eDiscovery AI vs Everlaw: how they compare in 2026
eDiscovery AI and Everlaw both use large language models to review litigation documents, but they are bought differently. eDiscovery AI is a plug in that classifies documents for relevance and privilege from inside Relativity; Everlaw is a complete ediscovery platform with its own review AI built in. Everlaw sits in the top two bands on nine of fifteen axes and eDiscovery AI on four of fifteen. The gap is almost entirely about data handling. Everlaw publishes an AI governance framework, states that customer data does not train models and that its language models run under zero data retention, discloses that document embeddings persist in a vector database, and holds FedRAMP Moderate authorization covering its generative features. eDiscovery AI publishes a privacy notice limited to its website and website terms with no data provisions, and nothing on how it handles the collections a firm sends it. Its counterweight is the review output itself: each classification restates the instruction it followed and explains the call, privilege results name the attorneys and the privilege elements present and draft a log entry, and it runs inside the Relativity workspace a team already uses.
At a glance
All 15 axes, side by side
The same grid applied to every vendor in the index, graded from public sources. Hover a grade to see what the letter means on that axis.
AI Centrality
How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.
The models are the product and there is nothing underneath them to license, which is the A band. Every capability the vendor sells is a classification or extraction produced by a language model: relevance coding against written instructions, privilege determination with named attorneys and privilege elements, early case intelligence, detection and extraction of personally identifiable information, and image comparison, filtering and facial recognition. There is no repository, processing engine, review platform or production tool being sold alongside them. The delivery architecture settles it rather than leaving it to inference: the product is a plug-in that receives a document set from Relativity, classifies it, and writes results back to a mapped field, so the platform of record belongs to somebody else and what this vendor supplies is the judgement applied to the documents. The vendor's own framing matches, positioning AI review as the successor to Boolean search and technology-assisted review rather than as a feature added to either. Verified 12 September 2026.
The models are the engine of a core capability layered on a platform that would function without them. Everlaw is an end to end ediscovery system covering upload, processing, search, review, production and trial preparation, and that platform existed and sold before the generative layer arrived in autumn 2024. What the models drive is substantial rather than peripheral, which is why this is not a C: Coding Suggestions automating first pass review, Deep Dive answering natural language questions across millions of documents, Review Assistant, Writing Assistant, custom extractions and deposition analysis. Fifth B on this axis. Worth recording for a reader that ediscovery has used machine learning for predictive coding long before generative AI, so the underlying platform was never model free, but the reviewable document system stands without the generative layer graded here.
Citation Accuracy and Hallucination Disclosure
Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.
Grounding is real and documented per document, and a figure is published without the test set that would make it testable, which holds this at B. The grounding is unusually concrete for this corpus because it is built into the output rather than described in the abstract: Relevance restates the instruction it was given when making each classification and then explains why it categorised the document as it did, and it does so consistently across every document and every classification. Privilege goes further, naming the attorneys it identified in a separate field, spelling out which elements of the privilege are present so a reviewer can see the strength of the call, and producing its reasoning as a draft privilege log entry. The published workflow ends in validation, the fourth step being to review the results and generate industry standard performance metrics. A figure is published: consistently achieves 90 per cent recall with excellent precision. **It does not reach A because the A limb asks for measured accuracy with the test set described and the failure modes named, and neither appears.** No corpus, sample, matter type or methodology is given for the recall figure, precision is characterised rather than quantified, and no failure mode is identified. The Pre/Dicta record at R15 is the contrast: it published a figure, described a test set of more than 50,000 motions across 94 districts, and named its own limits. Verified 12 September 2026.
Grounding is real and documented with the retrieval architecture described, short of published figures. The mechanism is stated concretely: Deep Dive is powered by a vector database holding embeddings of the customer's own documents, answers are synthesised from facts extracted from specific documents, and results cite source documents with access to the underlying materials so users can verify outputs when checking their work. The vendor states plainly that it has taken steps to reduce hallucinations. Coding Suggestions is claimed to deliver recall and precision that rivals eyes on review, which uses the right metrics for the task and is the correct frame for first pass review quality. Searched the product pages, the AI Assistant framework page, the support knowledge base and the vendor blog on 29 Aug 2026 and located no numeric recall or precision figure, no accuracy rate, no test set, no published evaluation methodology and no independent benchmark participation. A claim to rival eyes on review with no measurement attached is exactly the kind of claim this axis exists to mark.
Autonomy and Oversight Model
What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.
The review surfaces and the route back to human judgement are published as part of the workflow, short of any stated threshold, which is B. Oversight is not asserted, it is built into the four published steps: the reviewer specifies the criteria, the model classifies, and step four is to review results to validate document classifications and generate industry standard performance metrics. That is a validation gate with a measurement attached, which is more than most records in this corpus offer. Underneath it, every classification carries a restated instruction and an explanation, so a reviewer can audit a call rather than accept it, and privilege output is expressly a draft, the log entry field being available to use or edit as needed. The named attorney field exists precisely so results can be validated. What is missing for A is the boundary. Nothing published states a confidence threshold, describes what the system does when a document is ambiguous, sets out an escalation path, or says what proportion of a set a reviewer should sample. No mode structure is described, and nothing distinguishes a classification the model is confident in from one it is not, which matters on privilege where a single wrong call can waive. Verified 12 September 2026.
A real published commitment with a documented verification surface, short of thresholds. Control is the first of three stated generative AI principles, alongside Confidence and Transparency, Privacy and Security. The verification path is described rather than asserted: the assistant cites source documents in its results for users to reference when checking their work so they can ensure accuracy and verify outputs, and Deep Dive provides access to underlying source materials. Architectural limits reinforce it: no link following or web browsing, and every task runs as a one time generation request rather than a persistent agent. Not located as of 29 Aug 2026: what proportion of a first pass review Coding Suggestions is expected to decide unaided, any threshold at which a document routes to a human, and what the vendor commits to when an output is wrong. That last matters here because the vendor claims automation rivalling eyes on review, which is a claim about replacing a human step rather than assisting one.
Operational and Outcome Evidence
Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.
