CaseMark vs Syllo: how they compare in 2026
CaseMark and Syllo both read a matter's record and produce work from it with citations back to the source: CaseMark drafts summaries, chronologies and documents across more than 880 agent skills, and Syllo tags documents for responsiveness and issues at scale and answers questions across the case record. The grid does not separate them. Each sits in the top two bands on ten of fifteen axes, level on seven with four going each way, and the tie hides two different kinds of proof. Syllo proves performance: a white paper of 21 March 2025 reports recall and precision measured by named firms' own teams on live reviews, including a Ballard Spahr review of more than 100,000 documents at an estimated 99.4 percent recall. CaseMark proves terms: its agreement states that customer data does not train its models and is deleted after 30 days, and it publishes a price of $100 per user per month. Syllo's terms license content to improve the service without naming training either way, and it publishes no price. Both cap their liability at $100 in their published terms.
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 mechanism and also the unit of charge. Workspace bills 100 dollars per user per month including 80 dollars of AI credits, with additional usage billed at market rates, so a customer is paying for inference directly. Access to more than 880 legal agent skills is the headline entitlement, the positioning is agent-native legal work, and case.dev exposes the same capability as eighteen API services. Remove the models and what remains is file storage and delivery. Checked 4 September 2026.
The models drive the capabilities the vendor leads with, on a platform whose conventional layer would still function without them, which is the B band. Agentic Issue Tagging coordinates multiple language models in distinct roles to run first and second level review with unlimited issue codes, and the open prompting engine answers questions across the case record and drafts summaries, deposition digests and chronologies. The vendor describes itself as a Litigation AI platform and says it was founded to build the data and workflow infrastructure into which language models are woven. Beneath that sits a full eDiscovery module that stands on its own: native and load file processing with deduplication, threading and family grouping, Boolean and metadata search, multi-panel review with bulk redaction, Bates stamping, load file generation and privilege log automation, plus case management covering chronologies, exhibits, binders and transcript designations. Strip the models out and a litigation support team still has a review and case management platform, which is why this is B rather than A. The Litigation Agent, which would push further toward the A description, is in early access and not credited. 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 is real and documented as a product property rather than a slogan. The platform is described as reading every page and drafting the document the matter needs while citing back to the source, deposition summaries are offered in page-line format, and medical chronologies are described as cited timelines from the records. Failure is acknowledged in the agreement rather than only in marketing: the content accuracy clause states that use of the Service may result in Output Data that does not accurately reflect people, places or facts and should be checked, and requires review for accuracy, incorrectness, completeness and offensiveness before reliance. What is missing is measurement. No accuracy figure is published anywhere, no test set is described, no evaluation is linked, and the vendor names no specific failure mode of its own beyond the general acknowledgment.
Measured accuracy is published for document classification and grounding is documented for generated output, short of measurement of the generated output itself, which is the B band. The white paper of 21 March 2025 reports that across the last ten completed responsiveness reviews in live litigation the lowest estimated recall was 93.4 per cent, the average 97.8 and the median 99.4, with median estimated precision of 85.9 per cent and average 79.7, and it gives per-matter figures with dataset sizes and the firm's own precision and elusion testing as the method. Every tagging determination carries an issue-specific rationale pinpointing the text that triggered it, and answers from the prompting engine come with hyperlinked citations to the case record. Three things hold it below A. The measurements cover responsiveness classification, not the accuracy of generated answers, summaries or digests, for which nothing is published. The failure modes the paper names are those of competing approaches, prompt overload, context windows and hallucination in linear review, rather than this system's own. And nothing states what the prompting engine does when the record does not support an answer. The benchmark is eighteen months old at this date. Grounding is to the customer's own record rather than to primary law, so citation status checking does not arise. 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.
Autonomy is claimed at scale and oversight is asserted without a mechanism. The product promises to go from gigabytes of discovery to a motion for summary judgment in under an hour, offers more than 880 agent skills and 880 workflows, and a published testimonial describes a transcript dropped at 9pm and a finished summary sent to a partner in the morning untouched. Against that the only oversight statement is the agreement's instruction that the customer should review and evaluate Output Data before relying on it. Nothing published describes what runs unattended against what a person approves, no threshold is stated, no review surface inside the product is named, and nothing addresses what happens after an output is wrong. Searched the home page, the pricing page, the security page and the terms of service on 4 September 2026; the workspace, solutions and workflows pages were not opened.
