Legartis vs ThoughtRiver: how they compare in 2026

Legartis sells the product as a Legal AI Workspace built on an agentic framework. Its named components are the Review Agent, the Legal Agent and the Playbook Creator Agent, and every capability a customer pays for runs through these agents. Since agentic workflows launched on 23 September 2026, the Legal Agent takes on complete legal processes, from incoming requests and contract review to negotiation, approvals, reporting and signature. The pricing page also lists e signature via DocuSign and a contract repository with status and versions. Workflows can be created by voice or in writing and reused across the workspace. Procurement and sales teams work in their own Focus Spaces, under rules and templates legal has approved.

Machine learning is the mechanism the customer pays for. A model reads the contract, answers the Lexible question set, scores the answers against a playbook and produces the risk rating, issues list and suggested redlines. Remove the models and there is no review, no issues list and no redlines, only a document store. The vendor has sold this as an AI product since 2016 and describes the current engine as a proprietary generative model on several large language models.

Legal experts define the playbook requirements against which Legartis evaluates contracts. Each requirement has a reliability indicator derived from test examples, and users can correct classifications to tune future reviews. The interface traces review results to a requirement and contract passage; legal questions carry citations to supporting sources. The quality framework names inaccurate instructions, missing context, interpretation differences and invented information as distinct failure modes. Customer dashboards expose performance within their own workflows, but a reproducible public test set and current accuracy results are not provided. The widely stated reduction in review time measures speed rather than legal accuracy.

Lexible uses lawyer labeled concepts to identify contract provisions and assess them against a playbook. Its accuracy explanation reports 97% F1, 96% precision and 97% recall, with testing three times weekly against 750,000 verified data points. Legal specialists verify training examples, and sampling covers different contract types and domains. These figures concern classification and do not measure every generated redline or answer. The underlying dataset, current model version and detailed failure results are not available for independent reproduction. The terms acknowledge that training can reduce accuracy, especially with incorrect labels, and that mistakes remain possible.

The system works against legally validated requirements and company playbooks, and legal experts rather than the model define what counts as correct. Legartis names four review tools. AI Quality Scores for each requirement appear in a dashboard, and in test sets the user reviews and corrects how the system understood a requirement. Detection of clauses outside expected patterns and a full audit trail complete the set. It states twice that judgment returns to a person. A human is the final checkpoint before any decision is made on an AI classification, and accountability stays with the organization, not the AI, when agents draft, review or flag clauses. The path for a misclassification is documented end to end, from detection through correction in the test set to the system applying that correction to similar cases. Since the agentic workflows launched on 23 September 2026, legal teams set the playbooks, templates, responsibilities and decision boundaries the Legal Agent follows. When a deviation exceeds a boundary, such as a liability clause outside the agreed position, the agent refers it to the responsible person for a decision. Legartis publishes no default boundary, and each customer sets its own.

The platform reads an agreement, identifies issues against a configured playbook and presents suggested redlines in Word. Users can share issues and changes with colleagues, while Lexible Assistant supports questions and remediation. The customer sets its risk policies and remains responsible for decisions based on the resulting analysis, including automated decisions. Terms expressly warn that output is not legal advice and may contain mistakes. These features provide a review surface and organizational standards. The system does not specify when it can approve an agreement without review or how it must escalate and correct an incorrect conclusion.

Two customers, dormakaba and TUEV NORD, have dedicated case study pages. Seven more appear as logos (Uniklinik, Emmi, Fiege, Oechsler, Framatome, Zuercher Kantonalbank and Arabelle Solutions). Legartis cites a DPA first review falling from 45 to 60 minutes to under 10, and contract review up to 85 percent faster. It also cites playbook creation effort down up to 98 percent and costs down more than 90 percent on portfolio risk analysis. Testimonials carry personal names, including Kim Weiler, Dr Marc Hansmann, Cedric Ruepp, Gordian Berger and Juerg Sommer. The testimonials do not give these individuals' organizations, so the figures they quote are not tied to a named customer.

A named customer gives figures, without a date or method. Joe Stephenson, partner and head of technology at Shoosmiths, is quoted on the demo page saying the platform reviewed complex supply agreements in under three minutes at above ninety percent accuracy, where lawyers took four hours at 86 percent. A second, unnamed customer calls ThoughtRiver a clear winner for reviewing its appointments. No date, sample size or method accompanies the Shoosmiths figures.

The home page and pricing page state no data sharing with third parties, GDPR compliance, ISO 27001 certification and hosting in Switzerland and Europe. The pricing page uses the narrower phrase no uncontrolled data sharing. The only published data document is a privacy policy last updated 6 November 2019. It addresses website and account data and does not address uploaded contracts, prompts or outputs. On training, the only published statements are on the pricing pages, which offer exclusion of the customer's data from model training on the paid Team to Enterprise plan only. Legartis does not address segregation between customers or between matters, states no retention or deletion position for contract content, and identifies no model provider. Nothing on its site mentions privilege, professional secrecy or work product.

