Casepoint

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 drive a core capability and the platform underneath them is a complete product without them. CaseAssist Active Learning is the engine of prioritised review: it learns continuously from reviewer decisions to surface likely-relevant documents first, and the vendor attributes its 95 per cent faster identification claim to it. The generative layer adds chat-based search over customer data, document summarisation and automatic classification. Strip all of it out and what remains is a full eDiscovery platform that would still be sold: legal hold, in-place preservation and collection from Microsoft 365, Teams, Slack and Google Vault, processing of more than 600 file types, review, production, FOIA workflows, chain of custody and the FedRAMP and Department of Defense hosting that is the company's central pitch. Two things fix this below the top band. The vendor bundles rather than meters the AI, stating that CaseAssist comes built in at no extra cost, so it is not the thing a buyer purchases. And the platform's own framing puts security architecture and workflow unification first and AI third. Graded level with UniCourt and above Docket Alarm and Bloomberg Law, where the machine learning sits further from the core.

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 documented and the measurement is absent, with one unusual strength and one unusual restriction on either side of the line. The strength is that the failure modes are named, and named in the agreement rather than in marketing: the terms tell the customer that AI features including machine learning, technology-assisted review and large language models have limitations affecting output reliability, and enumerate five, being variation in training data quality and diversity, errors from human or machine inputs such as vague prompts, variation in natural language including nuance and sarcasm, factual inaccuracy of models or datasets, and lack of common sense. Almost nothing else in this pull names its failure modes at all, let alone contractually. Grounding itself is real: retrieval-augmented generation is named, answers are drawn from the customer's own data with source citations, and transparency logs are said to record every AI decision, so a reviewer can open the document behind an answer. What is missing is any number. No accuracy figure, recall or precision statistic, validation protocol or test description was located on 31 Aug 2026, which is a conspicuous gap in eDiscovery specifically, where technology-assisted review validation statistics are the currency courts examine. The restriction cuts the other way and belongs on this axis: the terms forbid a customer from disclosing software performance benchmark results to any third party without written consent, which forecloses the independent testing this axis asks about. One limb does not apply rather than failing: a citator or good-law check is out of scope for a platform that searches the customer's own collected documents rather than primary law.

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.

A stated design posture with contractual backing, short of the failure path. The posture is explicit: the vendor says technology should amplify human judgment rather than replace it, describes human-in-the-loop processes where appropriate, and commits in the agreement that it prioritises developing and deploying AI features with human-in-the-loop functionality. Control sits with the customer in a way few vendors state: all generative features are optional and the customer controls adoption. Review surfaces are real and specific, being source citations on generated answers plus transparency logs said to document every AI decision, and active learning is human-supervised by construction since the model is trained by the reviewers' own calls. The agreement also instructs the customer not to rely solely on any output for a purpose with material consequences or affecting anyone's rights. What is not published, checked across the AI strategy page, the platform page, the pricing page, the security page and the full terms on 31 Aug 2026: no threshold at which a feature declines to answer, no described behaviour when the data does not support a result, and no route for reporting or correcting a wrong output. The oversight model describes who is in the loop and not what happens when the loop fails.

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 deepest production evidence located anywhere in this pull, and it is attributed. Named organisations with named individuals and titles appear on the platform page: Marriott International through its senior vice president and assistant general counsel, Mayo Clinic, the West Virginia Office of Technology through its chief information officer, Kramer Levin through two named litigation support managers, Beveridge & Diamond through its chief information officer, Larkin Hoffman through two named staff, and Lewis Roca through a named partner. The pairing that matters most is Lewis Roca, an Am Law 200 firm named alongside a quantified result of more than 90 per cent reduction in document review time attributed specifically to the advanced analytics and AI. Further quantified case studies sit behind their own pages: a 57 per cent reduction in discovery-related costs at a Fortune 500 construction company, a FOIA response-time reduction at a major defence agency, a time-to-insight reduction at a firm that won a motion, and a federal regulatory agency migrating up to a petabyte from a legacy on-premises system. Deployment scale is stated at more than 250 government agencies and roughly 80 per cent of federal agencies. Two qualifications belong in the record. The case studies themselves were not opened on 31 Aug 2026, so their method and dates are unverified. And the pattern across the set is that the quantified studies mostly describe organisations by category rather than name, while the named organisations mostly give qualitative quotes, with Lewis Roca the bridge between the two.