Real review surfaces and a stated validation route, short of a control structure with a threshold or a categorical constraint, which is the B band. The white paper describes the review system as coordinating multiple language models that decide how to conduct the review within guidelines set by users. The product page states that case teams validate the outcomes of the agentic issue coding to ensure defensibility for outgoing productions, every determination carries a written rationale pointing to the triggering text, tag conflict detection is built into quality control, and the white paper's case studies show the validation in practice as precision testing on tagged sets and elusion testing on the null set. Customer and Case Administrator roles in the terms control who can act on a case. What the A band asks for is not published: no threshold at which the system stops or escalates, no mode or tier carrying a stated constraint on what its output may be used for, and nothing on what happens after a coding error reaches a production. The home page headline claims the platform can autonomously manage any matter end to end; that describes the Litigation Agent, which is in early access, and no oversight structure is published for it. 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 substance, short of measurement. Named customers span law firms and enterprises: Nelson Mullins, Ogletree Deakins, Pathway Law Firm, Planet Depos and Fortz Legal Support on the legal side, and Accenture, PwC, Apple, MLB and Google as enterprise logos, with a stated base of more than 7,000 attorneys and customers. One testimonial is fully attributed with an operational account rather than a platitude: Collin Ritsema, Chief Operating Officer of Fortz Legal Support, describing a 400-page transcript submitted at 9pm and returned by morning cited and formatted, and passed to a partner without editing. What holds this at B is that no figure is attached to any named customer, nothing is dated, no case study is published, and the enterprise logos are presented as customers without any indication of which product or scale of use.
Named firms, a date, figures for what changed and a method a reader can assess, which is the A band. The white paper of 21 March 2025 lists practitioner contributors from seven firms, Ballard Spahr, Mayer Brown, Nixon Peabody, Outten & Golden, Pillsbury Winthrop Shaw Pittman, Quinn Emanuel and Royer Cooper Cohen Braunfeld, and reports matter-level results measured by the firms' own teams. Ballard Spahr's trial team ran precision and elusion testing on a review of more than 100,000 documents against more than 25 issue codes and found estimated precision of 95.56 per cent and recall of 99.4 per cent, after about three hours of set-up. Royer Cooper Cohen Braunfeld ran a head-to-head on a random sample of just under 16,000 documents from more than 150,000: the managed review team using continuous active learning reached estimated recall below 67 per cent at over two dollars a document, and the system reached recall of 93.44 per cent and precision of 69.81 per cent. Outten & Golden reports precision of 84.09 per cent and estimated recall of 100 per cent on 12,543 documents with 28 tags, and Quinn Emanuel a zero-shot review of more than 40,000 documents at 98.69 per cent recall and 92.83 per cent precision. The limits are stated here so the grade is read correctly: the paper is vendor published, the matters are not named, practitioner contributors wrote in their individual capacities, and the paper's own footnotes disclose that one technical contributor is a research advisor and one practitioner a product advisor to the vendor. 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.
Substantive published commitments across most of what this axis asks, with two clear gaps. Training is prohibited in the agreement itself, which states that customer data is not used to train the company's AI models or sold to third parties, and the security page repeats it without qualification. Retention and deletion are the strongest part and are unusually specific for this market: customer data is retained for thirty days and then permanently deleted from systems, with extended retention available to enterprise customers on request. Access control is documented through least privilege by default, roles and permissions restricting access to sensitive data, and single sign-on with multi-factor authentication required on all accounts, alongside AES-256 at rest and TLS 1.2 in transit and comprehensive audit logging. Isolation is asserted and, at enterprise tier, offered as dedicated compute with isolated inference pool options. What is absent is privilege and work product treatment, which appears nowhere on a product ingesting depositions and medical records, and any position on what a model provider may retain, since no model provider is identified at all.
Substantive written commitments, including express treatment of privilege and work product, defeated on training, which is the B band. The Legal Disclosure and Subpoena Policy, incorporated by the terms at section 9, states that much of what users store is confidential and privileged under the attorney-client privilege and the work product doctrine, argues that the customer rather than the vendor is the proper target of legal process, and commits to written notice so a customer can seek a protective order. The terms add that Content is encrypted in transit and at rest, that the vendor acquires no rights beyond those needed to offer the service, that Content is not shared with third parties except service providers, legal process and security monitoring, that vendor personnel may not view Content without permission or a support, security or legal reason under internal controls (sections 13.1 and 16.6.2), and that Content may not be used outside provision of the service (16.6.4). Segregation is documented at case level through Case and Customer Administrators, and the product page states matter-level isolation is standard. What fails the A band is training: section 8.4.1 licenses Content for providing or improving the service, and no commitment against model training is published anywhere. The position of any third-party model provider is not stated. The customer Licensing Agreement, which prevails over the terms, is not published. 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 real position is published on advice versus tooling, short of the full treatment. The agreement carries a dedicated No Legal Advice clause stating that the company does not provide legal advice and is not engaged in the practice of law, that content generated through its artificial intelligence tools is for informational purposes only, that no attorney-client relationship is established, that the services do not cover all legal situations, and that the reader should seek advice from a qualified attorney licensed in their jurisdiction. That is more than a boilerplate footer. Jurisdiction limits are addressed twice and unusually plainly, with the service offered for use only by persons located in the United States and a restriction on transferring GDPR personal data without consent. Users must be at least 18. A separate restriction prohibits representing Output Data as human-generated when it is not, which is a professional honesty provision most vendors omit. What is missing is the competence and supervision dimension: nothing addresses the professional duties a lawyer owes when using the tool, and no rule of professional conduct or ethics guidance is named.
