Epiq AI vs Lighthouse: how they compare in 2026
Epiq AI and Lighthouse are the AI lines of two large eDiscovery service providers: Epiq sells review, privilege and fact research applications on its Laer platform, and Lighthouse sells LighthouseIQ alongside Spectra, a self service platform over Relativity. On the totals they are one axis apart. Epiq AI sits in the top two bands on fifteen of fifteen axes and Lighthouse on fourteen of fifteen, with identical grades on ten. The count hides where the strength sits. Lighthouse holds three A grades to Epiq's one, on privilege, data stewardship and deployment, and the sharpest difference is training. Lighthouse's Spectra agreement bars it from using client data, or the inputs and outputs of its AI, to train any AI service. Epiq's Software Terms tell the client that Epiq may use client data to fine tune its generative AI tools for that client's own environment, with no opt out stated. Lighthouse also publishes a subprocessor list with a location for each and a 48 hour breach notice. Epiq's extra axis is commercial: it prices per matter with unlimited reclassification at no added fee.
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 product is the AI, which is the A band for a product-level record. Epiq AI is a line of applications whose work is done by the models: Epiq AI for Review turns review protocols into prompts and classifies documents for responsiveness, issues, PII and privilege; Epiq AI for Privilege classifies privilege and drafts logs; Epiq Assist answers natural-language questions over a dataset; all run on the Epiq Laer platform, which orchestrates models, agents and reviewers. Without the models there is no product. Epiq's wider eDiscovery platform and services, and resold third-party AI such as Leah, are outside this record. Verified 18 September 2026.
The models drive the capabilities the vendor now leads with, on an eDiscovery platform and service that runs without them, which is the B band. LighthouseIQ's four applications answer natural-language questions from a matter's documents, draft chronologies and case strategy, code documents for responsiveness and issues, and identify privilege and draft privilege logs, using predictive and generative AI; Spectra runs integrated machine-learning and AI services inside its self-service workflows. Underneath are processing, hosting, Relativity review and production, and a large forensics and managed review business, all of which operate without the models. Verified 18 September 2026.
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 and validation are real and the method is documented, short of published accuracy figures with a test set, which is the B band. Epiq AI for Privilege gives reasoning and line-level citations for each call and Epiq Assist answers with links to supporting documents. Epiq's validation article sets out how results are tested: recall, precision, elusion and richness estimated from random samples with confidence intervals, and per-tag recall and precision trade-offs from k-fold cross-validation, which it says are not a substitute for a control set. The only outcome figures (review up to 90 per cent faster, more than 80 per cent of review automated) are speed claims with no accuracy measurement behind them. Verified 18 September 2026.
Grounding is real and documented with a described retrieval method, short of accuracy figures an outsider can test, which is the B band. The IQ Answers launch blog says answers draw only on the documents ingested for the matter, handle one question at a time to avoid context drift, and follow a three-step process of interpretation, layered retrieval (predictive relevance scoring, semantic similarity and context-preserving retrieval) and synthesis, with each response carrying highlighted source text and exemplar documents so every statement is traceable. The GenAI defensibility Q&A describes sample-based validation with attorney review and prompt versioning. The only accuracy figure, that IQ Priv's privilege descriptions were 12 per cent more accurate than contract attorneys in blind comparisons, gives no test set, and the IQ Answers page describes 'no hallucinations', which is a claim rather than a measurement. Verified 18 September 2026.
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.
Human validation is built in and described, with per-tag cut-off guidance, short of published stopping points or an error route, which is the B band. Classifications carry reasoning and citations for reviewers to check, each issue tag comes with guidance for choosing a cut-off score, Epiq recommends sample validation of the responsive and non-responsive sets before production, and review services pair the AI with reviewers who validate its results. The Software Terms state that generative outputs are preliminary and that the client is responsible for verifying them. Where the AI routes low-confidence material, and what happens when validation fails, are not published. Verified 18 September 2026.
A written commitment to attorney validation with real review surfaces, short of thresholds or a route when the AI is wrong, which is the B band. The GenAI defensibility Q&A says prompt engineers set AI instructions with counsel, attorneys review results on sample populations to confirm accuracy, and iterative feedback refines the prompts; IQ Review shifts work from document-by-document review to targeted validation and quality control, with summaries and rationales for each call, and IQ Priv produces editable log lines for attorney review. The vendor states that AI does not replace human reviewers. No confidence threshold, stopping point or published escalation path is stated, and IQ Review is run by Lighthouse consultants rather than configured by the customer. Verified 18 September 2026.