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.

Every element this axis asks for is in the agreement, which is the first time in this pull that is true. Privilege is named rather than implied: the confidentiality clause defines Confidential Information to include information protected by statute or regulation such as Attorney Work Product, Attorney-Client Privilege, personally identifiable information and protected health information. A dedicated clause governs AI specifically. It classifies AI Inputs and Outputs as both Customer Data and Confidential Information; limits their use to providing, securing and supporting the services; commits that AI features will enforce the customer's own access controls and matter or workspace permissions, prevent cross-tenant and cross-matter retrieval, and log administrative access; and commits that Casepoint personnel will not review AI Inputs or Outputs except as necessary for customer-requested support or a security incident, under least-privilege, time-bounded access with audit logging. Segregation is therefore stated at the level this buyer segment requires, at matter level, in writing. Support access is governed to the same standard, with no standing or default remote access, enablement only by the customer, named individual support personnel under multi-factor authentication, and audit logs made available to the customer. One tension is recorded rather than ignored: a separate clause grants Casepoint a transferable licence over Customer Data and usage data to improve its products and services, which reads broadly, though the AI clause carves AI Inputs and Outputs out of exactly that use. A buyer should read the two together and understand that the carve-out protects prompts and outputs specifically rather than the whole collected corpus.

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.

A real position on tooling versus judgement, published where it binds, and silent on the professional framework around it. The agreement tells the customer that it should not rely solely on any output from the software including AI features for any purpose that may have material consequences or affect the rights of any person or entity, and elsewhere that the customer alone is responsible for its reliance on the results of its use, including their completeness, accuracy and content. The competence and supervision dimension is present rather than assumed, through the enumerated AI limitations, the human-in-the-loop commitment and the vendor's stated position that technology should amplify human judgment rather than replace it. What is absent, checked across the AI strategy page, the platform and security pages and the terms on 31 Aug 2026: no bar or ethics authority is named anywhere, including ABA Formal Opinion 512 and any state guidance; no jurisdictional limit is stated; and nothing addresses the professional obligations of the reviewers who will act on the output. One limb does not apply and is not penalised: the product has no consumer-facing surface, since it is sold to legal departments, agencies and firms and reached through an authenticated platform, so consumer disclosure is not a question it raises.

AI Governance and Bias Disclosure

Published governance over model behaviour: 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 framework whose substance sits in the agreement rather than on the poster, and a complete silence on the second half of this axis. The public framework is TRUST, with five named principles, and read alone it would be principles rather than mechanism: transparency about data use, methodologies and an opt-out; reliability framed as transparent, auditable and defensible; user-centricity through human-in-the-loop processes and source citations; security; and a stated refusal to chase hype. What lifts it is that the commitments recur in binding form. The terms limit the purposes for which AI Inputs and Outputs may be used, restrict personnel review of them, require enforcement of access and matter permissions inside AI features, and mandate logging of administrative access. Transparency logs recording every AI decision are an auditable artifact rather than a claim, and a separate clause gives customers above a stated spend an annual audit right over privacy and security controls. That is a mechanism a buyer can hold the vendor to. What is missing is everything about bias. No fairness or uneven-output evaluation, no testing regime described, no accountable owner named, and no acknowledgement anywhere that a system which decides the order in which a human reviews documents can be uneven across custodians, languages or document types. For a review-prioritisation engine, that is the governance question, and it is not addressed.

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.