The terms place professional responsibility on the user and widen the audience, while the marketing describes the product in terms of legal analysis and strategy, which is the C band. Section 5.4 of the terms states that the service is not intended for use solely by attorneys, judges, paralegals or others in the legal profession and may be used by anyone holding a licensing agreement, and that the vendor has no responsibility to determine or monitor whether a user is engaged in the unlicensed practice of law. Section 5.8 makes each user responsible for ensuring they are authorised to practise where they use the service for the practice of law, and section 15 disclaims any warranty that content provided through the service is accurate. Against that, the product is marketed for work product generation, legal analysis and case strategy, and the home page headline speaks of autonomously managing a matter. No statement says the product does not give legal advice, nothing addresses a supervising lawyer's competence or supervision duties, and no jurisdictional limit is stated beyond a US-only account rule that the terms tie to privacy and data protection law rather than to practice. 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.
No governance position is published for a system that drafts filings from the case record. There is no responsible AI statement, no governance framework, no named owner accountable for model behaviour, no description of pre-release evaluation or testing of output quality, no published results, and nothing whatever on bias or uneven performance across record types, matter types or populations. The nearest published sentence is the agreement's assurance that the service is continuously updated and improved to ensure reliability and accuracy, which is an aspiration rather than a mechanism. The security estate is substantial but answers a different question: penetration testing, vulnerability assessment and audit logging are security controls, not model governance. Searched the home page, the security page, the pricing page, the terms of service and the privacy policy on 4 September 2026.
An evaluation practice is published without any governance framework, owner or pre-release testing regime around it, which places this nearest the C band. What exists is real: the white paper describes how outputs are validated in live matters, through precision testing of tagged sets and elusion testing of the null set, reports aggregated results across ten reviews, and names a Carnegie Mellon Language Technologies Institute researcher as a technical contributor. The work with Syllo page says the AI systems are designed for rigorous human review and feedback. The terms separate Beta Previews from the audited service, stating early access features are not subject to the same security measures and auditing. None of that is governance in the band's sense. No responsible AI policy or principles page was located, no individual or function is named as accountable for model behaviour, nothing describes what is tested before a model or feature is released, and uneven output across matter types or parties is not addressed. The band's C description speaks of principles without a mechanism, and this record is closer to the reverse, a measurement practice without stated principles; C is the nearest fit. 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.
Substantive published policy covering most of the ground. Retention is a stated period rather than a gesture, at thirty days followed by permanent deletion, with enterprise extensions on request. Access control is specific: least privilege by default, roles and permissions over sensitive data, single sign-on and multi-factor authentication required on all accounts, AES-256 at rest, TLS 1.2 in transit, and comprehensive audit logging of all user activities and system events. Testing cadence is published rather than asserted, with annual external penetration testing, quarterly vulnerability assessments and daily dependency updates, alongside background checks and annual mandatory security training for all employees. Some suppliers are named, being AWS and Azure for hosting and Stripe for payment processing at PCI Level 1. Two elements of the set are missing: no incident or breach notification practice was located anywhere, and no subprocessor list covering the AI layer is published. The privacy policy that would ordinarily carry the first was last updated 23 July 2024.
Substantive published controls across most limbs, short of incident practice and a processor list covering the AI, which is the B band. The security section of the Work with Syllo page states encryption in transit and at rest in line with NIST SP 800-57, automatic backups, regionalised storage, continuous vulnerability scanning and penetration testing, and multi-factor authentication. The terms name where Content is stored, Amazon Web Services virtual servers and Amazon S3 controlled by the vendor and ElasticSearch for indexing (section 13.3), restrict personnel access to Content (13.1), require single-person credentials and user notice of a compromise within one business day (2.1), reserve deletion of a cancelled user's personal knowledge bases within 90 days with residue in encrypted backups (2.3.2), and aim for 30 days' notice before a feature is discontinued so users can export (16.3). Three limbs fail. No commitment to notify customers of a security incident was located. No retention period is stated for prompts, outputs or case content during or after a subscription, and section 2.1.8 places export on the user with no obligation on the vendor after termination. No list names who processes content for the AI features. The product Privacy Policy and Data Security Policy, which the terms place at app.syllohq.com/policies, return an empty page to automated retrieval and could not be read on this date; they are the documents most likely to move this row. 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.
Liability is addressed only through a standard limitation clause that disclaims the exposure the product creates. The cap is the striking feature: cumulative liability for any damages arising from or related to the site, platform or services is limited to a maximum of one hundred US dollars, with the agreement stating expressly that more than one claim does not enlarge the limit. The service is provided as-is and as-available with all warranties disclaimed including accuracy and non-infringement, and no uptime guarantee is offered. The indemnity runs in one direction only, from customer to company, with no intellectual property or any other indemnity given to the customer. No insurance position was located. What keeps this off the floor is that the allocation is published, specific and readable before signing rather than absent; a buyer can establish exactly what recourse exists, and the published answer is one hundred dollars.