The headline result is quoted with no basis stated, and no customer is named on the surfaces read, which is C. The published claim is a performance one rather than a deployment one: consistently achieves 90 per cent recall with excellent precision, and reviews across hundreds of thousands of documents in hours at a fraction of the time and cost of manual review. No matter, no client, no engagement and no volume is attached to any of it, and no method is stated, so a reader cannot tell whether the figure comes from a benchmark, a customer project or an internal test. One structural fact is real and is recorded rather than credited as a customer: the product is listed on the Relativity App Hub and delivered as a Relativity plug-in, which is a partner relationship with the dominant platform in this lane and is graded on the integration row. Named here as the route by which this row would move and not opened under the sufficiency discipline: the vendor publishes a case studies category alongside fact sheets, whitepapers, videos and books, and a named engagement with a figure and a method in any one of them would lift this to B or A. Verified 12 September 2026.
Adoption figures stand where deployment outcomes would go. Published and specific: EverlawAI Assistant was developed with input from nearly 3,000 users, more than 150 customers use the generative features, and the platform is used by numerous United States government agencies including United States attorneys' offices. Those are real adoption numbers rather than logo walls, and the government adoption is corroborated by the FedRAMP and GovRAMP authorisations, which are matters of public record. But adoption is not outcome. Searched the product pages, the vendor blog, the newsroom and the support knowledge base on 29 Aug 2026 and located no named customer paired with figures and a date, no case study a reader could assess, and no measured result from any deployment.
Privilege and Confidentiality Posture
How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.
Nothing is published on how client confidences are handled by a product built to ingest them, which is the D band, and the absence is established rather than untested. Both published documents were read in full on the date shown. The Privacy Notice, last updated 8 January 2025, scopes itself in terms: it applies only to information collected on or through the vendor's websites or through other interactions with the vendor and its clients. It addresses names, addresses, identifiers, browsing data and marketing preferences. It says nothing about the customer documents submitted for classification. The Terms and Conditions, last updated 23 May 2024, are a generic website terms-of-use template covering content ownership, acceptable use, service interruption and severability, and contain no data provisions at all. So nothing published addresses training on client documents, retention of a reviewed collection, segregation between matters or between clients, or what any model provider may retain. **The irony belongs on the record as a fact rather than a charge: the product's principal function is to identify privileged material so it can be withheld, and the vendor publishes no statement about how it protects the privilege of the documents it reads.** Verified 12 September 2026.
The most complete confidentiality posture on the index, and the only one where an outside authority has tested the generative features specifically. All three limbs are addressed in terms a buyer can read in advance. Training: the vendor states data submitted and responses received are not used to train models across customers and are not shared between customers, and separately that the enterprise language models it uses adhere to a zero data retention policy under which data is used solely to generate a response, deleted on completion, and never used to train their models. Segregation: stated between customers, and reinforced architecturally because the models hold no persistent knowledge of a customer's case. Retention: disclosed honestly and at the right level of detail, including the part most vendors would omit, that the vector database powering Deep Dive stores numerical embeddings created from customer documents and that this storage is necessary for retrieval. Privilege is named directly rather than implied: the vendor describes environments where sensitive case data, privileged communications and personally identifiable information must be safeguarded, and FedRAMP authorisation of the generative features means a third party assessment organisation has tested that claim. Short of nothing material on the three limbs; the one gap, no stated retention period for the vector database, is recorded on the retention signal row.
UPL and Professional Responsibility Posture
Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point. Where the advice line is not the duty a product raises, the axis is read through the nearest professional duty it does raise: judicial conduct rules and the reviewing duty for products sold only to courts, and the duty to bill for time actually spent for products that draft time entries.
Nothing published addresses the advice line for a product that makes privilege determinations, which is the D band. R15 governs which limbs bite, and one of them is answered: the audience is unambiguous, the product being sold to legal teams and eDiscovery service providers and delivered inside Relativity, so there is no consumer-facing ambiguity to resolve. Every other limb is unaddressed. No statement was located that the vendor is not a law firm, that a classification is not a legal determination, or that a privilege call requires attorney review before it is relied on to withhold a document. Nothing addresses supervision, competence or the reviewer's own professional obligations, and the two published documents, both read in full, contain no disclaimer of any kind on the point. That gap is sharper here than on most records in this corpus, because a privilege designation is a legal conclusion with a waiver consequence attached, and the published workflow puts the model in the position of making it first. The validation step in the workflow does place a human after the machine, and it is graded on the Autonomy row as a control rather than as a professional-responsibility position, which is what it is. Verified 12 September 2026.
The audience is professional and the position is unstated. Users are law firms, corporate legal departments and government legal teams including United States attorneys' offices, with no consumer surface located, and the product is a document review platform rather than an advice tool, so the advice line question arises less sharply here than for a research or drafting product. Searched the product pages, the AI Assistant framework page, the support knowledge base and the vendor blog on 29 Aug 2026 and located no published position on advice versus tooling, no treatment of competence or supervision duties, and no jurisdiction limits. Recorded at C because the position is inferable from what the product is rather than published.
AI Governance and Bias Disclosure
Published governance over model behavior: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
No governance position is published for a system whose output affects legal outcomes, which is the D band. There is no responsible AI page, no set of principles, no framework named, no certification claimed, nobody identified as accountable for AI decisions, no description of what is evaluated before a model changes, and no testing regime beyond the per-project performance metrics a customer generates for itself. Neither published document touches AI at all. Bias is unaddressed and is squarely live on this estate rather than theoretical. The vendor ships **facial recognition** as a named product feature, alongside detection and extraction of personally identifiable information, applied to litigation collections that contain images of real people; facial recognition is the single AI capability with the most documented demographic performance disparity, and nothing published addresses accuracy across populations, thresholds, or what a false match in a document review would mean. The relevance and privilege classifiers raise a quieter version of the same question, since a model that reads correspondence to infer legal significance can perform unevenly across writing styles, languages and communication formats, and nothing published considers it. Verified 12 September 2026.