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.
Real deployment evidence with figures, without named customers and method, which is the B band. Epiq says 130 clients, including global corporations and law firms, adopted Epiq AI within its first year, and its case studies report a health insurer's 50 per cent savings, a media company saving US$1.5 million on a regulatory deadline, and an Am Law 25 firm deploying Epiq AI for Review; all are anonymised in their titles. Epiq AI won the 2026 Legalweek Leaders in Tech Law award for best use of AI in eDiscovery and litigation. Verified 18 September 2026.
Real deployment evidence with figures or names, not both with a method, which is the B band. The Spectra page quotes Penningtons Manches Cooper by name without figures, reports an unnamed large healthcare organisation saving $500,000 in its first year with 95 per cent fewer documents sent to review, and states that Spectra customers saved $2,951,280.50 in July 2026, calculated as native data kept out of Relativity hosting multiplied by the standard per-GB hosting charge. LighthouseIQ is described as applied across 1.4 billion documents with outputs accepted by the DOJ and FTC, without naming the matters. Verified 18 September 2026.
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.
Confidentiality is written into published agreements, with training limited but not excluded and privilege not named, which is the B band. The Epiq Software Terms (revised February 2026) treat every Discovery Document, including material under protective order, as confidential without the usual exceptions, limit use of client data to providing the software, require notice before compelled disclosure, and require return or destruction at termination at the client's choice; but section 15 lets Epiq use client data to fine-tune its generative AI tools for that client's environment. The Epiq AI FAQ says it does not train public models with client data, and the Responsible AI page says the Azure OpenAI service behind Epiq's summarisation keeps prompts and outputs from OpenAI and other customers. Neither agreement addresses privilege or work product by name. Verified 18 September 2026.
The confidentiality position is written into a published agreement, with privilege addressed directly, which is the A band. The Master Spectra Subscription Agreement (2 April 2026) bars Lighthouse from using client data, metadata, or the inputs or outputs of its AI services to train or develop any AI service, its own or a third party's, bars sending client data to any publicly available or shared AI model without written authorisation, gives privilege its own section confirming that privileged communications and work product remain protected and that work is done at counsel's direction, commits to notice before compelled disclosure, and requires deletion of client data within 30 days of termination. Spectra documents matter creation and role-based access so users see only the matters they need. The contract governs Spectra; for LighthouseIQ the same no-training and no-retention position appears in the launch blog rather than a published agreement. Flagged for sampling. Verified 18 September 2026.
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.
A clear contractual position on advice versus tooling, short of jurisdiction limits, which is the B band. The Epiq Software Terms say that nothing Epiq does is the practice of law, a legal opinion or legal advice, and that Epiq has told the client it is not representing it; the Epiq Global Terms of Service say Epiq is not a law firm. The Software Terms add that generative outputs are preliminary, untested and unverified, and that the client bears responsibility for verifying them. The product is sold to law firms and legal departments. No jurisdictional limits are stated. Verified 18 September 2026.
A plain contractual position on advice versus tooling, short of jurisdiction limits, which is the B band. The Spectra agreement states that Lighthouse is not engaged in the practice of law, provides no legal advice and creates no attorney-client relationship, that its services represent its interpretation as a legal technology provider, and that work on privileged material is done at the direction of the client or its counsel; it also describes optional matter conflict checks. The GenAI defensibility Q&A says attorneys govern and validate AI outputs and confirm they reflect their professional judgement. The product is sold to law firms and legal departments, and no jurisdictional limits are stated. Verified 18 September 2026.
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.
A published governance structure with a named body, short of testing results, which is the B band. Epiq's Responsible AI page says a Responsible AI Committee oversees the development and deployment of its AI, that bias assessments are carried out to identify and reduce bias, and that practices are reviewed regularly, with employee training and a route for staff to raise concerns; Epiq AI Labs, led by the founders of Laer AI with Cornell's Center for Data Science for Enterprise and Society as a founding member, researches the technology. No assessment results, release testing or findings about uneven performance are published. Verified 18 September 2026.