The published terms address liability only through a disclaimer, a nominal cap and an indemnity running to the vendor, which is the C band. Section 15 provides the service on an as is, as available basis, disclaims any warranty that content provided through it is accurate, reliable or correct, excludes all categories of damages, caps the vendor's total liability to an end user at one hundred dollars, and excludes liability for breaches at third-party providers. The indemnity runs the other way: each user indemnifies and defends the vendor for their use, their content and their breaches. Nothing addresses AI output specifically, no vendor indemnity is offered, and no insurance position is published. The terms bind end users and state that the customer's Licensing Agreement prevails where the two conflict, so the terms a firm actually negotiates may differ, but that agreement is not published and a buyer cannot read it before a sales process. 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.
Connectivity is published as a platform layer without any named practice system behind it. case.dev is presented as the infrastructure everything CaseMark runs on, available as an API across eighteen services, with public developer documentation at docs.case.dev, and Source is built to deliver files into court reporting and litigation support workflows with branded client portals. That is real integration surface. What is absent is the thing this axis asks for: no document management system, case management platform, e-discovery tool, word processor or e-signature product is named anywhere on the surfaces read, so an implementer cannot see what CaseMark connects to as opposed to what it exposes. The developer documentation was not opened in this pass and the workflows page listing 880 workflows was not opened either; both are named so the limit is visible, and they are the cheapest available upgrade on this record.
The product stands alone by design, which is the D band read literally. The vendor argues on its own platform page that API integration and acquisition have both failed to solve fragmented litigation tooling, and positions a single environment holding the whole record as the answer. What it supports is interchange rather than integration: native file and load file ingestion with deduplication, threading and family grouping, load file generation and Bates stamping at production, and import of standard court-reporting formats in text, audio and video. Those are the formats data arrives and leaves in, and they mean moving a matter onto the platform is routine, but no connection to any external system is named. No document management system, no Microsoft Word or Outlook add-in, no review platform, matter management, e-billing or court filing integration was located on the surfaces read, and no API or developer documentation is published. The grade records a stated design choice, not a gap the vendor has left unexplained. 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.
The deployment model is stated clearly with partial residency detail. Residency is unambiguous and named to the provider: all data is stored in the United States within AWS and Azure environments, with both providers stated to include guarantees on intrusion detection and physical security. Tenancy appears as a published tier difference rather than a claim, with dedicated compute and isolated inference pool options offered on the enterprise plan against the shared default, which answers what changes between tiers better than most records on this axis. What is missing is the distinction the top band turns on: nothing states where processing or model inference happens as against where data is stored, and no region choice is offered, the United States being the only location and also the only market the service is sold into.
Residency is offered and the hosting provider is named, with the deployment options themselves left unspecified, which is the B band. The security section states that multiple secure deployment options are available, that data storage is regionalised based on customer needs, and that for firm-level engagements the vendor will work with a firm's security team to customise deployment to its requirements. The terms name the infrastructure: Content is stored on Amazon Web Services virtual servers and Amazon S3 storage controlled by the vendor, with ElasticSearch for indexing (section 13.3), and user accounts may not be used outside the United States, a restriction the terms tie to privacy and data protection law (sections 2.1 and 5.6). What is missing for A is specificity. The deployment options are not named, so whether single-tenant, private cloud or on-premises arrangements exist cannot be established, no regions are listed, and nothing distinguishes where model inference runs from where data is stored. 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.
Certification is real, stated and scoped, and the route to evidence is better named than most records in this band. SOC 2 Type II is claimed and its coverage described as an independent audit of security controls, availability and confidentiality, and HIPAA compliance is claimed with a BAA available on the enterprise plan. The artifacts available are itemised rather than gestured at: SIG Lite 2024, CAIQ questionnaires, penetration test summaries and attestation letters, all stated to be available on request, which is a request flow rather than a sales gate. Testing cadence is published at annual external penetration testing and quarterly vulnerability assessments. A trust centre operates at trust.casemark.com and was not opened in this pass. It holds at B because no certifying body or auditor is named, no report date, observation period or certificate number is published, and on the third-party verifiability test a buyer cannot check either claim against a register without contacting CaseMark.
A named attestation with stated scope and a named auditor, short of any route to the report, which is the B band. The compliance section states SOC 2 Type II certification in Security, Confidentiality, Availability and Privacy, compliance with the AICPA SOC standards for service organisations, and third-party validation by Prescient Assurance. Naming four of the five trust services criteria and the attesting firm is more than most records at this band carry. What keeps it off A is access and currency: no trust centre was located on the site or by search on this date, no report period or issue date is given, and no request route, self-serve or otherwise, is described. Continuous vulnerability scanning and penetration testing are stated on the same page, without summaries. 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.
