Elicit

What Elicit is

Elicit is a research platform combining scientific literature search, data extraction and referenced outputs. Its three main entry points are Research Agent, Research Reports and Systematic Reviews: respectively, a flexible session, a topic overview and a more structured review process.

The unit of work need not be a conversation. It can also be a table in which each row represents a study and each column a defined question. This structure makes it possible to discuss what was found, what is missing and where each value came from before writing conclusions.

Reference documentation: Which Elicit workflow should I use?.

Origins, company and development

Elicit originated at Ought, a nonprofit research organisation. In September 2023 it announced its independence as a public benefit corporation and a nine-million-dollar seed round led by Fifty Years. Its cofounders are Andreas Stuhlmüller, chief executive, and Jungwon Byun, chief operating officer.

The original project proposed breaking reasoning into tasks that could be supervised, such as search, classification and summarisation. Ought also released Interactive Composition Explorer, an open library for exploring language-model programs. This background does not make Elicit’s complete commercial service open-source software.

• 2024: the team explained its research into semantic search and lexical expansion using SPLADE.
• 2025: it expanded review and report workflows and announced a 22-million-dollar round in February at a 100-million-dollar valuation. Keyword search followed in October.
• 2026: it introduced API and MCP capabilities in July, an expanded Research Agent in August, and collaboration, shared-library and routine features during August and September.

This evolution matters when interpreting research on the product: an evaluation of an earlier version describes that configuration, rather than every subsequent capability.

Reference documentation: Elicit raises $9M and becomes a PBC · Elicit Series A · Elicit team · Elicit at Ought · From grep to SPLADE · Introducing Keyword Search · Elicit API and MCP · Introducing Elicit Research Agent · AI-powered research is now multiplayer · Elicit Library shared research hub · Introducing Routines.

Semantic search and reproducible searches

Semantic search attempts to retrieve work related to an idea even when it does not use the query’s exact words. In 2024 Elicit described its work with SPLADE, which expands lexical representations to combine meaning relationships with more interpretable retrieval. This is a dated engineering account, not an unchanging specification of today’s engine.

Keyword search lets users edit a proposed query, inspect its history and work across Elicit’s paper bank, PubMed or ClinicalTrials.gov. The query used can appear in the report’s methods. This helps document what was searched and refine synonyms or terms producing irrelevant results.

A useful strategy starts with known papers that should be retrieved. If they are missing, review concepts, restrictions and sources before increasing volume. Finding many publications does not establish that the relevant evidence has been located; retrieving a few highly relevant results does not establish completeness either.

Reference documentation: From grep to SPLADE · Introducing Keyword Search.

Sources it can use

The current Research Agent page advertises access to over 138 million papers and more than 545,000 trial records. It also describes uploaded documents and public sources. The record count does not indicate how many full texts are accessible or guarantee uniform coverage across disciplines.

The expansion documented in November 2026 includes patents, Open Targets, DepMap, openFDA information and ChEMBL. These sources represent different kinds of material: a patent describes claims, a reporting database collects submissions, and a published trial provides findings from a particular study design.

To investigate a prevention question, separate these evidence types before synthesising them. A trial registration without results should not count as a demonstration of effectiveness, nor should a spontaneous report be treated as an incidence rate. Provenance should remain visible even when several sources appear together in one table.

Reference documentation: Elicit research agent · Where Elicit looks beyond research papers.

Screening and structured reviews

Systematic Reviews organises setup, collection, title-and-abstract screening, optional full-text screening, extraction and reporting. Setup accepts a question, criteria and context; Fast and Thorough modes adjust effort. Some advanced options depend on the plan and consume additional usage.

Before screening, turn each criterion into a decision that can be checked. “Adult workers performing office tasks” is more interpretable than “relevant population”. Decide what happens when an abstract omits a detail: excluding it for missing information can remove an eligible study whose full text supplies the answer.

A defensible review retains decisions and reasons. If a criterion changes after results have been seen, recording the change and reassessing affected earlier papers prevents selection from depending on preferences absent from the protocol.

Reference documentation: Systematic Reviews in Elicit.

Turning papers into an evidence table

In extraction, each column question should specify exactly what is sought. “Primary outcome at the final follow-up” and “any outcome mentioned” will produce different tables. Sabentis proposes separating population, design, size, intervention, comparator, duration, measure and limitations, adapting fields to the actual question.