A published governance method with substance, short of testing results or an accountable owner for the product, which is the B band. The GenAI defensibility Q&A, given by the Senior Director of AI and Analytics, sets out how generative workflows are governed: prompt engineers work with counsel, attorneys validate outputs on sample sets with qualitative and quantitative metrics, and prompts are documented and versioned; it also says predictive workflows follow validated technology-assisted review practice. The vendor says LighthouseIQ was pressure-tested on live matters over three years. No release testing results, named accountability for the AI applications, or findings on uneven performance across languages or data types are published. Verified 18 September 2026.
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.
Real published commitments on data handling with gaps, which is the B band. The Software Terms commit to breach notification without undue delay with assistance to the client, confidentiality obligations on staff and agents, no subcontracting of personal data processing without permission, and return or destruction of personal data at termination; the Global Terms of Service allow 30 days to export data after termination, after which Epiq may delete it. The security page describes identity and access management on Zero Trust principles, 24/7 managed network monitoring and regular penetration tests. No subprocessor list, fixed notification window or retention period for AI inputs and outputs is published. Verified 18 September 2026.
Retention, deletion, access, subprocessors and incident practice are all published and specific enough to hold the vendor to, which is the A band. The Spectra agreement's Information Security Addendum commits to notify a security incident within 48 hours with a description, likely consequences, mitigation and a contact, to multi-factor access controls, central tamper-protected logging and annual third-party penetration tests, and the Data Processing Addendum commits to destroy or return personal data within 60 days of termination, with client data destroyed within 30 days under the main agreement. The Processor Compliance Statement adds encryption at rest and in transit and a 30-day rolling backup retention, and the published subprocessor list names every subprocessor, affiliate and review-staffing firm with its role and processing location, with a dated change log and notice and objection rights for new ones. The launch blog says the AI workflows retain no customer data. Verified 18 September 2026.
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.
A real published position with indemnities and caps, and an express disclaimer of AI output, which is the B band. The Software Terms warrant that the software performs substantially as documented, with correction or termination and a pro-rata refund as the remedy; indemnify the client for infringement, material breach, gross negligence, security breaches the client did not contribute to and legal non-compliance; and cap liability at six months of fees for the software concerned. The Global Terms of Service cap liability at 12 months of fees, with carve-outs for confidentiality breach and gross negligence. Both disclaim any warranty of accuracy for AI outputs, which the client must verify. Verified 18 September 2026.
A real published position on liability with an indemnity and a cap, short of anything specific to AI output, which is the B band. The Spectra agreement warrants that the services conform to their documentation, with refund, repair or replacement as the remedy; indemnifies the client against third-party claims that the Lighthouse IP infringes; caps each party's liability at fees paid in the prior 12 months, with carve-outs for fraud, wilful misconduct, gross negligence, death or personal injury and third-party claims; and gives a termination right with a refund if availability or response levels are missed for three consecutive months. Nothing addresses an incorrect AI answer, privilege call or log line. Verified 18 September 2026.
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.
Real integrations with the review platforms legal teams use, described at the level of what they do, which is the B band. Epiq AI Accelerators add translation, image analysis, OCR and agentic automation inside RelativityOne and Relativity Server; Epiq AI for Review and Epiq Assist are built into Epiq Discover; and access runs through Epiq Access single sign-on in the Epiq Service Cloud, with Microsoft, AWS and Relativity named as technology partners. Epiq lists an Epiq MCP page in its navigation, which the fetch tool would not retrieve and search did not surface, so it was not read. No implementer documentation is published. Verified 18 September 2026.
Real integrations into the review platforms legal teams use, with the flow described, short of implementer documentation, which is the B band. Spectra runs matter setup, processing and culling, then promotes data into Relativity Server or RelativityOne for review and carries it through production and export, with Workflow Manager and Native File Manager controlling what reaches active hosting; IQ Review delivers its results directly in Relativity. The estate also covers Microsoft 365 and Purview and Google Vault linked-file workflows. Technical documentation is provided to customers through Spectra rather than published. Verified 18 September 2026.
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.