Nothing is published about the model supply chain a customer inherits. No model is named, no model provider is identified, no subprocessor list covering the AI layer exists, and no commitment to notify customers of a change to any of it was located. AWS and Azure are named, but as the environments in which data is stored, and infrastructure never answers this axis: naming where a model runs is not naming whose model it is. The enterprise tier offers dedicated compute with isolated inference pool options, which confirms that inference happens somewhere without saying on whose models. This is the largest disclosure gap on an otherwise well-documented record, and it sits against a customer base that includes AmLaw firms and large enterprises. Searched the home page, the security page, the pricing page, the terms of service and the privacy policy on 4 September 2026.
The architecture is described in unusual detail and no provider or model is named, which is the B band. Product pages state the platform is powered by a combination of proprietary Syllo models and foundational language models. The white paper goes further, describing an ensemble of varying-sized language models performing distinct roles, strategising, determining next steps, quality control, extracting learnings from documents, synthesising across documents, summarising and resolving disagreement between models, with telemetry routing harder material to higher-end models and simpler material to cheaper ones. A buyer learns how the system is built. The buyer does not learn whose models these are: no provider, model family or version is named on any surface read, nothing states where inference runs as distinct from the AWS storage named in the terms, and nothing addresses what a model provider may retain. On change, the terms run the other way, reserving the right to modify any feature at any time with or without notice (section 13.7). 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.
Real pricing is published for part of the range with the upper tiers withheld. Workspace Pro is published at 100 US dollars per user per month, with the entitlement itemised: 80 dollars of AI credits per user each month, document analysis and AI chat, access to more than 880 legal agent skills, one gigabyte of document storage and email support. The charging model is explained rather than left implicit, with included AI usage scaling with the monthly commitment and additional usage billed at then-current market rates. Teams and Enterprise are quoted rather than priced, and what each adds is listed, covering shared matter space, single sign-on, storage, a dedicated success manager, a BAA and dedicated compute. Pricing is organised by product across Workspace, Structured Summaries and Source, and the Summaries and Source tabs were not opened in this pass. The agreement adds that subscription price changes take effect only at the next term with 60 days written notice, and that pay-as-you-go purchases are non-refundable.
No price, unit of charge, tier or inclusion list is published, which is the D band. Searched the home page, the Work with Syllo page including its pricing section, the Unified Platform and Agentic Issue Tagging pages, the Terms of Service and the white paper on 18 September 2026. The pricing section names two buyer segments and routes both to a sales conversation: in-house teams are offered the platform on an enterprise basis, and law firms on a firm-wide enterprise partnership. Firm-wide describes the scope of a licence rather than a unit a buyer could price. The only figures on the estate are competitors' costs, the white paper reporting a managed review team at more than two dollars per document, alongside claims of review at a fraction of the time and cost. No pricing row is written. 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.
Segment coverage is described with substance and the practice edge is left open. Buyer segments are named and tiered explicitly: solo practitioners and individual legal professionals, growing firms at two to ten users, and larger legal teams at ten or more, with a separate published route for court reporting and litigation support businesses through Source. The customer set corroborates the breadth, spanning defence firms, a plaintiff firm and corporate legal functions. Record types supported are stated concretely through the summaries range, covering depositions, medical records, transcripts, recordings and contracts. The jurisdictional boundary is stated plainly and twice, with the service sold within the United States and intended for US-based audiences only. What is missing is the practice dimension: no practice area is named as supported anywhere, in-house and government use are not addressed separately, and nothing says which matter or record types the product does not handle. The solutions and workflows pages were not opened.
Segments and matter types are described with evidence behind them, short of the boundaries, which is the B band. The vendor names AmLaw 100 firms and boutique litigation firms, and in-house teams at corporations managing complex litigation portfolios, with a dedicated in-house page. The white paper evidences the breadth with named firms from AmLaw defence practices to a plaintiffs' employment firm, and lists matter types the system has been used on: antitrust, environmental, contract, employment, patent, bankruptcy, mass tort, construction, investment and shareholder disputes, M&A, automotive, real estate and insurance coverage litigation, across datasets from thousands to more than two million documents. The scope is litigation throughout. What is left open is where it stops: nothing addresses government or court use, nothing names matter types or data types it handles poorly, and the one stated boundary, that accounts may not be used outside the United States, is framed as a data protection rule rather than a coverage limit. 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 prohibition sits in the agreement rather than only on a policy page. The Content section of the terms of service grants CaseMark a license to copy, distribute and display User Data only for the purpose of providing the Service, and then states that it does not use customer data to train its AI models or sell it to third parties. The security page repeats the position without qualification, stating that CaseMark does not train AI models on customer data and that documents, transcripts and work product remain the customer's. No de-identification, anonymization or aggregation carve-out appears anywhere, and no opt-in or opt-out mechanism is described because none is needed.
The published terms bound the vendor's use of customer content to providing and improving the service, and no surface names model training in either direction. Section 8.4.1 grants a limited license to use Content for providing or improving the service, fixing issues and support, or as the user or customer instructs; section 16.6.4 states the license does not permit using Content outside provision of the service; and section 8.1.3 says the vendor acquires no other rights.
