S

Syllo

The capabilities the vendor leads with run on models, and the conventional layer underneath would still work without them. Agentic Issue Tagging coordinates several 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 calls itself a Litigation AI platform, founded to build the data and workflow infrastructure into which language models are woven.

Beneath that sits an eDiscovery module that stands on its own, with native and load file processing, Boolean and metadata search, bulk redaction and privilege log automation. Case management covers chronologies, exhibits, binders and transcript designations. Strip the models out and a litigation support team still has a review and case management platform.

Accuracy is measured and published for document classification, and generated output is grounded but not measured. The white paper of 21 March 2025 reports that across the last ten completed responsiveness reviews in live litigation, estimated recall was lowest at 93.4 percent, averaged 97.8 and had a median of 99.4. Median estimated precision was 85.9 percent and the average 79.7, with figures per matter, dataset sizes, and the firms' own precision and elusion testing as the method.

Every tagging determination carries a rationale for the issue that pinpoints the text behind it, and answers from the prompting engine carry hyperlinked citations to the case record. No figures are published for the accuracy of generated answers, summaries or digests. The failure modes the paper names belong to competing approaches, such as prompt overload, context windows and hallucination in linear review, not to this system.

The benchmark is eighteen months old at this date. Grounding is to the customer's own record, not primary law, so checking citation status does not arise.

Review surfaces and a validation route are real, and no threshold or use constraint is published. The white paper describes the review system as coordinating several language models that decide how to conduct the review within guidelines users set. The product page says case teams validate the outcomes of agentic issue coding to ensure defensibility for outgoing productions, and tag conflict detection is built into quality control.

The white paper's case studies show that validation as precision testing on tagged sets and elusion testing on the null set. Customer and Case Administrator roles control who can act on a case. Nothing says at what point the system stops or escalates, what any mode's output may be used for, or 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, in early access, and no oversight structure is published for it.

Named firms, a date, figures for what changed and a method that can be assessed are all published. The 21 March 2025 white paper lists practitioner contributors from seven firms, among them Ballard Spahr, Royer Cooper Cohen Braunfeld and Quinn Emanuel, and reports results per matter measured by the firms' own teams. Ballard Spahr's trial team tested a review of more than 100,000 documents against more than 25 issue codes and found estimated precision of 95.56 percent and recall of 99.4 percent, after about three hours of setup.

Royer Cooper Cohen Braunfeld compared approaches on a random sample of just under 16,000 documents from more than 150,000. Its managed review team using continuous active learning reached estimated recall below 67 percent at over two dollars a document, against the system's recall of 93.44 percent and precision of 69.81 percent. The paper is published by the vendor, the matters are not named, and the contributors wrote in their individual capacities.

Its footnotes disclose that one technical contributor is a research advisor and one practitioner a product advisor to the vendor.

Privilege and work product are treated expressly in writing, and training is left open. 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. The terms add that Content is not shared with third parties except service providers, legal process and security monitoring.

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). Segregation is documented by case. The license in section 8.4.1 covers improving the service, and no commitment against model training is published. The position of any outside model provider is not stated.

Professional responsibility is placed on the user and the audience is widened, while the marketing speaks of legal analysis and strategy. Section 5.4 of the terms says 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. It adds 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 being authorized to practice where they use the service for the practice of law. 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 the US account rule.

A measurement practice is published, and no governance framework, owner or testing before release surrounds it. The white paper describes validating outputs in live matters and reports aggregated results across ten reviews, and it 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 that early access features are not subject to the same security measures and auditing. No responsible AI policy or principles page is published, and no person or function is named as accountable for model behavior. Nothing describes what is tested before a model or feature is released, and uneven output across matter types or parties is not addressed.

Controls are published across most of this ground, and incident practice and a processor list for the AI are missing. The Work with Syllo security section states encryption in transit and at rest in line with NIST SP 800-57, automatic backups, continuous vulnerability scanning and penetration testing, and multifactor authentication. The terms require credentials for one person only and user notice of a compromise within one business day (section 2.1), and aim for 30 days' notice before a feature is discontinued so users can export (16.3).

No commitment to notify customers of a security incident is published, and 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, returned an empty page to automated retrieval and could not be read on this date.

Liability is handled through a disclaimer, a nominal cap and an indemnity running to the vendor. Section 15 provides the service as is and as available, disclaims any warranty that content provided through it is accurate, reliable or correct, and excludes all categories of damages. It 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, from each user to 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 give way to the customer's Licensing Agreement where the two conflict, so negotiated terms may differ, but that agreement is not published.

The product stands alone by design. On its own platform page, the vendor argues that API integration and acquisition have both failed to fix fragmented litigation tooling, and offers a single environment holding the whole record instead. What it supports is interchange: 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 formats make moving a matter onto the platform routine, but no connection to an external system is named. No document management system, Microsoft Word or Outlook plugin, matter management, electronic billing or court filing integration is published, and no API or developer documentation is published.

Residency is offered and the hosting provider is named, and the deployment options are not specified. The security section says multiple secure deployment options are available, data storage is regionalized to customer needs, and for firm engagements the vendor will work with a firm's security team to customize deployment. The terms name the infrastructure: Content is stored on Amazon Web Services virtual servers and Amazon S3 storage the vendor controls, with ElasticSearch for indexing (section 13.3).

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). The options themselves are not named, so whether dedicated, private cloud or on premises arrangements exist cannot be established. No regions are listed, and nothing separates where model inference runs from where data is stored.

A SOC 2 Type II attestation is named with its scope and auditor, and no route to the report exists. The compliance section states SOC 2 Type II certification in Security, Confidentiality, Availability and Privacy, compliance with the AICPA SOC standards for service organizations, and third party validation by Prescient Assurance. No trust center is published on the site or by search on this date, no report period or issue date is given, and no request route of any kind is described.

The architecture is described in detail, and no provider or model is named. Product pages say the platform runs on a combination of proprietary Syllo models and foundational language models. The white paper describes an ensemble of language models of varying size in distinct roles, such as strategizing, quality control and resolving disagreement between models, with telemetry routing harder material to more capable models and simpler material to cheaper ones.

No provider, model family or version is named on any surface read, nothing says where inference runs as distinct from the storage named in the terms, and nothing addresses what a model provider may retain. On change, the terms reserve the right to modify any feature at any time with or without notice (section 13.7).

No price, unit of charge, tier or list of inclusions is published. The pricing section of the Work with Syllo page names two customer segments and routes both to a sales conversation: corporate legal teams are offered the platform on an enterprise basis, and law firms a firmwide enterprise partnership. Firmwide describes the scope of a license, not a unit anyone could price. The home page, platform and tagging pages, the terms and the white paper were also searched, and the only cost figures on the site describe competitors.

Segments and matter types are described with evidence, and the boundaries are not. The vendor names AmLaw 100 firms, litigation boutiques and corporate legal teams managing complex litigation portfolios, with a dedicated page for the last. The white paper shows the breadth with named firms from AmLaw defense practices to a plaintiffs' employment firm. It lists thirteen kinds of matter the system has been used on, among them antitrust, patent and mass tort, across datasets from thousands to more than two million documents.

The scope is litigation throughout. Nothing addresses government or court use or names matter or data types it handles poorly, and the one stated limit, the US account rule, is framed as data protection rather than coverage.