Deployment is described in real terms without the regions, which is the B band. Epiq says it runs Tier 3+ data centres in ten locations worldwide, offers private cloud capabilities in the Epiq Service Cloud and region-specific sites across 14 countries, and that the Epiq Laer platform uses local models to control throughput and cost. The locations of the data centres, where AI processing happens relative to storage, and residency options for Epiq AI are not published. Verified 18 September 2026.
Deployment options, the regions available and where processing happens as distinct from storage are all published, which meets the A band. The subprocessor list names Microsoft Azure in the US and UK and co-located data centres in Seattle, Phoenix, London, Frankfurt and Dornach, says RelativityOne runs in the customer-specified region, and gives a processing location for every subprocessor and affiliate, including off-hours support from India and review staffing in named countries on a case-by-case basis; the security page describes region-controlled cloud workflows that keep storage and computing within defined boundaries, and the LighthouseIQ workflows run in a private Azure environment. What changes between Spectra and managed-service deployments is not set out. Flagged for sampling. Verified 18 September 2026.
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.
Named certifications without scope, dates or reports, which is the B band. Epiq's security page shows ISO, Cyber Essentials, TISAX and UKAS marks and says it is aligned with HIPAA, NIST and FISMA and that third parties assess its programme, with Bitsight ratings giving clients visibility of its posture. The page does not say which ISO standard is held or its scope, still shows the EU-US Privacy Shield mark invalidated in 2020, and offers no reports or trust centre. Verified 18 September 2026.
Real certifications with a named certifier, short of evidence a reader can open, which is the B band. The Processor Compliance Statement says the ISO 27001 information security management system is certified by A-LIGN Assurance, US operations hold SOC 2 and UK operations Cyber Essentials, and the Spectra agreement commits to an annual assessment against ISO 27001, the HIPAA Security Rule and SOC 2 Type 2 with a summary provided on written request. The security page links the ISO certificate and a security overview in a Trust Center at resource.lighthouseglobal.com, which renders only in the browser and could not be read, so scope and dates are not confirmed. Verified 18 September 2026.
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 architecture is described and one provider is named, short of a model list or change notice, which is the B band. The Responsible AI page says Epiq's summarisation uses the Microsoft Azure OpenAI Service and sets out that service's data terms; the Epiq AI FAQ says the Epiq Laer platform orchestrates multiple models and agents and uses local models strategically. Epiq has announced partnerships with AWS and DeepJudge. Which models power each Epiq AI application, their versions and how model changes are notified are not published. Verified 18 September 2026.
The architecture and hosting are described without the model providers, which is the B band. The vendor describes IQ Fabric (processing, cognition and orchestration), combines predictive models trained on attorney coding with generative models, says IQ Answers uses legal-tuned models and that IQ Review and IQ Priv run in a private Azure-based environment, and its subprocessor list names Microsoft Azure for hosting and Daypack Research for AI workflow management tools, with notice before new subprocessors. No model, model provider or version is named. Verified 18 September 2026.
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.
The pricing unit and structure are published without a figure, which is the B band under R10. Epiq AI is sold per matter, with unlimited prompts and protocol changes and reclassification at no additional fee, and is also offered to managed-services clients at a predictable monthly cost; the Software Terms describe prepaid and monthly service fees with 12-month auto-renewing subscriptions. No price is published. Verified 18 September 2026.
The shape of the offer is visible and the numbers are not, which is the C band. The Spectra agreement sets out a free Tier 1 support level and a paid Tier 2 expert tier, fees set in an order form, automatic annual renewal and rate increases on 90 days' notice, and the Spectra page offers a savings calculator and describes lower-cost Nearline storage; the per-GB hosting charge appears only inside the calculator's disclaimer. No rate, tier price or LighthouseIQ pricing is published. Verified 18 September 2026.
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.
Segments and matter types are described with substance, short of stated limits, which is the B band. Epiq AI is sold to corporate legal departments and law firms and positioned for investigations, risk assessments, HSR filings and second requests, merger reviews, post-execution M&A contract review, PII and PHI identification, early case assessment, incoming production review, deposition preparation and litigation review, with versions for antitrust, compliance and plaintiff class actions; Epiq says it classifies documents in any language. Where the AI is weaker, by language, data type or matter type, is not stated. Verified 18 September 2026.