Most of the ground is covered contractually and with more precision than anything else in this pull, and one element is missing. Incident practice is the standout: the agreement commits to written notification within seventy-two hours of a validated security incident, with periodic updates covering the cause, the mitigation under way, the anticipated impact on the customer and the expected remediation timeframe. Deletion and return are specified, with an export of all Customer Data on termination in a common format, a seven-day window, and an explicit statement that backup media copies may persist under standard procedures. Access control is described at several levels: role-based access controls, zero-trust network architecture with continuous threat detection, encryption in transit and at rest, and a remote-support regime with no standing access and customer-controlled enablement. Personnel controls are unusually specific, with ten-year criminal background checks, OFAC screening, and social security, employment and education verification. Disaster recovery and backup plans are reviewed and tested annually with copies available on request. What holds this below the top band is the subprocessor list: the agreement commits Casepoint to maintaining a list of material third-party components and subprocessors, and no such list was located as published on 31 Aug 2026. One research limit to state: the privacy policy and the data processing agreement were not opened, and either could carry the list.

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.

The strongest published liability position in this pull, and it still stops short of standing behind the output. What a buyer gets in writing: a general cap at fees paid under the relevant order in the six months preceding the claim; a materially higher cap of five times annual fees plus direct damages for breaches of confidentiality, data breaches, gross negligence and wilful misconduct, which is the risk that actually matters when a vendor holds a firm's collected evidence; named carve-outs for personal injury or death, fraud and anything not excludable by law; a defence and indemnity from Casepoint against claims that the software infringes a US copyright or misappropriates a trade secret, with a defined procedure and named remedies of modification, licensing, substitution or refund; an affirmative sixty-day warranty that the software will perform substantially in accordance with its documentation; a services warranty of professional and workmanlike performance; and a 99.5 per cent uptime commitment with a published service-credit formula. That is considerably more than a disclaimer. What is not there is the thing this axis asks for last. No warranty attaches to AI output, and the agreement instead disclaims reliance on it and places responsibility for accuracy on the customer. No insurance position was located. And the general indemnity runs from the customer to Casepoint on data, legality and breach caused by customer systems. The allocation is precise, readable before signing, and asymmetric on the question of whether the AI was right.

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, named connections into the systems where the data actually lives, described at the level of what they collect rather than how they are configured. Named on the vendor's own pages: Microsoft 365, Microsoft Teams, Slack and Google Vault, with collection performed in place and directly from the source rather than through manual export by the customer's IT function, and support for more than 600 file types on processing. An API is described as automating collection from those platforms. The most current addition is the Casepoint MCP Server, offered as a way to connect an organisation's own approved AI systems to the platform, which is an integration surface pointed outward at the buyer's AI stack rather than at its document systems. What is absent is the other half of a firm's estate: no document management integration is named, with nothing located for iManage or NetDocuments, and nothing for matter management, e-billing or court filing. No public API documentation was located, and the terms restrict use of APIs not supported by Casepoint. Two research limits: the data connectors page and the MCP Server page were not opened on 31 Aug 2026, and either could carry the configuration and direction detail this grade turns on.

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.

The deployment model is stated clearly and residency is answered for some data and not for the rest. Three delivery options are described in the agreement rather than merely implied: Casepoint-hosted software with a 99.5 per cent uptime commitment; customer-operated or third-party-hosted installation, permitted subject to named conditions on notification, secure storage and flow-down of terms; and separately certified government environments. Which certifications attach to which environment is stated, with SOC 1, SOC 2, SOC 3, ISO 9001, ISO 27001 and the NIST publications as the base commercial offering and GovRAMP, FedRAMP Moderate, FedRAMP High and Department of Defense Impact Levels 4, 5 and 6 for certain environments and applications, so a buyer can tell what changes between tiers. Residency is committed for one defined class of data: for export-controlled material the agreement requires storage and processing only within the United States in Casepoint-controlled environments with no replication to non-US locations, and restricts access to US persons. What is not published is the general case. No region list for the commercial offering, no tenancy model statement, and no distinction between where data is stored and where it is processed outside the export-control clause. A Canada government offering is referenced in the navigation without a residency statement attached.

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.

