Legartis vs ThoughtRiver: how they compare in 2026
Legartis and ThoughtRiver are both European AI contract review tools for in house legal teams, scoring contracts against a customer's playbook and proposing redlines in Word. Legartis sits in the top two bands on eleven of fifteen axes and ThoughtRiver on eight of fifteen. Both give customer contracts a role in model training that a buyer should read first. ThoughtRiver's data processing addendum lists machine learning training and feature enhancement among its purposes, with no opt out located. Legartis offers an opt out of model training only on its Team and Enterprise plans. Legartis leads on control and price. It shows a quality score for each playbook requirement, lets users correct the system in test sets, and keeps a human as the final checkpoint. It publishes its prices, from a free tier to CHF 300 per user a month, with Enterprise quoted. ThoughtRiver's counterweight is data handling and measurement. It publishes a subprocessor list, per customer encryption keys and breach notice, and an accuracy of 97.3 percent across its question set.
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.
Every capability the buyer pays for is an agent. The product is sold as a Legal AI Workspace built on an agentic framework, and the named components are the Review Agent, the Legal Agent and the Playbook Creator Agent. The conventional software around them, contract repository, version control, e-signature via DocuSign, approval workflows and negotiation, is marked Coming soon across the pricing matrix, so what a customer can buy today is the model layer. Remove the models and there is no document management or workflow system left underneath. Checked 4 September 2026.
The machine learning is the mechanism the buyer pays for. The product is contract pre-screening: 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. Home page, demo page and platform description read 6 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, documented, and unusually candid about failure. The AI Quality System page describes the mechanism directly: the AI works against legally validated requirements rather than free prompts, what counts as correct is defined by legal experts rather than the algorithm, and every requirement in a playbook carries its own quality score derived from test sets. Unusual-clause detection is described as catching clauses that fall outside expected patterns even where the playbook does not define them. The FAQ names four error types the vendor says occur most often, and identifies hallucination among them as particularly critical because it occurs unpredictably, which is a franker statement than most records on this axis carry. What is missing is a published accuracy figure an outsider can test: the quality score is a per-customer, in-product instrument rather than a measured result, the test sets are named as the mechanism but never described, and the 85 per cent figure quoted throughout is a speed claim rather than an accuracy one. Grounding to primary authority does not bite on a contract review product and is not counted either way.
A measured accuracy figure is published, short of a described test set and named failure modes. The demo page states 97.3 per cent accuracy across 2,500 pre-trained questions and an average review time of 8.2 minutes, and a Shoosmiths partner is quoted that the platform reviewed complex supply agreements in under three minutes at above ninety per cent accuracy against qualified lawyers at 86 per cent over four hours; no test set, method, date or failure-mode statement accompanies either figure on the surfaces read. An Accuracy page exists in the navigation and was not opened, and would lift this to A if it describes the test set. The primary-authority limbs do not apply to a contract reader. Demo page and home page read 6 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.
The control structure is published rather than asserted. What constrains the system is stated: it works against legally validated requirements and company playbooks, and what counts as correct is defined by legal experts rather than the model. The review surfaces are named and specific: an AI Quality Score visible per requirement in a dashboard, test sets in which the user reviews and corrects how the system understood a requirement, unusual-clause detection, and a full audit trail. The route back to human judgement is stated twice and in terms: a human is kept as the final checkpoint before any decision is made on an AI classification, and accountability stays with the organisation rather than the AI when agents draft, review or flag clauses. The misclassification path is documented end to end, from detection through correction in the test set to the system applying that correction to comparable cases. No numeric threshold at which an agent stops and hands back is published; the per-requirement quality score is the surface that does that work in this product's own idiom, and it is named here rather than treated as equivalent.