Segments and matter types are described with substance, short of stated limits on the AI, which is the B band. The vendor serves corporate legal departments and law firms in financial services, technology, pharmaceutical, healthcare, telecommunications, manufacturing and construction, with specialties in antitrust and HSR second requests, multidistrict litigation, DSARs and Microsoft 365, and Spectra covers investigations, DSARs and compliance reviews, and routine litigation. IQ Answers works on any data with document text, and LighthouseIQ claims no limits on document types or volume; nothing states where the AI applications perform less well. Verified 18 September 2026.
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?
The published agreement reserves a right to train on client content, limited in scope. Section 15 of the Epiq Software Terms and Conditions lets Epiq use client data to fine-tune or adapt its generative AI tools for the limited purpose of improving outputs specific to that client's environment; no opt-out is stated. Elsewhere the same terms restrict client data to providing the software, the Epiq AI FAQ says Epiq does not train public models with client data, and the Epiq Global Terms of Service let Epiq use customer data to improve the services for that customer.
The published agreement prohibits training on customer content. The Master Spectra Subscription Agreement (2 April 2026) bars Lighthouse from using client data, metadata, or the inputs or outputs of its AI services to train or develop any AI service, its own or a third party's, directly or indirectly, except to provide the services to that client, and bars sending client data to any public or shared AI model without written authorization.
The agreement governs Spectra; the LighthouseIQ launch blog states the same no-training position for IQ Review and IQ Priv.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
Retention is addressed without a period for AI inputs and outputs. The Global Terms of Service give 30 days to export data after termination, after which Epiq may delete it, and the Software Terms require return or destruction at the client's choice on termination; the Azure OpenAI service behind summarization is described as stateless. How long prompts and outputs are kept during a matter is not stated.
Fixed periods are published. Client data is destroyed within 30 days of termination under the Spectra agreement (personal data within 60 days under its DPA), backups run on a 30-day rolling retention, and the LighthouseIQ launch blog says the AI workflows retain no customer data. Clients manage matter hosting and archiving within Spectra, but no setting for the retention of AI inputs and outputs is described.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
Access control is claimed without documented matter-level segregation. The Epiq Service Cloud page describes flexible user permissions under Epiq Access single sign-on, and Epiq AI is bought per matter, but no page describes how access to one matter's data, prompts or outputs is walled from other matters or users.
The product keeps its own permission model at matter level. Spectra lets clients create matters, provision users and assign access by role, with delegated matter administration and less unnecessary access to review environments, and the Information Security Addendum says the client assigns each authorized user an account and role that controls access. Review itself runs in Relativity, whose permissions apply there.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
The Software Terms commit to written notice before disclosing confidential information under a court or agency order, and cooperation in seeking a protective order; the Data Processing Terms commit to inform the client and give it an opportunity to object where a law requires disclosure of personal information, unless notice is prohibited. The Global Terms of Service give prior notice where permitted. No transparency report is published.
The Spectra agreement commits to prompt written notice of a court order, regulatory requirement or subpoena, where permitted by law and feasible, so the client can seek a protective order, and its DPA commits to immediate notice of any request to disclose client personal data with a copy of the request. No transparency report is published.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
Searched the Epiq AI pages, Responsible AI page, Software Terms, Global Terms of Service and security page on 18 September 2026. The products work on the client's own matter data; no external legal corpus is described.
Searched the Spectra agreement, Privacy Policy, Spectra and security pages and the LighthouseIQ launch material on 18 September 2026. IQ Answers draws only on the documents ingested for the matter; no external legal corpus is described.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
Searched the same surfaces on 18 September 2026. Citations point to documents in the matter, not to case law, and no check of cited authority is described.
Searched the same surfaces on 18 September 2026. The products cite matter documents rather than case law, and no check of cited authority for subsequent history is described.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
The product exposes its uncertainty for classification. Each issue tag comes with recall and precision trade-off metrics from k-fold cross-validation and guidance for choosing a cut-off score, and Epiq recommends sample validation of the responsive and non-responsive sets with confidence intervals. How Epiq Assist behaves when the dataset holds no answer is not described.
Searched the same surfaces on 18 September 2026. IQ Answers restricts itself to the ingested documents and one question per result, and responses carry source passages, but no behavior for when no supporting document is found, and no confidence signal shown to the user, is described.
Fabricated Citation Record
Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?