The grant is an improvement right that never names training, which is recorded here because it is the tension: the Work with Syllo page adds that lawyers and firms play a critical role in shaping and improving Syllo's AI systems. The customer Licensing Agreement, which prevails over these terms, is not published, and the product privacy policy could not be read on this date.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
A specific period is published in the agreement and the customer cannot change it on standard terms. The Data Retention and Deletion clause states that customer data is retained for thirty days, after which all customer data is permanently deleted from CaseMark's systems. The one qualification is that exceptions can be made for enterprise customers requiring extended retention, which lengthens rather than shortens the window and is arranged by contacting the company rather than configured by the customer. Zero retention is not offered as a setting.
Retention is acknowledged in the terms without a period for prompts, outputs or case content. Section 2.3.2 says that when a user cancels an account the vendor may retain the content of that user's personal knowledge bases as needed for legal obligations, disputes and enforcement, and reserves the right to delete it within 90 days, with some information possibly remaining in encrypted backups. Section 2.1.8 places export on the user and states the vendor has no obligation to provide case or knowledge base content after the licensing agreement ends.
Nothing states how long prompts, generated answers, rationales or case records are kept during a subscription or after it ends, and no customer setting for retention is described. The product privacy and data security policies could not be read on this date.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
CaseMark maintains its own permission model and describes it at feature level. The security page publishes least privilege by default, user roles and permissions restricting access to sensitive data, and single sign-on with multi-factor authentication required on all accounts, alongside comprehensive audit logging of all user activities and system events. The home page adds that data is encrypted and isolated, and the enterprise tier offers dedicated compute with isolated inference pool options.
The model is CaseMark's own rather than one inheriting a document management system's access control at query time, and the customer administers its own roles. Nothing published addresses segregation between individual matters within a customer account.
The product maintains its own documented permission model organized by case, which a firm keeps aligned with its own walls. The terms define Cases, each with a Case Administrator who controls access, and a Customer Administrator with ultimate control over every case and knowledge base at the customer (sections 1, 2.2 and 6); a user granted access to a case reaches all of its content through every case application, and content posted to a case or shared knowledge base is automatically visible to others with access (8.2.2).
Personal knowledge bases can be kept private. The product page states that role-based permissions and matter-level isolation are standard. Nothing connects this model to a firm's document management permissions, so walls are set inside the platform rather than inherited.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
The privacy policy addresses disclosure under legal process directly, stating that CaseMark may share information in response to a request for information where it believes disclosure is required by or in accordance with any applicable law or legal process, including lawful requests by public authorities. So the customer is told data can leave. No commitment to notify the customer of such a request was located anywhere, and no discretion over notice is reserved either.
Searched the terms of service, the privacy policy, the security page and the home page on 4 September 2026. Recorded for context: the privacy policy carrying this clause was last updated 23 July 2024 and still describes products, CaseMark Workflow and CaseMark Productivity, that no longer match the current range.
The Legal Disclosure and Subpoena Policy, which the terms incorporate at section 9, commits to written notice to a current or former customer of a civil subpoena or a criminal subpoena, warrant or similar process, so the customer can seek a protective order, and to furnish only what the vendor determines is legally required while seeking confidential treatment. The exception is where notice would breach a court order or could expose the vendor to criminal liability.
The policy argues that the vendor is an inappropriate target because the information is often privileged and remains in the customer's control under Federal Rule of Civil Procedure 34(a)(1). Users who are not customers get no notice commitment. No transparency report was located.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
No located public material identifies a corpus behind the product's output, and the product's design makes the question narrow: CaseMark reads the record the customer uploads, depositions, medical records, transcripts, recordings and contracts, and drafts from it rather than retrieving external legal content. No database, publisher, jurisdiction or license basis is named on any surface. Searched the home page, the summaries and features navigation, the security page, the pricing page and the terms of service on 4 September 2026.
Searched the home page, Unified Platform, Agentic Issue Tagging and Work with Syllo pages, the white paper and the terms on 18 September 2026. No primary law corpus is identified. The product works over the customer's own case record, and its citations point into that record. The home page headline lists legal research among the platform's functions, but no surface names a source of case law, statutes or regulations behind it.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
Nothing on any located surface addresses whether authority is checked for subsequent history. The product does not retrieve primary law: it works on the customer's own record to produce summaries, chronologies and drafted documents, citing back to the source material supplied. The question therefore does not bite on this product class and the honest value is the absence rather than a penalty. Searched the home page, the product navigation, the security page and the terms of service on 4 September 2026.
Searched the same surfaces on 18 September 2026. No citator or subsequent-history check is described. Output cites documents in the customer's case record rather than legal authority, so the question arises only for the legal research function the home page headline names, and nothing published says whether authority returned there is checked.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
No located public material describes what the product does when it cannot ground an output. The agreement acknowledges that output may not accurately reflect people, places or facts and instructs the customer to review it for accuracy, incorrectness and completeness before relying on it, which is an acknowledgment of risk and an instruction to the reader rather than a description of system behavior. No abstention path, no no-answer state and no confidence or grounding score surfaced to the user is described anywhere.