The modes and the review surface are published, short of the full control structure. The product runs the review and produces a Digital Issues List with a risk rating for every contract, routes and prioritises matters on that basis, and presents suggested redlines for a lawyer to accept inside Word; the playbook a customer configures sets what is flagged, and the vendor describes the product's purpose as automating routing and prioritisation decisions in busy legal functions. What is not published is the threshold at which a contract is treated as clear without human review, or a stated route back after a wrong answer beyond the reviewer's own correction. Home page, G2 feature description and platform coverage read 6 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 verified measurement. Two customers are named with dedicated case study pages, dormakaba and TUEV NORD, and seven more appear as logos: Uniklinik, Emmi, Fiege, Oechsler, Framatome, Zuercher Kantonalbank and Arabelle Solutions. Figures are published and specific, including a DPA first review falling from 45 to 60 minutes to under 10, contract review up to 85 per cent faster, playbook creation effort down up to 98 per cent, and cost reductions above 90 per cent on portfolio-wide risk analysis. Testimonials carry personal names, including Kim Weiler, Dr Marc Hansmann, Cedric Ruepp, Gordian Berger and Juerg Sommer. Two things hold this at B rather than A. The named individuals carry no organisation on the pages read, so a reader cannot tie a quoted figure to a named customer. And the two case studies were not opened in this pass, so they are credited for existing and named as unread rather than counted for their contents, which would be crediting a document by its title.
A named customer with figures, short of a date and method. Joe Stephenson, partner and head of technology at Shoosmiths, is quoted on the vendor's demo page that the platform reviewed complex supply agreements in under three minutes at above ninety per cent accuracy where qualified lawyers took four hours at 86 per cent; a second customer is quoted, unnamed, that ThoughtRiver was a clear winner for reviewing its appointments. No date, sample size or method accompanies the Shoosmiths figures. Demo page and home page read 6 September 2026.
Privilege and Confidentiality Posture
How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.
Confidentiality is asserted in general terms across marketing surfaces and is not supported by any readable commitment about contract content. The home page and pricing page state no data sharing with third parties, GDPR compliance, ISO 27001 certification and hosting in Switzerland and Europe, and the pricing page uses the narrower phrase no uncontrolled data sharing. Behind those claims, the only published data document is a privacy policy last updated 6 November 2019, which addresses website and account data and never addresses uploaded contracts, prompts or outputs at all. Nothing published addresses training on contract content except the pricing matrix line offering opt-out of model training, which is available only on the Team and Enterprise plans. Nothing addresses segregation between customers or between matters, nothing states a retention or deletion position for contract content, no model provider is identified, and privilege, professional secrecy and work product are not mentioned on any surface. Searched the home page, AI Quality System page, pricing page, privacy policy and legal notice on 4 September 2026.
Segregation and encryption are substantive, and the training position is adverse, which is the finding. The security page states that contract documents and derived data are segregated from other customers and encrypted with a unique key per account, and the DPA deletes or returns personal data on termination. Against that, the DPA annex lists machine-learning training and feature enhancement among the purposes for which uploaded contract data is processed, with no opt-out located, and the published sub-processor table names no language-model provider while the demo page states the engine is built on several industry-leading LLMs, so who receives contract text for inference is not stated. Nothing addresses privilege or work product. DPA, security page and demo page read 6 September 2026; the terms of use were not opened.
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 tooling versus advice, and it sits in FAQ prose rather than in any agreement. Asked who is liable if the AI makes a mistake, the vendor answers that when a lawyer sends a signed document it makes no difference how it was created, that they remain liable, that the AI is an assistive tool, and that legal responsibility stays with the person who approves the result. A companion answer states that accountability stays with the organisation rather than the AI. That is a genuine advice-line position. Two limbs are open. No jurisdiction limit is stated anywhere despite the product being sold across every European language and legal system. And the product actively invites non-lawyers to work in it, with procurement and sales teams reviewing contracts independently in line with legal standards and a dedicated FAQ confirming it, which is the professional-responsibility surface that most needs a published scope-of-use statement and does not have one.
No advice line or supervision statement was located on the surfaces read. The product is marketed as delivering detailed advice that guides users through remediation and as serving in-house counsel, business and procurement teams, without a located statement that outputs are not legal advice or that a qualified lawyer should supervise them; the terms of use exist in the footer and were not opened, and are the rebuttal route. Home page and platform descriptions read 6 September 2026.