The agreement treats customer contract data as confidential and restricts disclosure by staff and subcontractors, with confidentiality continuing after termination. Customer accounts have separate encryption keys and segregated contract data. Anonymization removes sensitive information before model processing. The same agreement grants continuing use of training data and anonymized contract text, and the DPA includes machine learning training among processing purposes. This is permission to reuse specified material, rather than a promise never to train. The terms do not cover internal matter walls or privilege handling directly, and ThoughtRiver gives no complete account of the underlying language model providers.

Asked in its FAQ who is liable if the AI makes a mistake, Legartis answers that when a lawyer sends a signed document it makes no difference how it was created. The lawyer remains liable, the AI is an assistive tool, and legal responsibility stays with the person who approves the result. A companion answer says accountability stays with the organization rather than the AI. Both positions sit in FAQ prose rather than in an agreement. No jurisdiction limit is stated, although the product is sold across every European language and legal system. Legartis also invites people who are not lawyers to work in the product. A dedicated FAQ confirms that procurement and sales teams review contracts on their own in line with legal standards. No statement on scope of use for those users is published.

The customer terms expressly say that output is not legal advice and must not be relied on as advice. They also warn that predictive technology makes mistakes and that incorrect training can worsen accuracy. Customers remain responsible for results, conclusions and automated or manual decisions. The license allows both internal business work and integrated managed services for legal clients, so law firms and other business users share that responsibility. Product review surfaces support human checking, but jurisdiction specific practice limits and an explicit treatment of lawyers' competence and supervision duties are not set out.

The AI Quality System comes in three named layers, each with a described mechanism. Playbooks and Legal Best Practices with Agent Memory form the knowledge base, AI Quality Scores come with feedback loops and targeted user tuning, and clauses outside expected patterns are detected. Three published articles cover auditable AI, explainable AI and governing agentic legal AI. Legartis distinguishes explainable AI, which shows how a decision was reached, from auditable AI, which checks whether output meets a defined standard. Nobody inside Legartis is named as accountable for AI governance, no testing regime before release is described, and no testing results are published. Nothing is published about uneven output across matter types or populations, although bias appears in a section heading. ISO 27001 is a security management standard and does not cover AI governance.

ThoughtRiver describes lawyer verification of training examples, annotation that accounts for contract context and sampling across contract types and legal domains. Its evaluation program runs three times each week against 750,000 verified data points and reports separate precision, recall and F1 measures. Customers can set playbooks for their own risks and review proposed changes before further use. These are concrete training and evaluation practices. The description does not identify an accountable AI governance owner, a mandatory release approval threshold or findings about uneven output across jurisdictions or populations. Its regulatory position is separate from those operating tests.

Data handling is governed by a privacy policy last updated 6 November 2019. It covers website and account data and does not address what happens to uploaded contracts, prompts or generated output after processing. Its retention section speaks generally, saying aggregated, anonymized or pseudonymized information may be kept indefinitely, and states no retention period for contract content. No deletion position or incident and breach practice is published. The policy names Turicode, VSHN and LinuxFabrik as third party processors for Application users. All three are Swiss infrastructure and operations suppliers, and no AI model provider is among them. Access control appears only as paid features in the pricing matrix, with role management, document controls and audit logs on the paid Team to Enterprise plan. Personal data is stated to be stored on servers in Zurich.

The DPA names infrastructure, authentication, document conversion, email and support providers, and requires equivalent safeguards for subprocessors. Personal data breaches must be reported without undue delay, with cooperation in mitigation. Customer data is encrypted in transit and at rest, separated by account and protected by a dedicated encryption key. Authentication supports single sign on and multifactor controls. At termination, personal data must be deleted or returned, with backup and legal exceptions. ThoughtRiver gives no in use retention period for prompts, no backup deletion deadline and no explicit list of the language models processing contract text.

The legal section of the site footer offers two documents, a legal notice and a privacy policy, and neither is a contract for the service. Legartis publishes no customer agreement. The legal notice is a website disclaimer. It disclaims the accuracy of information on the website, excludes liability for losses from using that information, and puts reliance at the user's own risk. It governs the website rather than the Application. No indemnity, liability cap, warranty on the service or its output, or insurance position is published. On who bears the loss, the only statement is an FAQ answer on the AI Quality System page. It says the lawyer who signs remains liable and accountability stays with the organization. That answer is marketing prose that puts risk on the customer, and it is not a contractual term binding Legartis.

ThoughtRiver indemnifies customers against claims that permitted use of the platform infringes third party copyright, subject to control of the defense and exclusions for unapproved combinations. It can replace or modify the platform to address infringement. General liability is limited to the order term's fees, or £1,000 during a trial, with the copyright indemnity excluded from that cap. Death or injury caused by negligence, fraud and other legally unavoidable liabilities also remain outside the limitation. Output errors carry no warranty or accepted liability, so the software indemnity does not amount to a guarantee of correct legal analysis.

The Word add in is the main integration, and customers on Microsoft Office 2019 or Office 365 can install the Legartis Word app. An open REST API is described as letting Legartis run as a standalone web app alongside any contract management solution. The pricing matrix lists API and MCP integrations on the paid Team to Enterprise plan. Legartis has no integrations page or developer documentation, and does not describe what moves between systems or in which direction. The pricing page now lists e signature via DocuSign, and the Team to Enterprise plan adds SSO.