Third A on this axis and the first earned on bias disclosure rather than certification alone. The vendor publishes a named AI Governance Framework as a standing document, structured on three stated generative AI principles of Control, Confidence, and Transparency, Privacy and Security. The framework describes specific governance decisions with their reasoning: third party technologies are evaluated from technical, legal, privacy and security perspectives before adoption, subprocessors including each language model are vetted with the appropriate internal teams, tasks run as one time generation requests, and link following and web browsing are excluded to keep the system closed loop, with the stated rationale that a model holding no persistent case knowledge cannot have it exfiltrated by prompt or data injection. Independent validation exists and is specific to the AI: a third party assessment organisation tested the generative features for FedRAMP authorisation, with an attestation letter, as part of the 2025 annual assessment. And the vendor makes the bias disclosure almost nobody on this index makes, stating plainly that if the case materials contain bias or toxicity that may be reflected in the output. Not located: pre release testing results, and a named individual owner of model governance.
AI Safety and Data Stewardship
Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.
Nothing is published on retention, deletion or access for a system that holds client documents, which is the D band, and both documents that would carry it were read in full. The Privacy Notice's security section is a single paragraph and is weaker than the norm rather than merely thin: it says the vendor takes a number of steps to protect against loss and misuse of information under its control, warns that no internet transmission is completely secure, states that information is sent at the user's own risk, and disclaims responsibility for illegal acts such as hacking by third parties. No encryption, access control, monitoring or incident practice is described anywhere. Its retention section is criteria-based rather than periodic and is scoped to personal information collected through the website, not to a reviewed document collection. There is no subprocessor list, no security page, no trust centre and no data processing addendum on any surface, and the site navigation, which renders in full and is therefore the page inventory under R20, contains no security page at all. The Terms and Conditions contain no data provisions. On a product that ingests entire litigation collections, that is the whole published record. Verified 12 September 2026.
Substantive published policy covering most of the ground, on an unusually broad certification base. Published: zero data retention at the language model layer with data deleted on completion, a closed loop architecture excluding link following and web browsing, one time generation requests rather than persistent agents, subprocessor vetting including each language model, in region AI processing for United Kingdom and European customers, and a certification set spanning FedRAMP Moderate, GovRAMP Moderate, SOC 2 Type 2, ISO 27001:2013, Cyber Essentials Plus, GDPR, HIPAA and UK G-Cloud. The vendor also discloses that the platform is developed and operated by United States citizens with its research and development team in California, which is a personnel sovereignty statement few vendors make. Not located as of 29 Aug 2026: a retention period or deletion control for the vector database that stores customer document embeddings, a published subprocessor list as distinct from a statement that subprocessors are vetted, and an incident or breach notification practice.
AI Liability and Recourse
What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.
Nothing is published on who bears the loss when the system is wrong, which is the D band, and the way that comes about is worth recording precisely because it is unusual. The Terms and Conditions were read in full. They run to about a dozen short clauses covering content ownership, access to external resources, acceptable use, no waiver, service interruption, service reselling, intellectual property, changes to terms, assignment, contacts and severability. **They contain no limitation of liability, no disclaimer of warranties, no indemnity and no cap of any kind.** Most records in this corpus sit at C because a standard limitation clause disclaims the exposure the product creates; here even that is absent, so nothing is allocated in either direction. Nothing addresses what happens if a relevant document is coded non-responsive and not produced, or if a privileged document is classified as non-privileged and produced, which is the failure mode with the largest consequence in this workflow and the one the product exists to prevent. No warranty attaches to the published 90 per cent recall figure. No insurance position appears. No customer agreement is published on any surface located. Verified 12 September 2026.
Searched the product pages, the AI Assistant framework page, the support knowledge base, the vendor blog and the newsroom on 29 Aug 2026. No published indemnity, liability cap, carve out, warranty on output or insurance position was located, and no customer terms of service or master agreement was located on the surfaces reached. Recorded as a pure absence on those surfaces. Rebuttable, and worth noting the shape: this vendor publishes more than almost anyone on this index about how its AI is governed and secured, and nothing about who bears the loss when an output is wrong. Governance and recourse are different questions and only the first is answered.
Practice Systems Integration Depth
How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.
One integration, and it is the one that matters in this lane, documented at the level of what actually moves. The product is delivered as a Relativity plug-in and is listed on the Relativity App Hub, which is a reviewed partner channel rather than a self-declared compatibility claim. What the integration does is described as a workflow rather than named: the reviewer selects a document set inside Relativity, supplies relevance, issue, privilege and category specifications, submits the documents to eDiscovery AI through the plug-in for classification, and receives results written back to a mapped field. The vendor states that the only configuration required is the prompt instructions and the field to map results to, and that no additional indexing or model training is needed to run the same instructions against a further document set, which is a real statement about how the connection behaves in use. What holds it off A is breadth and documentation. No other platform is named, so a firm on Everlaw, Reveal, DISCO or Nuix has nothing published to work from; no API, developer documentation or connector specification was located; and nothing describes authentication, volume limits or how results are versioned. Verified 12 September 2026.
Searched the product pages, the AI Assistant framework page, the support knowledge base entry point and the vendor blog on 29 Aug 2026. No integrations page was reached, no named connector for document management, practice management or productivity tools was located, and no API documentation was located. Recorded as an absence on the surfaces reached rather than as a finding about the product. Two things a reader should weigh. First, the platform is deliberately end to end, covering upload through production and trial preparation, so it substitutes for a workflow chain rather than connecting into one, which reduces how much integration it needs. Second, this axis was not the focus of the research pass and an ediscovery platform of this scale almost certainly documents data ingestion routes somewhere. Rebuttable with a single link and flagged for revisit.
Deployment Model and Data Residency
Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.
Cloud delivery is implied and neither the tenancy model nor the region is stated, which is C in its exact terms. Delivery is clear enough in outline: documents are submitted from a customer's Relativity workspace to eDiscovery AI for classification and results return to a mapped field, so processing happens on the vendor's side and the product is a hosted service reached through a plug-in. Nothing beyond that is published. No cloud provider is named, no region or data centre is identified, no residency option is offered or refused, no tenancy model is described, and nothing states whether a customer's collection is isolated from any other. Nothing distinguishes where documents are held during processing from where model inference happens, and no on-premises or in-tenant option is mentioned. The Privacy Notice contains the only geographic statement located, that the vendor operates in various countries including the United States and countries in the European Union and European Economic Area and that personal information may be accessed from countries outside the United States. That is a statement about website personal information rather than about a document collection, and it is recorded rather than credited. Verified 12 September 2026.