AI Governance and Bias Disclosure
Published governance over model behavior: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
A published governance framework with real substance. The AI Quality System is set out as three named layers with a described mechanism for each: playbooks and Legal Best Practices with Agent Memory as the knowledge base, AI Quality Scores with feedback loops and targeted user tuning, and detection of clauses falling outside expected patterns. It is supported by three published deep-dive articles on auditable AI, explainable AI, and governing agentic legal AI, and the vendor draws a distinction most do not, that explainable AI shows how a decision was reached while auditable AI verifies whether the output meets a defined standard. What is absent is what separates this band from the top: nobody inside Legartis is named as accountable for AI governance, no pre-release testing regime is described, no testing results are published, and nothing at all is published about uneven output across matter types or populations despite bias appearing in a section heading. The ISO 27001 certification is a security management standard and does not answer this axis.
An AI-regulation statement exists and was not read, and no framework, testing regime or accountable owner was located on the surfaces read. The footer links a document titled the EU AI Act and how it relates to ThoughtRiver, which was not opened and is the rebuttal route; the demo page states continuous training by in-house lawyers since 2016, which is a development practice rather than a governance disclosure. No ISO 42001 or equivalent, pre-release testing description or statement about uneven output across contract types was located. Footer, demo page and home page read 6 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.
The governing document is a privacy policy last updated 6 November 2019, on a product now sold as agentic AI, and it covers website and account data without ever addressing what happens to uploaded contracts, prompts or generated output after processing. That is this band in its own words. No retention period is stated for contract content; the retention section speaks generally and states that aggregated, anonymised or pseudonymised information may be retained indefinitely. No deletion position is published. No incident or breach practice is published. Third-party data processors are named for Application users as Turicode, VSHN and LinuxFabrik, all Swiss infrastructure and operations suppliers, with no AI model provider among them. Access control appears only as paid features in the pricing matrix, with advanced role management, document-level permissions and audit logs reserved to Enterprise. Personal data is stated to be stored on servers in Zurich. Searched the home page, AI Quality System page, pricing page, privacy policy and legal notice on 4 September 2026.
Substantive published policy covering most of the ground, with one gap. Retention and deletion: the DPA deletes or returns personal data on termination unless backed up or legally required, with no period stated for backups. Access control: Auth0 authentication with single sign-on and multi-factor, TLS 1.2 in transit, encryption at rest under a per-customer key, a web application firewall. Sub-processors: a published table of five with location, data categories and operations and the transfer mechanism for the one outside the UK and EU. Incident practice: the DPA commits to breach notification without undue delay. The gap is that the sub-processor table names no language-model provider while the product is stated to run on several, and the security page states no external processors are used for customer data. DPA and security page read 6 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.
No customer agreement is published at all. The footer's legal section offers exactly two documents, a legal notice and a privacy policy, and neither is a contract for the service. The legal notice is a website disclaimer: it disclaims the accuracy of information on the website, excludes liability for losses caused by use of that information, and states that reliance is at the reader's sole risk. It governs the website rather than the Application, which is the distinction drawn where a website terms of use was held not to grade the platform. No indemnity, no liability cap, no warranty on the service or its output, and no insurance position was located. The only statement located about who bears the loss is an FAQ answer on the AI Quality System page saying the lawyer who signs remains liable and that accountability stays with the organisation, which is marketing prose allocating risk to the customer rather than a term a buyer can hold the vendor to. A buyer cannot read the allocation of loss before signing because there is nothing published to read. Searched the home page, AI Quality System page, pricing page, privacy policy, legal notice and site footer on 4 September 2026.
No liability position was located on the surfaces read. The DPA addresses processor obligations and audit rights without an indemnity, cap, warranty or insurance position, and the security page describes controls; the terms of use, linked in the footer, were not opened on 6 September 2026 and are the rebuttal route that would replace this grade on a read. This records what is locatable on the date and not a finding that no position exists. DPA and security page read 6 September 2026.