Current independent attestation, with scope named and verifiable without asking the vendor anything. The security page lists fourteen certifications and authorisations, each with explanatory text rather than a bare badge: FedRAMP High and FedRAMP Moderate, Department of Defense Impact Levels 4, 5 and 6 with authorities to operate from the Defense Information Systems Agency, GovRAMP, SOC 1 Type II issued under SSAE 18, SOC 2 Type II, SOC 3, the Shared Assessments SIG questionnaire self-assessed annually, ISO 9001:2015, ISO 27001:2022, NIST 800-53, NIST 800-171 and the EU-US Data Privacy Framework. The government authorisations are the decisive point for this axis, because a FedRAMP authorisation and a DISA authority to operate are third-party assessed and independently listed by the authorising bodies, so a buyer can confirm the status without a sales conversation and without relying on the vendor's own page. Scope and period are stated for the SOC 2 Type II, described as an independent third-party assessment issued annually against the AICPA Trust Services Criteria covering security, confidentiality, availability, processing integrity and privacy over a twelve-month period. The agreement then restates the whole set contractually and distinguishes which certifications apply to the base commercial offering from those applying to certain environments, which is a scope statement most vendors never make. Two honest gaps: no auditor is named for the SOC or ISO work, and no report date appears on the page, with the reports and the SIG available to customers on request under the audit clause.

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 the providers are not. What is published: retrieval-augmented generation named as the mechanism for generative answers, with the retrieval corpus identified as the customer's own data rather than an external body of content; the technology classes named in the agreement as machine learning, technology-assisted review and large language models; and a contractual commitment that Casepoint will ensure proper licensing of third-party software, patch critical vulnerabilities in it, and maintain a list of material third-party components and subprocessors. That commitment is more than most vendors make, and the list itself was not located as published on 31 Aug 2026. What a buyer cannot establish: no model or provider is named anywhere, no version is given, no inference location is stated as distinct from the data residency commitments, and no notice obligation attaches to a change of model or training data. One structural feature cuts across this axis and is worth recording: the separately offered MCP Server is described as connecting the organisation's own approved AI ecosystem to the platform, which lets a buyer supply models it has already vetted and shifts part of the supply-chain question onto its own governance rather than answering it.

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.

A page titled Transparent Pricing that contains no price, and the gap between the two is the finding. What the pricing page does publish is the shape of the deal: pricing is described as tailored across an extensible platform and specialised government products, priced to the applications and use cases actually taken, with the explicit statement that a customer does not pay for a product it does not use. The purchasable units are named across the site as Casepoint eDiscovery, Legal Hold, FOIA, Filestore, Casepoint AI, Investigator, the MCP Server and Data Connectors, alongside five separately named government products, so the modular structure is visible. The agreement adds real billing mechanics a buyer can read in advance: usage-based fees calculated on the greater of maximum actual usage or the agreed cap, with the units named as data storage, processing and user counts; net-thirty payment; interest at 1.5 per cent per month; one-year initial terms with automatic annual renewal and sixty days' notice to prevent it; and suspension rights at forty-five days past due. What is nowhere published is a number. No rate, no per-gigabyte or per-user figure, no band and no starting point, and the single call to action on the pricing page is a demo request. Checked the pricing page, the platform page and the full terms on 31 Aug 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 use cases are documented in unusual detail and the boundary is left open. Coverage is published along two axes. By market: federal civilian, federal defence, state, local and education, Canada, and corporate and enterprise. By team: corporate legal, government legal, FOIA and public records, government audit and investigations, and information technology and information security. A law firm solution page sits alongside them, so all three buyer types are addressed rather than assumed. The use cases are enumerated rather than gestured at: legal hold, investigations, litigation, eDiscovery, data subject access requests, third-party subpoena responses, FOIA and public records requests, and congressional inquiries. Depth claims are quantified, with more than 600 file types processed and named collection sources. What is missing is any statement of where the platform stops. No practice area is excluded, no data source is identified as unsupported, no matter size or volume threshold is named, and the Canada offering appears in the navigation without a description of how its coverage differs. Checked the platform page, the pricing page, the security page and the site navigation on 31 Aug 2026; the individual solution pages were not opened.