Deployment model is stated clearly with real residency options, short of the full picture. Cloud native delivery is stated throughout, and residency is offered concretely rather than described: in region AI processing is available for United Kingdom and European customers, introduced explicitly to meet data security and sovereignty requirements and keep data within local boundaries, with the vendor confirming the same zero retention and no training terms apply as in the United States implementation. Government deployment is authorised separately under FedRAMP Moderate and GovRAMP Moderate in the United States and G-Cloud in the United Kingdom. Personnel location is also disclosed, the platform being developed and operated by United States citizens with research and development in California. Not located as of 29 Aug 2026: a tenancy model, a named hosting provider, an enumerated region list, and any statement of where processing happens for customers outside the named regions.
Security Certifications and Trust Center
Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.
No independent security attestation was located, and the absence is established through the page inventory rather than assumed. The site navigation and footer render in full and are therefore the inventory under R20: Solutions, Partners, Thought Leadership, Company, Contact, with Privacy Policy and Terms and Conditions in the footer. There is no security page, no trust centre and no compliance page. Note for a future pass that the seed list recorded a security page on the site as present but unconfirmed; **it was checked and no such page exists in the current navigation**, and the Privacy item in the Solutions menu is a product for detecting and extracting personally identifiable information, not a policy surface. No certification of any kind is claimed anywhere: no SOC 2, no ISO 27001, no HIPAA, no badge, no auditor, no report and no request route. No penetration testing or vulnerability disclosure programme is described. The only security statement on the estate is the Privacy Notice paragraph graded on the stewardship row, which describes no control and disclaims responsibility for third-party hacking. There is accordingly no access flow to grade under R5, because there is nothing to request. Verified 12 September 2026.
The strongest certification position on the index, and the only one where the generative AI features themselves have been independently authorised. FedRAMP Moderate Authorization and GovRAMP Moderate Authorization are held, and critically the generative features were included in the 2025 FedRAMP Annual Assessment where a third party assessment organisation independently tested and endorsed the authorisation, with an attestation letter issued. FedRAMP is a continuous authorisation regime with annual reassessment rather than a point in time report, and its status is a matter of public record rather than a vendor claim, which is a materially stronger form of evidence than any private attestation elsewhere on this index. Alongside it: SOC 2 Type 2, ISO 27001:2013, Cyber Essentials Plus, GDPR and HIPAA compliance, and UK G-Cloud status. The vendor is stated to be the first ediscovery provider to have its full portfolio of generative AI features FedRAMP authorised. Short only of a published report request route or trust portal, which was not located on the surfaces reached.
Model Supply Chain Disclosure
Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.
The vendor refers to advanced AI without identifying anything underneath it, which is C. The language across the estate is capability-level and consistent: advanced AI technology, powerful and efficient AI technology, generative AI, and AI review technology that understands grammar and context. No model is named, no version is given, no provider is identified, no hosting arrangement for inference is stated, and nothing commits to notifying a customer when any of it changes. There is no subprocessor list on any surface, and neither published document names a processor. One architectural fact is disclosed and is worth crediting as far as it goes: the vendor states that no additional indexing or model training is required to run the same instructions against a further document set, which tells a reader the product does not fine-tune on the customer's collection and works by instruction rather than by training. That is a statement about method, not about supply. The gap is material on this product because documents leave the customer's Relativity workspace and are submitted to the vendor for classification, so a firm is asked to send a litigation collection to an unnamed model. Verified 12 September 2026.
The architecture and the terms binding it are disclosed clearly while the components are not named. Published: enterprise grade large language models from what the vendor calls highly reputable AI service providers, operating under a zero data retention policy with data deleted on completion and never used to train their models, plus a vector database storing embeddings to power retrieval, plus a statement that all subprocessors including each language model are vetted with the appropriate internal teams and that technology is chosen on the quality of the foundation models and on providers' willingness to support the vendor's data commitments. A buyer therefore knows the shape of the chain and the commercial terms governing it, which is more than most disclose. Searched the product pages, the framework page, the support knowledge base and the newsroom on 29 Aug 2026 and located no named model provider, no named vector database provider, no published subprocessor list, and no commitment to notify customers when the supply chain changes.
Commercial Transparency
Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.
No pricing information is published at any level, including the unit of charge, which is the D band. There is no pricing page in the navigation and none was located anywhere on the estate. No plan or tier is named, no rate or band appears, no minimum or term is stated, and nothing indicates whether the product is charged per document, per gigabyte, per matter, per classification run or by subscription, which is the first question a litigation support manager would ask of a review tool. The only commercial route published is a demonstration booking, repeated as the call to action on every page, with a telephone number and a support address as the sole contacts. Under R10's closing discipline a page that only invites a sales conversation is an absence and belongs in this note alone, so no VendorPricing row is written for this record. Recorded because it sharpens the point rather than softens it: the vendor's central commercial claim is that review is completed at a fraction of the time and cost of conventional manual review, and no figure anywhere allows a buyer to test that against what a contract-attorney review would actually cost. Verified 12 September 2026.
Checked the product pages, the vendor blog, the newsroom and the support knowledge base entry point on 29 Aug 2026. No pricing page was located, no rate is published, no unit of charge is stated and no tier structure appears on the surfaces reached. Access to the generative features specifically is directed through a customer success manager or a request for a call, which is a sales gated route. No free trial or self serve entry point was located, and no third party pricing figure was located either.
Firm and Practice Coverage
Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.
Coverage is described by capability rather than by who the product serves, and the breadth is claimed without being evidenced, which is C. What is published is a product taxonomy: four suites covering early case intelligence, review, privacy and multimedia, each with named components. The buyer is identified only in general terms, as legal professionals and legal teams, with a Partners page indicating that eDiscovery service providers are a channel, and the Relativity plug-in delivery implying a firm or provider already running that platform. Beyond that, nothing. No practice area is named, no matter type is identified as suited or unsuited, no firm segment or size is addressed, and no distinction is drawn between what a law firm, a corporate legal department and a service provider would each use. There is no jurisdictional statement of any kind, which matters more than usual on a product making privilege determinations, since privilege doctrine and the treatment of in-house counsel communications differ materially between United States federal practice, state practice and other jurisdictions, and nothing published says which the privilege classifier is built for. Verified 12 September 2026.