A numerical value needs context: group, time point, unit and denominator. A cell containing “20” is unusable if the reader cannot tell whether it means participants, a percentage or weeks. Retaining a supporting passage and study identifier helps resolve discrepancies without repeating the entire search.

Checks should include difficult cases, such as multiple arms or follow-up periods, before processing the whole collection. Repeated errors can indicate a poorly defined column as well as an extraction failure. Correcting the definition and reassessing affected papers is more consistent than fixing isolated cells.

Reports and working with agents

Research Reports provides an organised synthesis of a question; Research Agent supports iteration with the material and outputs such as tables, figures or presentations. The August 2026 announcement describes an environment that combines sources and generates results with traceable references.

Choosing the output format at the start makes the work more useful. For a decision meeting, a table separating finding, uncertainty and implication may be more useful than a long report. For a formal review, the synthesis must retain the study selection and methodological decisions supporting it.

When reviewing a report, follow the chain from conclusion to study: a reference can identify the right paper without supporting the precise accompanying claim. Separating the authors’ results, the agent’s interpretation and the team’s interpretation avoids attributing a later-added conclusion to the paper.

Reference documentation: Introducing Elicit Research Agent · Elicit Reports · Which Elicit workflow should I use?.

Library, projects and collaboration

The library supports personal and shared collections that Research Agent can consult and expand. Projects can be associated with those collections so that teams retain reference sources alongside the context of their work.

The 2026 collaboration features add shared sessions, output editing and version history. Skills stores instructions for a reusable method; collaborative sessions and shared editing are restricted to particular organisational offerings.

For this continuity to be useful, a collection should state its scope and review date. Saving a paper means it is available to the team, rather than that it has passed critical appraisal. Distinguishing “awaiting assessment” from “included in the synthesis” prevents a growing library from becoming a list of implied conclusions.

Reference documentation: Elicit Library shared research hub · AI-powered research is now multiplayer.

Routines, API and integration

Routines, announced in September 2026, configures recurring daily or weekly work. Users review instructions and schedules, can pause a routine and inspect its results in an agent session. This provides a way to update research whose evidence changes over time.

The API and MCP server offer search, reports and systematic reviews for external workflows; access is tied to Pro and higher offerings. Outputs can be incorporated into other applications while retaining references and process states.

For literature monitoring, Sabentis proposes distinguishing a “new paper” from “evidence that changes the conclusion”. A routine should record the period searched, duplicates and results that leave the conclusion unchanged. For integrations, execution limits and a review before distributing conclusions help prevent an automatic update from being mistaken for an approved technical decision.

Reference documentation: Introducing Routines · Elicit API and MCP.

Access, plans and languages

The offering reviewed distinguishes Basic, Plus, Pro, Scale and Enterprise, with features varying by segment and subscription. Basic supports exploration; systematic reviews, exports, collaboration and additional capacity depend on the plan. Budget for the complete search, extraction and revision workload, rather than only the first query.

Elicit is optimised for English documents and English-language use. Its help centre warns that quality in other languages varies and may consume more usage; it is not presented as a translation service. This profile is available in Spanish, Catalan and English, which does not imply equivalent product coverage in those languages.

For topics with relevant local literature, retain searches in the original language and document sources outside the main workflow. Translating only the question does not by itself resolve a coverage gap.

Reference documentation: Elicit pricing · Languages other than English.

What published evaluations show

Figures need to be read by task. In March 2025, Elicit published a screening evaluation based on 58 reviews, reporting 93.6% sensitivity and 62.8% specificity. This assessed classification of a paper set, rather than the ability to find every paper from scratch.

In an independent 2025 study covering four evidence syntheses, Lau and Golder observed average search sensitivity of 39.5% for Elicit, compared with 94.5% for the original reviews. It also found additional studies. The authors considered it complementary under the configuration evaluated.

Bianchi and colleagues compared extraction across 20 trials and seven variables. Information was particularly incomplete for nuances of intervention effects: 95% of those fields were rated partially equivalent. Their conclusion called for human checks of completeness and accuracy.

The company’s own Reports evaluations acknowledge small samples, paid evaluators and reuse of evaluators across rounds. These conditions should accompany any reading of their findings.

These studies can inform a local evaluation but cannot assign one accuracy figure to today’s product. Retrieval, exclusion, extraction and synthesis are different stages. A pilot should identify where errors arise, the effort required to correct them and what information would be lost without review.