Practice Systems Integration Depth
How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.
Real integrations exist and are named, short of documented depth. The Word add-in is the substantive one and its prerequisite is stated: if the customer already runs Microsoft Office 2019 or Office 365, the Legartis Word app can be installed. An open REST API is published as allowing Legartis to run as a standalone web app combined with any contract management solution, and the pricing matrix lists API and MCP integrations on all paid tiers, with custom integrations reserved to Enterprise. What is missing is depth: no integrations page, no developer documentation and no description of what actually moves between systems or in which direction. Three further connections are listed but marked Coming soon, DocuSign e-signature, sign-in with Google and Microsoft, and approval workflows, and are not credited, since a capability a customer cannot use today is intent rather than evidence.
Integrations are referred to without documentation an implementer could use on the surfaces read. The vendor describes remediation inside Microsoft Word and an Integrations and Security page exists in the navigation; iManage, HighQ, Outlook and Power BI are named by a third-party directory and are not credited. Nothing read describes what syncs or in which direction. Home page and navigation read 6 September 2026; the integrations page is the rebuttal route.
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.
Residency is stated clearly and repeatedly and is the same on every tier: hosting in Switzerland and Europe on the product and pricing pages, and personal data stored on servers located in Zurich per the privacy policy. That publishes the region limb and clears the band below, where neither tenancy nor region is stated. What is not addressed is processing location as distinct from storage location, which matters more here than usual because no model provider is named anywhere, so where inference actually runs is unknown to a buyer. Tenancy is not addressed either: nothing states whether the platform is single or multi-tenant or how customers are separated, and the only related published detail is document-level permissions offered as an Enterprise feature. No deployment options or tiers vary the hosting arrangement.
Region and processing location are stated and the tenancy model is described as segregation. The security page states that data resides in Azure data centres managed and secured by ThoughtRiver in a single specified region, and the sub-processor table places Azure hosting in the United Kingdom; contract documents and derived data are segregated per customer under a unique key. Whether a customer may elect a region other than the UK, and whether infrastructure is shared, are not stated, and the inference location for the language models is not addressed. Security page and DPA read 6 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.
ISO 27001 is stated consistently rather than decoratively: it appears in the trust block on the home page, in the security section of the AI Quality System page, as a line item available on all four tiers in the pricing feature matrix, and as a certification badge in the site footer. That is more than an unsupported badge, which is what separates this from the band below. It stops short of the top on every accessible-evidence limb. No certifying body is named, no certificate number is published, no scope statement or statement of applicability is offered, no issue or expiry date appears, there is no trust centre or portal of any kind, and no route to obtain a report exists at any access tier. Applying the third-party verifiability test, a buyer cannot check the claim against the certifying body's register without contacting Legartis, so the loop does not close. No penetration testing is mentioned. Checked 4 September 2026.
Certification is real and identified by number, short of a report route. The security page states ISO 27001 certification under certificate number 21188-ISMS-001, with encryption at rest and in transit, a web application firewall, per-customer keys and Auth0 multi-factor authentication. No certifying body, coverage period or route to the certificate or any SOC report is published, and no trust centre was located. Security page read 6 September 2026.
Model Supply Chain Disclosure
Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.
The vendor describes its own layer and never identifies what sits underneath it. Marketing distinguishes the agentic AI framework from generic language-model output and contrasts Legartis with ChatGPT, which tells a reader that third-party models are in use without naming one. No model is named, no model provider is named, and no notice of change to either is committed. The third-party data processors published for Application users, Turicode, VSHN and LinuxFabrik, are Swiss infrastructure and operations suppliers rather than model providers, and infrastructure never answers this axis. Where inference runs is not stated; hosting in Switzerland and Europe describes where the platform sits. The clearest evidence that third-party or self-hosted training occurs at all is indirect, in the pricing matrix line offering opt-out of model training on the two upper tiers.