Segment coverage is described with substance and the public sector position is unusually well defined. Segments named: law firms, corporate legal departments and government agencies at federal, state and local level, with United States attorneys' offices named specifically. Use cases are enumerated concretely for the public sector, covering litigation, investigations and Freedom of Information Act requests, and for the private sector covering document review, evidence based writing, coding and deposition analysis. Practice scope is clear and consistently stated as ediscovery, investigation and litigation end to end from data upload through production to trial preparation, with no claim to transactional or advisory capability. Geographic coverage extends to the United Kingdom and Europe with in region processing. Not located as of 29 Aug 2026: firm size segmentation, industry breakouts, and any statement of what the platform is not built for.
The 12 legal signals, side by side
Recorded rather than graded. These are the questions a practitioner has to answer before a tool touches a client matter, and the answers are taken from public material only.
Client Data in Training
Can material a lawyer puts into this product be used to train a model?
A published policy exists, it grants an improvement right, and it never names training, which is this value's first shape. Both published documents were read in full on the date shown. The Privacy Notice, last updated 8 January 2025, lists among its purposes the use of personal information for analytical purposes and to research, develop and improve programs, products, services and content. That is an improvement right that says nothing about model training either way.
Two scope points belong in the reading and neither rescues the position. The notice scopes itself to information collected on or through the vendor's websites or through other interactions with the vendor and its clients, so it does not clearly reach the document collections submitted for classification at all. And it names as a source of personal information the vendor's clients with whom it provides services, so client-supplied material is contemplated somewhere in its scope without being addressed on this question.
The Terms and Conditions, last updated 23 May 2024, are a website terms-of-use template with no data provisions. **No customer agreement is published anywhere on the estate.** One adjacent architectural statement is recorded and not credited: the vendor states that no additional model training is required to run the same instructions against a further document set, which describes method rather than a commitment.
The question is answered at both layers, which few records here manage. At the vendor layer the support documentation states that the data a customer submits and the responses they receive are not used to train models across customers and are not shared between customers. At the provider layer the published AI governance framework states that the enterprise language models used adhere to a zero data retention policy, under which data sent for processing is used solely to generate a response, is deleted on completion, and is never used to train their models.
The vendor sets out the consequence rather than leaving it implied: because customer data does not train or fine tune the models, the models hold no inherent proprietary knowledge of a case that could be exfiltrated through prompt or data injection. Recorded at policy never on the strength of both statements together.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
No located public material states how long submitted documents, prompt instructions or returned classifications are retained. The Privacy Notice does carry a retention section, and it does not reach this question: it is scoped to personal information collected through the website, sets no period, and gives only criteria, being the length of the ongoing relationship, any legal obligation, and whether retention is advisable in light of the vendor's legal position.
Nothing addresses the document collection itself. Nothing states whether a customer's classified set is deleted at the end of a matter, whether the instructions written for a review are retained, whether returned classifications and privilege log drafts persist on the vendor's side after being written back to Relativity, or whether a customer can require deletion. The Terms and Conditions contain no data provisions at all.
The question has particular force on this product because documents leave the customer's own review platform and are submitted to the vendor for processing, so the retention answer determines whether a second copy of a litigation collection exists and for how long. Both published documents were read in full and no customer agreement is published.
Retention is disclosed honestly at both layers with a period stated at neither. At the model layer the answer is zero: data sent to the language models is deleted on completion and is not retained once the request finishes. At the platform layer the vendor volunteers what most would omit, that a vector database stores numerical embeddings created from customer documents and that this storage is necessary for retrieval to work.
Disclosing that derived representations of client documents persist, and explaining why, is a materially more candid answer than the zero retention headline alone would give. Searched the framework page, the support knowledge base, the product pages and the newsroom on 29 Aug 2026 and located no retention period for the vector database, no customer control over it, and no deletion commitment for embeddings when a case closes. Recorded at disclosed without a period on that basis.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
No located public material addresses walls or matter-level segregation. Nothing published describes a permission model, user roles, access groups, conflict screening, or separation between one customer's document set and another's. The Privacy Notice, read in full, says only that the vendor takes a number of steps to protect against loss and misuse of information under its control, and the Terms and Conditions contain no access provisions.
One point of architecture is recorded because it bears on where the question sits rather than answering it: the product runs as a plug-in to Relativity, so the review platform's own permissions and workspace boundaries govern who can see a document inside the customer's environment, and what is unaddressed is what happens on the vendor's side once a set has been submitted for classification. Nothing published states whether collections from different customers or different matters are isolated during processing, and no subprocessor or infrastructure statement exists that would let a reader infer it.
Established as an absence on the surfaces read: the four solution pages, the review workflow and both published legal documents.
Segregation between customers is asserted and segregation within a customer is not addressed. The vendor states that data submitted and responses received are not shared between customers, and the architecture supports that claim, since the models hold no persistent knowledge across requests. What was not located, after searching the framework page, the support knowledge base and the product pages on 29 Aug 2026, is any published detail on how separation is enforced between users, teams or matters inside a single customer, which is where walls actually operate.
That question has real weight for this product type: an ediscovery platform holds the whole document universe for a matter, and a firm running two adverse matters needs to know retrieval cannot cross between them. No document management integration was located whose permissions could be inherited. Recorded at claimed but not documented.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
Disclosure to authorities is addressed and customer notice is absent, which is this value. The Privacy Notice states twice, once in its general sharing section and again in the California section, that the vendor may share personal information to comply with a court order or other legal process or requirements, or to protect the safety of users or others, or to protect its own rights or the rights of others. It goes further than most in the same breath, adding that it may share or use personal information for any other legally permitted business purpose upon its sole discretion.
