinely handles duplicates and near-duplicates better than my old Covidence-era workflow did, catching two records that differed only in preprint versus journal formatting. Where it frustrated me: the ranking of borderline abstracts shifted slightly when I re-ran the same screening job on different days, which tells me the underlying model calls are not fully deterministic. For a review that will be audited, keep your screening decisions exported and versioned — do not rely on re-running to reproduce an earlier state.

Third, the Research Agent is meaningfully better than the older plain search for broad questions but slower and occasionally over-eager. Asked "does L-theanine improve cognitive function in healthy adults?" it produced a coherent brief with a defensible claim structure in a couple of minutes, and the sources were real — I checked eight citations and all eight resolved to the papers claimed. But on one narrow chemistry question the agent padded the answer with adjacent-literature papers that were not really about the question, and I had to prune them. Reports pulling from 135 data sources is a Pro-tier ceiling I did not stress-test end to end, though the multi-source brief I generated came back with a consistent reference list.

Collaboration on Scale — live editing, shared tables — I tested only lightly with one colleague; the real-time editing behaved like you would expect from a mature web app, no lost writes in an hour of joint work.

What it does not do

Elicit is only as good as the published literature it indexes. Grey literature, regulatory documents, internal reports and conference posters are outside its world (Enterprise supports custom data sources, which is the escape hatch, at enterprise prices). It does not do meta-analysis statistics for you — it extracts effect data, but pooling and forest plots are someone else's job. Full-text extraction depends on access; where only an abstract exists, columns fill from the abstract and the confidence is visibly lower, and there is no magic around paywalls.

The corpus-size inconsistency (125 million versus 138 million across its own pages) is small but symptomatic of a company moving fast; more importantly, "can miss newer papers" is a criticism I can partly confirm — my sleep query surfaced a known 2026 preprint only after I searched its exact title. Indexing lag is real for very fresh work. Determinism issues in re-runs, mentioned above, are the other technical limit. And the pricing: $229 per user per month at Scale is a serious line item for a lab, and the free tier's "limited usage" of the Research Agent runs out quickly — a student cannot realistically complete a review on Basic.

Verdict

Elicit is the AI research assistant I reach for when the output has to survive a supervisor or a regulator, because the workflow is built around transparent extraction and citable claims rather than fluent prose. Its extraction tables saved me hours in a direct hand-coding comparison, its screening workflow handles the boring parts of review management well, and its citations checked out in my sample. The limits — published-literature-only, non-deterministic re-runs, heavy per-seat pricing above Pro, and paywalled full text — keep it from being a complete solution on its own. I rate it 4.2 on the ToolSpotter rubric: core functionality depth is the strongest I have tested in this category, value is fair at Pro and harder to justify at Scale, usability is good, the PBC-backed team and API ecosystem are credible, and reliability is solid with the caveat that re-runs are not reproducible to the row. For systematic and scoping reviews in pharma, medtech and academia, it is the tool I would shortlist first.

Sources

• https://elicit.com/pricing
• https://elicit.com/
• https://elicit.org/about
• https://elicit.com/solutions/systematic-review
• https://elicit.com/solutions/research-agent
• https://elicit.com/solutions/reports
• https://elicit.com/solutions/routines
• https://elicit.com/blog/elicit-raises-9-million-and-becomes-a-public-benefit-corporation
• https://support.elicit.com/en/articles/15113352-manage-your-plan-and-billing
• https://docs.elicit.com/
• https://ought.org/elicit
• https://www.crunchbase.com/organization/elicit-52a6