The vendor refers to the models without identifying what sits underneath, and its own documents disagree. The demo page states the proprietary Lexible generative model is built upon several industry-leading LLMs, the sub-processor table names no language-model provider, and the security page states that no external processors are used for customer data; a separate marketing statement refers to a zero-data-retention policy with AI partners. No provider, model, inference location or change-notification commitment is named, and the contradiction is the finding a buyer would need resolved. Demo page, security page and DPA read 6 September 2026.
Commercial Transparency
Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.
The most complete pricing disclosure in this pull. Four named plans are published with rates, units and volume entitlements: Free at one seat and two NDA reviews a month; Professional at CHF 250 or 250 euro per user per month, or CHF 2,500 or 2,500 euro per user per year excluding VAT, covering 120 contracts a year from one seat; Team at CHF 300 or 300 euro per user per month, or CHF 3,000 or 3,000 euro per user per year, covering 250 contracts a year from five seats; and Enterprise at custom pricing for 500 or more contracts a year from ten seats. A full feature matrix runs across all four tiers, and the FAQ restates the numbers in prose. Both a monthly and an annual rate are given for each paid tier, currencies are stated in both Swiss francs and euros, and VAT treatment is explicit. Free and Professional are purchasable self-serve without contacting anyone. What implementation adds is named as a set of add-ons rather than priced: additional reviews, seats, custom playbooks, custom dashboards, and Best Practice Playbooks for NDA, DPA, SaaS, purchase and commercial lease agreements are each listed On request.
A free tier is published and the paid figure is not on the surfaces read. The home page offers a free 28-day trial and a pricing page exists in the navigation; the pricing page was not opened on 6 September 2026 and is the rebuttal route in either direction. No unit of charge, tier or figure appears in the material read. Home page and navigation read 6 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.
Coverage is described with substance and its outer edge is left open. The buyer is identified precisely and is not a law firm: legal departments together with the procurement and sales teams around them, each with a dedicated solutions page, plus a construction industry page. Contract type coverage is evidenced rather than claimed, with Best Practice Playbooks published for NDA, DPA, SaaS agreement, purchase agreement and commercial lease agreement, and all contract types available on paid tiers against NDA only on the free tier. Twelve industries are listed. Language coverage is stated per tier, German, English and French on Free and all European languages on paid plans. What is not stated is where the product stops: coverage is claimed for companies of any size and industry with no boundary given, law firms as a segment are neither claimed nor excluded, and government use is not addressed.
Segment and coverage are described with substance; the boundaries are partly stated. The buyer is in-house legal with business and procurement users, and law firms as partners and customers; the question set is stated at 2,500 pre-trained questions with playbooks configurable per customer, and the company has operated from the United Kingdom since 2016 with Azure hosting there. The vendor states the product is trained by its in-house lawyers, which fixes its coverage to the contract types those questions address; no contract type or jurisdiction is named as unsupported on the surfaces read. Demo page, home page and DPA read 6 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 pricing feature matrix lists opt-out of model training as a product feature, and the security FAQ on the same page confirms it: Team and Enterprise plans offer opt-out of model training. The entitlement is tiered, so customers on the Free and Professional plans have no opt-out available to them. No agreement, policy or trust page states what the default is, what is trained on, or whose models are involved; the privacy policy last updated 6 November 2019 does not address contract content at all. The commitment located is a paid product setting rather than a published term.
The published agreement names machine-learning training as a purpose for which customer data is processed, and no opt-out was located. The data processing addendum's annex states that processing of personal data uploaded in documents includes contract risk review, data extraction, ML training, feature enhancement and customer support, and its fuller text describes training of properties created within the platform and product feature enhancement.
Two readings are carried: training of the Lexible properties a customer configures on its own contracts, which is the product working for that customer, and product feature enhancement generally, which is training for the vendor's benefit; the clause covers both. A vendor marketing statement refers to a zero-data-retention policy with AI partners so that customer data is never used to train AI models, which concerns the model providers rather than ThoughtRiver's own use; the agreement governs. Surfaces checked 6 September 2026.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
No located public material states how long prompts, uploaded contracts or generated output are retained. The privacy policy, last updated 6 November 2019, addresses website and account data only; its retention section states generally that information is kept as long as necessary and that aggregated, anonymized or pseudonymised information may be retained indefinitely, but nothing in it applies to contract content. The pricing matrix lists a contract repository with storage and version control, marked Coming soon, without any retention period.