Nothing anywhere commits the vendor to notify a customer of a demand, reserves discretion over whether to notify, promises to seek a protective order, or sets any timeframe. No transparency report exists. The scope limit is recorded and does not change the value: the notice governs personal information collected through the website rather than a submitted document collection, so for the litigation material the product actually processes the position is not addressed at all, no customer agreement being published. On a vendor holding a second copy of an opponent's production, that is a live question.
Searched the framework page, the support knowledge base, the product pages, the vendor blog and the newsroom on 29 Aug 2026, and no published customer agreement or data processing agreement was reached. No clause committing to notify a customer of a government or law enforcement request for their data was located, and no transparency report was located. Worth recording as context rather than as an answer: this vendor serves United States government agencies including attorneys' offices under FedRAMP and GovRAMP authorization, which makes the question of what happens when a government requests data from the platform a live one for its private sector customers, and nothing located addresses it.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
The working corpus is identified unambiguously and nothing behind it is, which is this value. What the product reads is the customer's own document set and nothing else: the reviewer selects the documents, writes the relevance, issue, privilege and category specifications, and submits that set through the Relativity plug-in for classification. There is no external legal corpus, no caselaw database and no reference collection, and the vendor makes no coverage claim that this signal would otherwise test.
The rights position on the customer side is not stated anywhere, no published document addressing what the vendor may do with a submitted collection, and no customer agreement exists on the estate. Behind the classification sit unnamed models: no provider, no version, and nothing about what they were trained on, which matters here because a privilege classifier's judgment about attorney-client communications is a function of what it learned privilege looks like.
One architectural statement is credited as far as it goes: the vendor states that no model training is required to run the same instructions against additional documents, so the customer's collection is not itself used to build a model for that review.
No primary law corpus is identified because the product does not hold one. Retrieval runs entirely against the customer's own case documents, uploaded into the platform and indexed as embeddings in a vector database, so the corpus is the evidence in the matter and its provenance is the discovery process itself. Searched the product pages, the framework page and the support knowledge base on 29 Aug 2026 and located no vendor supplied legal corpus, no license basis and no update cadence, and none would be expected.
Same architectural shape as the contract platforms on this index, where the absence describes the product design rather than a disclosure failure.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
No located public material addresses whether authority is checked for subsequent history, and on this product class the question does not arise in its usual form. The product cites no cases, statutes or regulations to a reader. It classifies documents from a litigation collection against instructions the reviewer writes, explains why it classified each one, identifies attorneys, sets out privilege elements and drafts privilege log entries.
Nothing it produces is a proposition about the state of the law that a lawyer would check for later treatment. The nearest adjacency is the privilege determination, which applies a legal doctrine rather than citing an authority, and nothing published states which jurisdiction's privilege law the classifier is built around or how doctrinal change would be reflected. That is recorded here rather than graded, and the jurisdictional half of it is graded on the Firm and Practice Coverage row.
Recorded so the row states the position rather than leaving a reader to infer it from the product category. The four solution pages, the review workflow and both published legal documents were read on the date shown.
Searched the product pages, the framework page, the support knowledge base and the vendor blog on 29 Aug 2026. No material was located addressing whether authority carries a treatment signal or whether subsequent history is checked, and no commercial citator license was located. Noted for context: this is an ediscovery and investigation platform whose corpus is the evidence in a matter rather than published case law, so a citator is outside its design entirely, notwithstanding that the Writing Assistant produces evidence based writing which may in practice sit alongside legal authority drawn from elsewhere.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
No located public material addresses what the product does when it cannot classify a document with confidence. What is published is explanation rather than abstention, and the distinction matters because the explanation is genuinely good. Every classification carries a restatement of the instruction the model was given and an account of why the document was categorized as it was, and privilege output sets out which elements of the privilege are present so a reviewer can see the strength of a call.
Those let a reviewer disagree with an answer; they do not describe the system declining to give one. Nothing published states that a document can be returned unclassified, flagged as borderline, routed for human decision, or accompanied by a confidence score. The privilege elements field is the closest thing to a graded output on the estate, and it grades the legal strength of the claim rather than the model's certainty about its own reading, so it is recorded here rather than treated as a confidence signal.
The gap is real on a workflow that runs across hundreds of thousands of documents in hours, where the ambiguous document is the one a reviewer most needs surfaced.
Searched the framework page, the support knowledge base, the product pages and the vendor blog on 29 Aug 2026. No published material describes what the product does when the documents do not support an answer, and no explicit no answer path or confidence signal exposed to the user was located. Two adjacent statements were considered and not treated as satisfying this signal. The vendor says it has taken steps to reduce hallucinations, which is an assertion about frequency rather than a described behavior.
And answers cite source documents so a user can verify them, which supports checking an answer that was given rather than telling a user when the corpus contained nothing. For a product whose central claim is that Deep Dive answers questions across millions of documents, what it says when the evidence is absent is a live question and is unaddressed.
Fabricated Citation Record
Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?
Searched on 12 September 2026, on the product name with a vendor qualifier and on the corporate name, against published trackers of decisions on AI-generated fabricated citations including coverage of the Damien Charlotin AI Hallucination Cases database and two independent sanctions trackers, for any court order, opinion or disciplinary record addressing fabricated or hallucinated legal citations produced by this product.
None located. This is a statement about the public record on that one subject as of the date shown, and under R119 this signal records fabricated citations and nothing else, so it is not a litigation history and no other proceeding involving the vendor would appear here. Two notes for a future reader. The vendor's name is the name of the practice area, so a bare search returns the general eDiscovery sanctions literature and every query needs a vendor qualifier.
And this product does not generate citations to authority at all, so the exposure this signal tracks is not the shape this product presents; its analogous failure mode is a misclassified privileged document, which no tracker records.
No court order, opinion or disciplinary record naming this product has been located as of 29 Aug 2026. Instrument searched: the AI Hallucination Cases database maintained by Damien Charlotin, which tracks court decisions worldwide involving AI generated hallucinated content and records the AI tool implicated where it is known. Also checked published 2026 sanctions summaries and secondary sanctions trackers. The entries located name filers, and in some rows other products, rather than this one.