Agreements enter through email, file upload or a Microsoft Word add in, with issues and redlines available for team collaboration. Microsoft 365 supports work in Word, Outlook and Teams, while Power BI receives contract data for analysis. HighQ and iManage are named connections for legal projects and document workflows. The platform also exposes the same public APIs used by its web application and follows the OpenAPI specification. These connections have identifiable purposes, but per connector field mappings, authentication setup, permission inheritance and error recovery steps are not described in the integration overview.

Hosting is in Switzerland and Europe, according to the product and pricing pages, and the arrangement is the same on every tier. The privacy policy says personal data is stored on servers in Zurich. Processing location is not addressed separately from storage. No model provider is named, which leaves the inference location open. Legartis does not say whether the platform is single tenant or multitenant, or how customers are separated. The only related detail is document level permissions, offered as an Enterprise feature. No deployment option varies the hosting arrangement.

Region and processing location are stated, and tenancy is described as segregation. The security page says data resides in Azure data centers managed and secured by ThoughtRiver in a single specified region, and the subprocessor table places Azure hosting in the United Kingdom. Each customer's data is held separately under its own key. ThoughtRiver does not say whether a customer may choose a region other than the UK or whether infrastructure is shared, and it does not address where the language models run inference.

An ISO 27001 certification is claimed. The claim appears in the home page trust block, the security section of the AI Quality System page, the pricing matrix and a badge in the site footer. No certifying body, certificate number, scope statement or statement of applicability, or issue or expiry date is published. Legartis has no trust center or portal, and offers no route to obtain a report at any access tier. With no certifying body named, the claim cannot be matched to a register without contacting Legartis. No penetration testing is mentioned.

ISO 27001 certification is stated and identified by number, without a route to the report. The security page gives certificate number 21188-ISMS-001 alongside encryption at rest and in transit. No certifying body, coverage period or route to the certificate or any SOC report is published, and no trust center is published.

Legartis describes its own agentic AI framework and does not identify the models underneath it. Its marketing distinguishes that framework from generic language model output and contrasts Legartis with ChatGPT, which implies third party models are in use without naming one. No model or model provider is named, and no notice of change is committed. The third party processors published for Application users, Turicode, VSHN and LinuxFabrik, are Swiss infrastructure and operations suppliers, not model providers. Legartis gives no location for inference. Hosting in Switzerland and Europe describes where the platform sits. The pricing matrix offers exclusion from model training on the paid Team to Enterprise plan, without explaining the default treatment of Free accounts.

Lexible combines a proprietary legal model with general purpose language models for contract interpretation and generated responses. ThoughtRiver describes supervised learning from expert labeled concepts, legal validation of examples and repeated evaluation. An anonymization layer protects sensitive text before analysis, and a private model option is offered. The hosting environment is Microsoft Azure. ThoughtRiver does not name its external foundation model providers, their versions, its routing decisions or where each model processes data. The subprocessor register covers infrastructure and support functions but does not resolve the generative model chain or commit to notice of each model substitution.

Legartis has a Free plan and a paid Team to Enterprise plan priced on request. Free provides one seat, two NDA reviews a month, basic agent access and German, English and French. The paid plan expands to all contract types and European languages, with custom playbooks, workflows, reporting, single sign on, role management and audit logs. Contract volume and team requirements determine the commercial discussion, and no paid rate or implementation amount accompanies the current plans. Training exclusion is a paid entitlement. The older Professional and Team rates no longer appear in either the current English or German plan matrix.

Professional starts at £15,000 a year in the United Kingdom for teams reviewing fewer than 50 contracts a month; Enterprise starts at £30,000 for higher volume. Outside the United Kingdom, the page displays starting rates of $20,000 and $35,000 respectively. Both list unlimited users, a private database instance, Word access, Lexible Assistant, single sign on and dedicated success and professional services. A trial lasts 28 days. The agreement bills order fees in advance and permits separate professional services charges. The published starting amounts therefore do not settle every implementation charge, volume adjustment or negotiated order term.

The buyer Legartis describes is the legal department and the procurement and sales teams around it, each with a dedicated page, plus a construction industry page. Law firms are neither claimed nor excluded. Best Practice Playbooks are published for NDA, DPA, SaaS agreement, purchase agreement and commercial lease agreement. Paid tiers cover all contract types, and the free tier covers NDA only. Twelve industries are listed. Language coverage is set per tier, with German, English and French on Free and all European languages on paid plans. Coverage is claimed for companies of any size and industry. Government use is not addressed, and nothing says where the product stops.

Corporate legal, procurement and business teams use ThoughtRiver to review incoming agreements, and the license also allows managed services delivered to legal clients. Configurable playbooks apply the customer's negotiation standards to recurring contracts. Professional and Enterprise packages distinguish teams below and above 50 reviews per month, with unlimited users in both. Named examples include complex supply agreements and professional appointments. The current site gives different totals for its trained concept library on different pages, so that total does not define a reliable coverage boundary. ThoughtRiver publishes no complete jurisdiction matrix and no list of unsupported contract types.