Searched the home page, the product navigation, the security page, the pricing page and the terms of service on 4 September 2026.
The review system exposes graded relevance and grounds each determination, without a documented abstention path. The product page describes traversing the record by documents' degree of relevance with an issue-specific rationale pinpointing the text behind each finding, and the white paper reports graded outputs such as highly responsive and likely responsive, with hot documents escalated separately. Nothing describes what the prompting engine does when the record does not support an answer, and no state is described in which the system reports that it cannot tell.
Fabricated Citation Record
Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?
The AI Hallucination Cases database maintained by Damien Charlotin was searched on 4 September 2026 on both the product name CaseMark and the corporate name CaseMark AI. No court order, opinion or disciplinary record naming the product was located. This records the state of the public record on that date and is not a finding about the 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 legal citations in output from Syllo or TLATech Inc. 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?
Public materials refer to professional responsibility in general terms without naming guidance. The agreement carries a No Legal Advice clause stating that CaseMark does not provide legal advice and is not engaged in the practice of law, that generated content is informational only, that no attorney-client relationship arises, and that the reader should seek advice from a qualified attorney licensed in their jurisdiction.
A separate restriction prohibits representing output as human-generated when it is not. Both engage professional responsibility as a subject. No bar association, regulator, rule of professional conduct or ethics opinion is named anywhere, and no mapping of product behavior to any published guidance exists.
Professional regulation is referred to in general terms without naming any ethics guidance. The terms say the vendor has no responsibility to monitor unlicensed practice, may act on instruction from an attorney disciplinary body, and make each user responsible for being authorized to practice (sections 5.4 and 5.8). The white paper cites the Sedona Conference TAR Case Law Primer and the Bolch Judicial Institute TAR Guidelines on the defensibility of technology-assisted review; those are court-facing standards for review methodology, not bar ethics opinions. No ABA or state bar opinion on generative AI is named on any surface read.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
Public materials claim time savings without addressing billing. The headline promise is gigabytes of discovery to a motion for summary judgment in under an hour, and the published testimonial describes a 400-page transcript submitted overnight and returned finished, saving the reviewer the work entirely. Nothing addresses what happens to a client bill when that time disappears, and no per-matter record of AI-assisted work is described as available for fee or disclosure purposes, although the platform does log all user activities and system events for security purposes.
CaseMark's own charging model is per user plus AI credits rather than per hour, which is a cost fact rather than an answer to this signal.
Time and cost savings are claimed for work that sits inside a lawyer and client fee relationship, without any treatment of billing or disclosure. The product pages claim review at twenty times the throughput of traditional methods and at a fraction of the time and cost, and the white paper reports a managed review team at more than two dollars per document alongside the system's results. One case study records that the firm educated its client on the option to use the platform and the client elected to proceed, and a partner refers to the amount the client paid for the review; that describes one firm's practice, not vendor guidance. Rationales and audit logs exist but nothing marks work as AI-assisted for fee purposes.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
The material exists behind a request and does not cover the AI layer. The security page itemizes what is available rather than gesturing at it, listing SIG Lite 2024, CAIQ questionnaires, penetration test summaries and attestation letters as available on request, which is a genuine route a firm can use. Some suppliers are named openly, being AWS and Azure for hosting and Stripe for payments. What is missing is the disclosure a client's AI clause actually asks for: no model provider is identified anywhere, no subprocessor register covering the AI layer is published, and no data processing agreement or forwardable client-facing pack was located at any access tier. Naming the cloud environments says where data sits, not whose models see it.
No subprocessor list, model provider statement or forwardable disclosure pack was located. The terms name where Content is stored, Amazon Web Services virtual servers, Amazon S3 and ElasticSearch (section 13.3), which is infrastructure and does not say whose models see client content. The SOC 2 Type II attestation is stated without a route to the report, and no DPA is published. A firm answering a client's AI clause could say where data sits but not which model providers process it.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Some elements of a record are available, short of a document-level export covering the AI. Outputs are described as citing back to the source, deposition summaries are offered in page-line format and medical chronologies as cited timelines, which covers sources retrieved. The security page publishes comprehensive audit logging of all user activities and system events, which is a trail of who did what, though it is presented as a security control rather than as a disclosable record and no export is described.
What is absent is the model dimension entirely: no model or provider is identified anywhere, so which system produced a given output could not be stated even if a record were exported, and no human verification step is documented. No disclosure guidance or template for a court was located.
Some elements of a disclosure record exist, short of a document-level export covering model, sources and human verification. Every tagging determination carries a written rationale pointing to the triggering text, audit logs are standard, the platform generates privilege logs and production load files, and exported chronologies and summaries carry direct links to the record. The white paper documents the validation method firms use, precision and elusion testing, which bears on defending a review.
Nothing states which model produced an output, and no export framed as an AI use or verification record for a court is described.
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 CaseMark if
- You need training and retention settled in the contract. CaseMark's terms state that customer data is not used to train its AI models or sold, and that it is retained for 30 days and then permanently deleted, with longer retention only for enterprise customers who ask for it.