This is a statement about the public record on the date shown and not a clearance. Note the exposure differs from a research tool: this product generates from case evidence rather than from case law, so its characteristic failure would be a mischaracterised document or a fact unsupported by the record rather than an invented citation, and that failure mode is far less likely to be cataloged in a hallucination database.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
No located public material engages with bar or ethics guidance. No bar opinion is named on any surface read, ABA Formal Opinion 512 does not appear, no state guidance on generative AI in legal practice is referenced, and nothing maps the product to any rule of professional conduct. Nor is professional responsibility referred to in general terms, which is what separates this from the tier above: neither published document mentions professional obligations, competence or supervision, and the product pages address efficiency, accuracy and cost rather than duty.
The absence is worth stating plainly because the product operates in the one part of legal AI where a professional-conduct framework already exists and is well developed: privilege review sits directly on the duty of confidentiality, and technology-assisted review has a decade of judicial and bar commentary on defensibility behind it. Named here as a surface not opened under the sufficiency discipline, and as the likeliest place the position would be found if it exists: the vendor publishes a substantial practitioner library of fact sheets, whitepapers, case studies and books on AI document review.
Searched the product pages, the AI governance framework page, the support knowledge base, the vendor blog and the newsroom on 29 Aug 2026. No engagement with any named ethics opinion or bar guidance was located, including ABA Formal Opinion 512 and state bar guidance. The vendor publishes a substantial AI governance framework and a set of generative AI principles, which govern its own conduct rather than engaging with the professional responsibility rules binding the lawyers who use it.
Also not located: engagement with the Federal Rules of Civil Procedure or with case law on technology assisted review, which would be the natural professional touchstone for an ediscovery product using AI for first pass review.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
Cost savings are the central claim and nothing addresses billing or disclosure, on work that sits squarely inside a fee relationship. The claims are repeated across the estate: document reviews completed in a fraction of the time and cost of conventional manual review, hundreds of thousands of documents reviewed in hours, and the important documents identified immediately at a fraction of the cost. Document review is the paradigm case this signal was written for.
It is among the largest line items in litigation, it is conventionally billed to the client either as attorney time or as a vendor cost passed through, and replacing it with machine classification changes that number by an order of magnitude. Nothing published addresses what happens to the bill, whether AI-assisted review is identified on an invoice or in a budget, or whether the client is told. No pricing is published at all, so a firm cannot even see what the substitution costs.
One route the vendor could take and does not is recorded: the workflow already generates per-project performance metrics for validation, which is the nearest thing on this estate to a per-matter record of AI-assisted work, and it is presented as a defensibility artifact rather than a billing one.
Savings are claimed and nothing is published on the client's side of the equation. The vendor's framing is throughput rather than hours: automating first pass review with recall and precision that rivals eyes on review, streamlining processes, eliminating manual tasks, and answering questions across millions of documents in seconds. Automating first pass review is the single largest displaceable cost in litigation support and the claim is squarely about replacing billable human review time.
Searched the product pages, the framework page, the support knowledge base and the vendor blog on 29 Aug 2026 and located no per matter record of AI assisted work intended for fee purposes, and no guidance on billing, fee or client disclosure treatment. Recorded at savings claims only.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
No located public material supports a client-side disclosure obligation, and none of the three artifacts this signal looks for exists. There is no subprocessor list on any surface. No model provider is named anywhere, so a firm asked by a client which third party processed its documents could not answer from anything published. There is no data processing addendum, no security exhibit, no certification and no client-facing disclosure pack that could be forwarded.
The value is not on-request, because nothing indicates such material exists behind a sales conversation: there is no security page, no trust center, no compliance contact and no request mechanism published, and the site navigation renders in full so that is the inventory. The two documents that do exist were read in full and neither serves: the Privacy Notice is a website notice that scopes itself away from client document collections, and the Terms and Conditions are a website terms-of-use template with no data provisions.
This matters more than the bare grade suggests on a product delivered through Relativity, because the firm's client is often a corporate legal department with its own outside counsel guidelines governing exactly this question.
Substantial diligence material is published openly and the specific artifact this signal names is not. Available without a sales conversation: a named AI governance framework describing the architecture and the terms binding the model providers, a support knowledge base answering the training question directly, and a certification set including FedRAMP Moderate and GovRAMP Moderate authorization whose status is a matter of public record, alongside SOC 2 Type 2, ISO 27001, Cyber Essentials Plus and G-Cloud.
A firm could evidence a great deal from that. What was not located as of 29 Aug 2026 is a subprocessor list: the vendor states it thoroughly vets all subprocessors including each language model, which is a statement about process rather than a disclosure of who they are, and no model provider is named anywhere located. No client facing consent or notification pack was located either. Recorded at on request on the strength of the published material falling short of the named artifacts.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Several elements of a record exist and no document-level export is described, which is this value, and the elements here are the strongest in this lane because defensibility is the discipline this product class already lives under. Per document, the output carries the classification, a restatement of the instruction the model was given, and an explanation of why the document was categorized as it was, produced consistently across every document and every classification.
Privilege adds more: the attorneys identified and named in their own field, the elements of the privilege present, and a document summary with the privilege reasoning drafted as a log entry. At the project level the published workflow ends in validation and the generation of industry standard performance metrics, which in this field means recall and precision against a sample, and which is precisely what a party is asked to produce when an AI-assisted review is challenged.
What is missing is the model. Nothing identifies which model produced a classification, no version or provider is named, and nothing records when the model or the instructions changed mid-review, so the one element a court would need to assess reproducibility is absent. No disclosure template, certification form or ESI-protocol guidance was located.
Several elements of a record are available, assembled from the product's design rather than offered as a disclosure artifact. Outputs cite the source documents they rest on, with access to the underlying materials, so what was relied on is traceable per answer, and an ediscovery platform necessarily maintains production and review histories at document level. Two elements are missing: no single per document export covering model used, sources retrieved and human verification together was located, and no model is named in published material so the model used could not be stated.
Recorded at partial record. Noted for a reader: courts have engaged with technology assisted review in ediscovery for well over a decade, so this is one of the few product categories on the index where a defensibility record has an established judicial context, and the vendor does not connect its AI disclosure material to it.
The questions both sides leave open
Derived from the records above rather than written, so it cannot favor either vendor. Take these into both conversations and ask each side the same question.