Searched the AI Hallucination Cases database maintained by Damien Charlotin and trade press reporting on 18 September 2026 for court records addressing fabricated or hallucinated content in output from Epiq AI. None located. This signal does not record litigation history of any other kind.
Searched the AI Hallucination Cases database maintained by Damien Charlotin and trade press reporting on 18 September 2026 for court records addressing fabricated or hallucinated content in output from Lighthouse products. None located. This signal does not record litigation history of any other kind.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
Searched the Epiq AI pages, Responsible AI page, agreements and validation article on 18 September 2026. Defensibility is framed around metrics accepted by regulators, courts and opposing counsel; no ethics opinion, court AI order or rule, or bar guidance is referred to.
Professional responsibility is referred to in general terms. The GenAI defensibility Q&A says attorneys govern and validate AI outputs so they meet legal standards and reflect professional judgment, and notes that responsiveness review lacks settled case law. No ethics opinion, court AI order or rule, or judicial or bar guidance is named on the surfaces read.
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 marketed without disclosure guidance. Epiq AI is sold per matter with unlimited prompts and reclassification at no extra fee, and its case studies report savings of 50 percent and US$1.5 million; nothing addresses how law firm customers should bill or disclose AI-assisted review to their clients.
The vendor's charge is set up to reach the end client as a matter expense. The Spectra agreement provides for invoicing through e-billing platforms using generic timekeeper codes such as E118, the LEDES expense code for litigation support vendors, which is how law firms pass vendor costs to clients. Nothing addresses how firms should disclose AI-assisted work or its cost to clients.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
Searched the Epiq Global Terms of Service, Software Terms, Data Processing Terms, Privacy Policy, Privacy Statement, Responsible AI page and security page on 18 September 2026. The Data Processing Terms are public, and Azure OpenAI is named for summarization, but no subprocessor list, model provider list or disclosure pack for client guidelines was located, and none is offered on request.
A current subprocessor list is published, with each subprocessor's role, processing location and parent, a dated change log (latest 14 September 2026), and a Data Processing Addendum and Information Security Addendum readable in the Spectra terms. No AI model provider is named: the list names Microsoft Azure for hosting and Daypack Research for AI workflow tools.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Parts of a defensibility record exist. Epiq AI reports recall, precision and elusion metrics with confidence intervals, gives reasoning and line-level citations for classification and privilege calls, and Epiq positions the metrics for regulators, courts and opposing counsel. No per-document record of the model used and the human verification is described as exportable.
Parts of a defensibility record exist, short of a document-level export. The vendor documents prompts and versions, validates outputs on attorney-reviewed samples with metrics, provides rationales and citations for AI calls, and says its workflows have been accepted by regulators including the DOJ and FTC. No per-document record of the model used and the human verification is described as exportable.
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.
- Primary Law Corpus Provenance
- Good Law Verification
Which one fits
Choose Epiq AI if
- You want AI review priced per matter without paying again every time the instructions change. Epiq sells Epiq AI per matter with unlimited prompts and protocol changes and reclassification at no additional fee, and offers managed services clients a predictable monthly cost. No figure is published.
- You need numbers to defend the review to opposing counsel. Epiq AI for Review reports recall, precision, elusion and richness from random samples with confidence intervals, gives each issue tag trade off metrics from cross validation with guidance on choosing a cut off score, and Epiq AI for Privilege gives reasoning and line level citations for each call.
- You want a named body overseeing the AI. Epiq's Responsible AI page describes a Responsible AI Committee overseeing development and deployment, bias assessments, staff training and a route for raising concerns, with research through Epiq AI Labs, whose founding members include Cornell's Center for Data Science for Enterprise and Society.
Choose Lighthouse if
- You need a contract term that your documents, prompts and results will not train anyone's AI. Lighthouse's Spectra agreement bars it from using client data, metadata, or the inputs or outputs of its AI services to train or develop any AI service, its own or a third party's, and bars sending client data to any public or shared model without written authorization.
- Your client audits every vendor that touches its data. Lighthouse publishes a subprocessor list naming each subprocessor, affiliate and review staffing firm with its role and processing location, with a dated change log, and its security addendum commits to notice of a security incident within 48 hours.