Searched the home page, the AI Quality System page, the pricing page, the privacy policy and the legal notice on 4 September 2026; there is nothing to quote because the position is absent rather than adverse.
Retention is acknowledged without a period. The DPA commits to delete or return personal data on termination except where it is held in backups or law requires retention, with no backup retention period stated and no statement about prompts or outputs during the term; the marketing reference to zero data retention with AI partners concerns the model providers. Surfaces checked 6 September 2026.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
The pricing matrix asserts advanced role management, described as team-based access control with document-level permissions, as an Enterprise-tier feature, and lists user management on paid tiers. No published material describes how any of it is enforced, and no surface addresses separation between customers or between matters. The marketing trust blocks state no data sharing with third parties and, on the pricing page, no uncontrolled data sharing, neither of which speaks to segregation inside the platform. Segregation is therefore asserted as a purchasable feature with no published detail behind it.
The product maintains its own separation model and documents it at customer level. The security page states that contract documents and derived data are segregated from other customers and encrypted with a unique key specific to the account, with authentication through Auth0 single sign-on and multi-factor. Nothing describes walls between matters or teams within a customer, and no document management system's access model is inherited. Surfaces checked 6 September 2026.
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 states that without explicit consent Legartis will not disclose personal data to third parties other than its named processors, unless this is required by law. Disclosure under legal compulsion is therefore addressed directly. No commitment to notify the customer of such a request was located, and no discretion over notice is reserved either. Searched the privacy policy, the legal notice, the home page, the AI Quality System page and the pricing page on 4 September 2026; the legal notice separately states that data transmitted by users is treated as confidential and will not be forwarded to third parties, which addresses voluntary sharing rather than compelled disclosure.
The published DPA commits to inform the customer where law requires ThoughtRiver to process personal data contrary to the customer's instructions, unless prohibited by law, which is the shape of a compelled-disclosure notice for personal data in uploaded contracts. The commitment is narrower than a general legal-process clause: it is framed around processing instructions and personal data, and the terms of use, where a broader clause would sit, were not opened on 6 September 2026 and are the rebuttal route. No transparency report is published.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
No located public material identifies the corpus behind the product's answers. The Legal Agent is described as performing legal research with references to legal sources and as being connected to relevant legal sources, but no database, publisher, jurisdiction or collection is named, and no license or rights basis is stated. The knowledge the vendor does describe in detail is the customer's own: playbooks, Legal Best Practices and Agent Memory built from company guidelines and contract history.
Searched the home page, the Legal Agent references on it, the AI Quality System page and the pricing page on 4 September 2026.
No located public material identifies a legal corpus behind the product's answers, and the product is not built on one: the engine answers the vendor's own Lexible question set, described as 2,500 questions authored and trained by its in-house lawyers since 2016, against the customer's contract text. The questions are the vendor's own work rather than licensed law, and no primary law source or update cadence is stated. Demo page and home page checked 6 September 2026.
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. This is primarily a contract review and drafting product, so the question bites only through the Legal Agent, which is described as answering legal questions and researching with references to legal sources. No treatment signal, currency check or citator relationship is described anywhere, and no primary law source is named. Searched the home page, the AI Quality System page and the pricing page on 4 September 2026.
No located public material addresses whether authority is checked for subsequent history, and the product does not retrieve or cite primary law; its output is risk ratings, issue lists and redlines on the customer's contract. Recorded as the honest value for a product without a citator function. Surfaces checked 6 September 2026.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
The AI Quality Score is a per-requirement reliability signal surfaced to the user in a dashboard, showing how well the system currently detects each requirement in a playbook, and the vendor states a buyer can see at any time where the AI stands. Unusual-clause detection separately flags clauses falling outside expected patterns even where the playbook does not define them. Neither is an abstention path: no located material describes the product declining to answer, marking an output as ungrounded, or stopping when it cannot support a conclusion.