- AI Liability and Recourse
- Commercial Transparency
- Good Law Verification
- Refusal and Uncertainty Behavior
- Bar Guidance Alignment
Which one fits
Choose eDiscovery AI if
- Your review already runs in Relativity. eDiscovery AI is listed on the Relativity App Hub and works as a plug in: a reviewer selects a set, writes relevance, issue, privilege and category instructions, submits the documents and gets classifications back in a mapped field, with no indexing or model training needed to rerun the same instructions.
- You want privilege calls a reviewer can audit, and a log drafted as you go. eDiscovery AI's privilege output names the attorneys it identified in their own field, sets out which elements of the privilege are present, and drafts a summary and privilege reasoning as a log entry that can be used or edited.
- You want to measure the review yourself before relying on it. eDiscovery AI's published workflow ends with validating the classifications and generating standard performance metrics, and each relevance call restates the instruction it was given and explains why the document was categorized as it was. It states 90 percent recall without publishing the test set behind it.
Choose Everlaw if
- You need to know what happens to your documents inside the AI. Everlaw states that submitted data and responses do not train models and are not shared between customers, that its enterprise language models run under zero data retention and delete data on completion, and that embeddings of your documents persist in a vector database to support search.
- Your client is a government agency or holds you to government grade security. Everlaw holds FedRAMP Moderate and GovRAMP Moderate authorization, with its generative AI features included in the 2025 annual assessment by a third party assessment organization, alongside SOC 2 Type 2, ISO 27001:2013 and Cyber Essentials Plus.
- You want one platform from upload to trial, with AI processing kept in your region. Everlaw covers upload, processing, search, review, production and trial preparation, and offers in region AI processing for United Kingdom and European customers under the same zero retention and no training terms.
In summary
eDiscovery AI
eDiscovery AI, based in Bloomington, Minnesota, applies large language models to document review inside an existing litigation workflow, delivered as a plug in listed on the Relativity App Hub. Its Review suite classifies documents for relevance against instructions a reviewer writes in plain language and identifies privilege, naming attorneys, setting out the privilege elements present and drafting log entries; further suites cover early case intelligence, personal data detection and image review including facial recognition. The AI Legal Index grades it in the top two bands on four of fifteen capability axes, with an A on AI centrality. As of 12 September 2026 the index located no security attestation, no published position on training or retention for submitted documents, no customer agreement and no pricing.
Everlaw
Everlaw is a cloud native ediscovery, investigation and litigation platform from Oakland, California, covering processing, search, review, production and trial preparation for law firms, corporate legal departments and government agencies including United States attorneys' offices. Its EverlawAI Assistant adds coding suggestions for first pass review, Deep Dive answers across millions of documents with citations to source, and writing and deposition tools. The AI Legal Index grades it in the top two bands on nine of fifteen capability axes, with A grades on privilege posture, AI governance and security certifications, including FedRAMP Moderate authorization covering its generative features. It states that customer data does not train models. As of 29 August 2026 the index located no pricing, no published customer terms and no retention period for document embeddings.
Questions buyers ask
eDiscovery AI vs Everlaw: which is better for AI document review?
On published evidence Everlaw sits in the top two bands on nine of fifteen AI Legal Index capability axes and eDiscovery AI on four of fifteen, mostly because Everlaw publishes how it protects documents and eDiscovery AI does not. The products answer different needs, though. eDiscovery AI adds relevance and privilege classification to a Relativity workspace a team already runs. Everlaw replaces the whole ediscovery platform, with review AI built in. The first question is which platform the review lives on.
Does eDiscovery AI train on client documents?
It does not say. eDiscovery AI's privacy notice covers information collected through its website and reserves a right to use personal information to research, develop and improve its products, without mentioning training, and its terms are a website template with no data provisions. No customer agreement is published. The vendor does state that no model training is needed to run the same instructions against more documents, which describes its method rather than a commitment. Checked on 12 September 2026. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 25, 2026. No vendor pays for placement.
What does Everlaw integrate with?
Everlaw is built to replace a chain of tools rather than connect into one, covering upload through trial preparation itself, and the index located no integrations page or API documentation as of 29 August 2026. The index also records two dated product changes: a Model Context Protocol integration with Harvey on 27 August 2026, letting Harvey users query Everlaw evidence, and a ChatGPT Enterprise plugin released on 17 September 2026. eDiscovery AI, by contrast, runs as a Relativity plug in. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 25, 2026. No vendor pays for placement.
Is Everlaw FedRAMP authorized?
Yes. Everlaw holds FedRAMP Moderate and GovRAMP Moderate authorization, and its generative AI features were included in the 2025 FedRAMP annual assessment, where a third party assessment organization tested them and issued an attestation letter. It also lists SOC 2 Type 2, ISO 27001:2013, Cyber Essentials Plus, HIPAA and UK G-Cloud. eDiscovery AI claims no certification of any kind, and no security page or trust center was located on its site. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 25, 2026. No vendor pays for placement.
What do eDiscovery AI and Everlaw both leave unpublished?
Who pays when the AI is wrong. Neither publishes a customer agreement, a liability cap, an indemnity or a warranty covering a missed responsive document or a privileged document produced by mistake. Neither names the language model behind its review, publishes a price or unit of charge, or says what its AI does when a document is ambiguous. Neither engages with bar guidance on AI, and both sell review savings without saying how they should appear on a client's bill. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 25, 2026. No vendor pays for placement.
Three readings to weigh. eDiscovery AI's low grades on confidentiality, stewardship and security record that its published privacy notice and terms cover its website rather than the document collections it classifies, and that no security page, attestation or customer agreement was located; a customer agreement may say more, and it is not public. Its 90 percent recall figure has no test set or method behind it. Everlaw's confidentiality grade rests on published policy and a FedRAMP authorization rather than a published customer agreement, and no retention period is stated for the document embeddings it keeps. eDiscovery AI was verified on 12 September 2026 and Everlaw on 29 August 2026. Neither vendor reviewed this page.
Neither vendor paid for inclusion, placement or a grade, and neither reviewed this page before it published. Everything above comes from public material on the dates shown. How the index grades.