Reference documentation: How Elicit evaluated Systematic Review · Lau and Golder: Four-case comparison · Bianchi et al.: Elicit versus human extraction · How Elicit evaluated Reports.

Data, security and usage rights

In October 2025 Elicit announced a SOC 2 Type II report available under confidentiality and described account and sharing controls. Its published policy covers account and usage data, retention and third-party services. Sharing a read-only link and making a report publicly discoverable are different decisions.

The API-specific terms prohibit sending personal data and state that input data is not used to train models without a separate opt-in. They also note that third-party sources retain their own licences. These conditions should not be confused with general permission to upload internal case files.

For an occupational safety review, published papers and anonymised examples reduce unnecessary dependencies. If a team wants to incorporate confidential documents, it should check the product, contract, permissions and output recipients first. A service security control does not establish the methodological quality of a report.

Reference documentation: Elicit SOC2 announcement · Elicit API terms · Elicit privacy policy.

Applications in occupational safety and health

For occupational safety and health, it can support reviews of ergonomic interventions, work organization or exposure to agents by helping prepare evidence tables. The team should define its question and inclusion criteria before interpreting results, reducing the risk of selecting only studies that favor an existing hypothesis.

Proposed Sabentis uses include comparing training interventions, examining organisational changes or preparing a table of exposure-assessment methods. In each case, separate population, intervention or factor, comparator and outcome. Two studies may use the same word for different measures: satisfaction with training is not the same as learning, and learning alone does not demonstrate fewer accidents. Define each extraction field before reading the results so attractive findings do not quietly change the question being reviewed.

Worked example and checks

For example, a researcher can find studies on breaks and musculoskeletal discomfort among office workers. They extract population, intervention, comparison and outcomes, then read the full texts to determine which conclusions fit their context and which remain uncertain.

For the review of breaks and musculoskeletal discomfort:

1. Formulate a bounded question and define population, task type, intervention, comparator and outcomes beforehand. Record exclusion criteria.
2. Find papers and resolve duplicates before screening. Retain searches, dates and decisions so the process can be reviewed.
3. Pilot extraction columns on a small set of articles: participants, duration, outcome measurement, comparator and stated limitations.
4. Open passages supporting each data point. Distinguish baseline from follow-up measurements and intervention from comparison groups.
5. Review relevant full texts and methodological quality before synthesising. Keep non-comparable studies and conflicting findings visible.

A useful output is a table with verifiable provenance and a synthesis explaining uncertainty. Counting favourable papers or combining results using different scales, durations or populations is insufficient. Where multiple reports describe the same underlying study, identify that relationship rather than treating every publication as independent evidence.

Interpretation and reproducibility limits

Elicit acknowledges that it may miss nuance and confuse what a figure refers to. It also distinguishes summarising findings from assessing a study’s methodological trustworthiness. An accurately extracted value can come from a design insufficient to support the conclusion of interest.

The same question can generate different outputs because of changed filters, updates or model variability. Retaining dates, criteria, queries, studies and the exported version makes it possible to reconstruct what informed a decision even when a later run changes.

Where only the abstract is available, say so and leave fields requiring the full text unresolved. Failure to find evidence does not demonstrate absence of an effect. The report should explain its gaps and the additional work that could reduce them.

Reference documentation: Elicit's limitations · Why Elicit answers vary.

Choosing and complementing the tool

Elicit fits work requiring an evidence table with defined variables and an explicit selection process. To explore how later research has used a paper, it can complement Scite. To review and consolidate extracted data, Excel supports independent checks and annotations.

In a pilot, compare automated extraction with a manually reviewed reference. Record errors by field, missing information and doubtful exclusion decisions. Assess time savings alongside the effort needed to correct and justify results. The pilot should include difficult tables or incomplete reporting, not only straightforward articles.

Frequently asked questions

Does the Systematic Reviews workflow make a review systematic?

The tool supplies structure, but quality depends on the protocol, search coverage, screening and study appraisal actually performed by the team.

What does an empty information cell mean?

The information may not have been found or available in the accessible material. It does not automatically mean that the study did not measure the variable.

Can it directly recommend an intervention for my company?

It can organise relevant evidence. Selection requires assessing applicability, local conditions and study quality, alongside the appropriate technical review.