The vendor names hallucination as one of four error types it says occur most often and describes it as particularly critical because unpredictable, without describing a behavior that follows from detecting one.
No located public material describes what the engine does when it cannot answer a Lexible question with confidence. The product assigns risk ratings and flags issues, and the vendor publishes an accuracy figure, but no abstention path, unanswered-question handling or confidence signal is described on the surfaces read; the Accuracy page was not opened. Demo page and home page checked 6 September 2026.
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 Legartis and the corporate name Legartis Technology AG. 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.
No court order, opinion or disciplinary record naming ThoughtRiver or Lexible was located as of 6 September 2026. The AI Hallucination Cases database maintained by Damien Charlotin was searched on both names together with a general search for court findings; results returned sanctions involving general-purpose chatbots and commentary, none of which names this product. This is a statement about the public record, not a finding about the product; a contract review tool that cites no authority carries a remote exposure on this signal.
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 any guidance. The AI Quality System FAQ addresses liability directly, stating that a lawyer sending a signed document remains liable however it was created, that the AI is an assistive tool, and that legal responsibility stays with the person who approves the result. No bar, chamber or regulator is named on any surface, no ethics opinion or professional code is cited, and no jurisdiction is identified despite the product being sold across European legal systems.
Searched the home page, the AI Quality System page, the pricing page, the privacy policy and the legal notice on 4 September 2026.
No located public material names an ethics opinion, bar rule or professional responsibility framework. The vendor states its questions are trained by in-house lawyers and offers law firms a pre-screening service line, but no guidance from any bar or regulator on lawyers' use of AI is named on the surfaces read; the terms of use and the EU AI Act statement were not opened. Home page, demo page and DPA checked 6 September 2026.
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 and cost savings without addressing billing. Published figures include contract review up to 85 percent faster, playbook creation effort reduced by up to 98 percent, a DPA first review falling from 45 to 60 minutes to under 10, and cost reductions above 90 percent on portfolio-wide risk analysis. The return on investment FAQ goes further and names lower external law firm costs and avoided headcount growth as the direct return.
Nothing addresses what happens to a bill when AI-assisted work compresses the time it takes, and no per-matter record of AI-assisted work is described. The buyer here is an in-house department rather than a firm billing a client, which is the usual direction this signal assumes.
Law firms are a named buyer and partner segment and the published position on the bill is a savings claim: the Shoosmiths quote sets three minutes of platform review against four hours of qualified lawyer time, and a 2019 vendor article pitched pre-screening as a new recurring revenue stream for law firms. Nothing addresses how AI-assisted review is recorded or disclosed on a client's bill. Demo page and home page checked 6 September 2026.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
A third-party processor list is published in the privacy policy and separates website visitors from Application users, naming Turicode, VSHN and LinuxFabrik for the latter. All three are Swiss infrastructure and operations suppliers and no AI model provider appears among them, so the list does not tell a client whose models see its content, and infrastructure alone does not satisfy this signal. The list also sits in a policy last updated 6 November 2019, so its currency cannot be established.
No data processing agreement, consent pack or client-facing disclosure material was located at any access tier, and no subprocessor change notification is committed.
A current sub-processor list is published: Microsoft Azure in the United Kingdom for infrastructure, Twilio SendGrid in the United States for account emails, Okta in the EU for authentication, Abbyy Vantage in the EU for document conversion and Zoho Desk in the EU for support, each with data categories, framework and processing operations, and the transfer mechanism for Twilio. The list names no language-model provider although the engine is stated to run on several large language models, so a firm cannot answer its client's model-provider question from it; that gap is why the row sits at this value. Surfaces checked 6 September 2026.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Some elements of a record exist, short of a document-level export. A full audit trail is stated in the trust blocks on both the home page and the AI Quality System page, the vendor states that a complete audit trail documents every correction step, and that every classification is traceable via the quality score and the audit trail so misclassifications stand out. Review history is listed as part of AI contract review, and audit logs appear in the pricing matrix as an Enterprise-tier feature.