- Your data must stay in a defined region. Lighthouse names Microsoft Azure in the United States and United Kingdom and data centers in Seattle, Phoenix, London, Frankfurt and Dornach, says RelativityOne runs in the region you specify, and runs its LighthouseIQ workflows in a private Azure environment.
In summary
Epiq AI
Epiq AI is the AI product line of Epiq, the New York legal services and technology company, built on technology from its 2024 acquisition of Laer AI and running on the Epiq Laer platform. It covers Epiq AI for Review, which classifies documents from review protocols, Epiq AI for Privilege, which drafts privilege logs with reasoning and line level citations, and Epiq Assist for fact research, with accelerators inside RelativityOne. The AI Legal Index grades it in the top two bands on fifteen of fifteen capability axes, with an A on AI centrality. Epiq says 130 clients adopted it in its first year. Its Software Terms let Epiq fine tune its generative tools on a client's data for that client's environment. As of 18 September 2026 the index located no named model for its review applications and no published price.
Lighthouse
Lighthouse is an eDiscovery, review and information governance company founded in 1995 and based in Bellevue, Washington, with funds affiliated with Lightyear Capital holding a controlling stake since 2019. It sells LighthouseIQ, whose four AI applications answer questions from a matter's documents, build chronologies, code documents for responsiveness and identify privilege, and Spectra, a self service eDiscovery platform over Relativity. The AI Legal Index grades it in the top two bands on fourteen of fifteen capability axes, with A grades on privilege posture, data stewardship and deployment. Its Spectra agreement bars training on client data and AI inputs or outputs, and its subprocessor list gives a processing location for each. As of 18 September 2026 the index located no published price and no named model provider.
Questions buyers ask
Epiq AI vs Lighthouse: which is better for eDiscovery AI?
The grid barely separates them: Epiq AI sits in the top two bands on fifteen of fifteen AI Legal Index capability axes and Lighthouse on fourteen of fifteen. Lighthouse publishes more on data protection, with A grades on privilege, stewardship and deployment and a contract barring training on client data. Epiq publishes more about how it prices AI review, per matter with unlimited reclassification, and reserves a right to fine tune its tools on a client's data for that client. The training terms are the clearest point of difference.
Does Epiq train its AI on client data?
Within limits, yes, and its terms say so. Section 15 of the Epiq Software Terms lets Epiq use client data to fine tune or adapt its generative AI tools for the limited purpose of improving outputs specific to that client's environment, and no opt out is stated. Epiq's AI FAQ says it does not train public models with client data, and its Global Terms of Service let it use customer data to improve the services for that customer. 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.
Where does Lighthouse store and process data?
Lighthouse's subprocessor list names Microsoft Azure in the United States and United Kingdom and data centers in Seattle, Phoenix, London, Frankfurt and Dornach, gives a processing location for every subprocessor and affiliate, including support from India and review staffing in named countries case by case, and says RelativityOne runs in the region the customer specifies. Its LighthouseIQ workflows run in a private Azure environment. Epiq describes data centers in ten locations worldwide without naming where they are. 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.
Do Epiq AI and Lighthouse work inside Relativity?
Both do. Epiq AI Accelerators add translation, image analysis, OCR and agentic automation inside RelativityOne and Relativity Server, and on 17 September 2026 Epiq announced Model Context Protocol servers connecting its applications to outside AI assistants. Lighthouse's Spectra promotes processed data into Relativity Server or RelativityOne for review and production, and its IQ Review delivers coding results directly in Relativity. 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 Epiq AI and Lighthouse both leave unpublished?
The models and the prices. Neither names the model behind its review and privilege applications, and neither publishes a rate. Neither says what its question answering assistant does when a dataset holds no answer, or offers a document level record of which model made a call and who verified it. Neither publishes accuracy figures with a test set, and neither addresses how its customers should disclose AI assisted review to their own clients. 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.
Two readings to weigh. Epiq's fine tuning right is limited to improving outputs for the same client's environment, and Epiq's FAQ says it does not train public models with client data; the terms are the binding text, and they permit the narrower use. Lighthouse's contractual bar on training sits in the Spectra agreement; for the LighthouseIQ applications the same position appears in a launch blog rather than a published agreement. Epiq's security page still shows the Privacy Shield mark, a transatlantic framework invalidated in 2020. Both records were verified on 18 September 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.