- You want to start at a known price. CaseMark publishes Workspace Pro at $100 per user per month, including $80 of AI credits, document analysis, AI chat, more than 880 legal agent skills and 1 GB of storage, with extra usage billed at market rates.
- You run court reporting or litigation support, or want to build your own tools. CaseMark's Source product delivers transcript bundles in twelve formats from one certification pass with synchronized deposition video, and its case.dev API exposes eighteen services with public developer documentation.
Choose Syllo if
- You need a recall figure you can defend. Syllo's March 2025 white paper reports that across its last ten completed responsiveness reviews in live litigation the lowest estimated recall was 93.4 percent and the median 99.4 percent, measured by the firms' own precision and elusion testing, with seven firms named as contributors.
- You want every coding decision explained. Syllo's Agentic Issue Tagging applies unlimited issue codes across document sets running to millions and writes a rationale for each determination that points to the text behind it, with tag conflict detection in quality control and case teams validating results before production.
- You want the vendor to push subpoenas back to you. Syllo's incorporated Legal Disclosure and Subpoena Policy treats stored material as privileged work product, argues that the customer is the proper target of legal process, and commits to written notice so a customer can seek a protective order.
In summary
CaseMark
CaseMark, from CaseMark AI, Inc. in San Francisco, reads a matter's record and drafts what the matter needs, citing back to source, through a Workspace with more than 880 legal agent skills, Structured Summaries for depositions and medical chronologies, Source for court reporting and litigation support businesses, and the case.dev API. The AI Legal Index grades it in the top two bands on ten of fifteen capability axes, with an A on AI centrality. Its terms state that customer data does not train its models and is deleted after 30 days, and Workspace Pro is published at $100 per user per month. It names Nelson Mullins and Ogletree Deakins among its customers. As of 4 September 2026 the index located no named model provider, no accuracy measurement and no AI governance position.
Syllo
Syllo is a litigation platform from TLATech Inc. of New York, founded in 2019 by two litigators and two Carnegie Mellon engineers, holding the case record from complaint through trial with case management, an eDiscovery module and AI analysis. Its Agentic Issue Tagging coordinates several language models to code millions of documents with a written rationale for each call. The AI Legal Index grades it in the top two bands on ten of fifteen capability axes, with an A on operational evidence: its March 2025 white paper reports recall and precision measured by named firms on live reviews. It states SOC 2 Type II attested by Prescient Assurance. As of 18 September 2026 the index located no published price, no named model provider and no statement against training on customer content.
Questions buyers ask
CaseMark vs Syllo: which is better for litigation AI?
Neither on the totals: the AI Legal Index places both in the top two bands on ten of fifteen capability axes. Syllo publishes measured review accuracy from named firms' live matters and is built for document review at scale. CaseMark publishes contract terms on training and deletion, a price, and a broader set of drafting and summary products. A team defending a review protocol has more to read from Syllo; a firm reviewing a vendor contract first has more from CaseMark.
How accurate is Syllo's document review?
Syllo's white paper of 21 March 2025 reports that across its last ten completed responsiveness reviews in live litigation, estimated recall ranged upward from 93.4 percent with a median of 99.4 percent, and median estimated precision was 85.9 percent, using the firms' own precision and elusion testing. Named examples include Ballard Spahr at 99.4 percent recall and 95.56 percent precision. The figures cover classification, not generated answers, and the matters are not named. 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.
Does CaseMark train AI on client files?
No, by contract. CaseMark's terms of service license customer data only to provide the service and state that it does not use customer data to train its AI models or sell it to third parties, and its security page repeats the position. Data is kept for 30 days and then permanently deleted, with longer retention arranged for enterprise customers. Syllo's terms license content to provide or improve its service without naming training. 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.
How much do CaseMark and Syllo cost?
CaseMark publishes Workspace Pro at $100 per user per month, including $80 of AI credits and access to more than 880 legal agent skills, with additional usage billed at market rates and its Teams and Enterprise tiers quoted. Syllo publishes no price or unit: in house teams buy on an enterprise basis and law firms through firm wide partnerships, both through a sales conversation. 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 CaseMark and Syllo both leave unpublished?
Whose models run the work. CaseMark names none, and Syllo describes an ensemble of proprietary and foundation models without naming a provider. Neither publishes a vendor indemnity or liability beyond $100 in its published terms, a security incident notice commitment, or a subprocessor list for the AI. Neither says what its AI does when the record cannot support an answer, and neither addresses how saved review time reaches 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. Syllo's accuracy figures come from its own white paper, the matters are not named, and its footnotes disclose that one practitioner contributor is a product advisor to the vendor; the figures cover document classification, not the accuracy of generated answers or summaries. Syllo's published terms bind end users and give way to a customer licensing agreement that is not published, and its product privacy and data security policies could not be read. CaseMark's privacy policy dates from July 2024 and describes products it no longer sells. CaseMark was verified on 4 September 2026 and Syllo 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.