No export of a per-document record is described, the model used for a given output is not disclosed to the customer anywhere, and no disclosure guidance or template for a court was located.
Some elements of a review record are available and no export of an AI verification record is described. The Digital Issues List records every issue the engine flagged against the playbook, its risk level and resolution status, and can be shared with colleagues or downloaded as a PDF, which is a per-contract record of what the AI found and what a person resolved. Nothing states that the model used, its sources and the human verification can be exported for a court, and the product produces no court-facing work product. Vendor-supplied feature description read 6 September 2026.
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.
- AI Liability and Recourse
- Primary Law Corpus Provenance
- Good Law Verification
Which one fits
Choose Legartis if
- You want to see how reliable the review is for each requirement. Legartis shows an AI Quality Score per playbook requirement, lets your team correct how the system read a requirement inside test sets, and flags clauses outside expected patterns even where the playbook is silent.
- You want to price it without a call. Legartis publishes a free tier for two NDAs a month, Professional at CHF 250 per user a month for 120 contracts a year, Team at CHF 300 for 250 contracts, and custom Enterprise terms.
- You work across European languages and want data in Switzerland. Legartis supports every European language on paid plans, hosts in Switzerland and Europe with data on Zurich servers, offers a Word add in, a REST API and an MCP integration, and states ISO 27001.
Choose ThoughtRiver if
- You want a measured accuracy claim and a named customer figure. ThoughtRiver states 97.3 percent accuracy across its 2,500 pretrained questions, and a Shoosmiths partner reports supply agreements reviewed in under three minutes at above 90 percent accuracy, against four hours at 86 percent by qualified lawyers.
- You need data handling commitments in a published addendum. ThoughtRiver's data processing addendum names five subprocessors with locations, commits to breach notice without undue delay and to deletion or return on termination, and each customer's data is encrypted under its own key in a UK Azure region.
- You want every issue in a contract on one list. ThoughtRiver's Digital Issues List rates the risk of every point to resolve against your playbook, tracks resolution status, can be downloaded as a PDF, and pairs with suggested redlines in Word and portfolio analysis.
In summary
Legartis
Legartis, from Legartis Technology AG of Zurich, is a legal AI workspace for in house legal departments and the procurement and sales teams around them. Its agents review contracts against customer playbooks, draft from them, answer legal questions and flag risks and deadlines across a portfolio, with a per requirement quality score users can tune. The AI Legal Index grades it in the top two bands on eleven of fifteen capability axes, with A grades on AI centrality, autonomy and oversight, and pricing, published from a free tier to CHF 300 per user a month. It hosts in Switzerland and Europe and states ISO 27001. As of 4 September 2026 the index located no published customer agreement and no named model provider.
ThoughtRiver
ThoughtRiver, from ThoughtRiver Ltd of Cambridge, England, is an AI contract pre screening and review platform for in house legal teams and the law firms that serve them. Its engine answers a network of about 2,500 legal questions built by its own lawyers, scores the answers against a playbook, and produces a risk rating, a Digital Issues List and suggested redlines in Word. The AI Legal Index grades it in the top two bands on eight of fifteen capability axes, with an A on AI centrality. It publishes a data processing addendum with a subprocessor list, states ISO 27001 with a certificate number, and hosts on Azure in the UK. As of 6 September 2026 the index located no named model provider and no published price.
Questions buyers ask
Legartis vs ThoughtRiver: which is better for in house contract review?
On published evidence Legartis sits in the top two bands on eleven of fifteen AI Legal Index capability axes and ThoughtRiver on eight of fifteen. Legartis publishes its full price list and a clear control structure with per requirement quality scores. ThoughtRiver publishes stronger data handling terms and a measured accuracy figure with a named customer. Teams that want to try a tool before a sales call have more to read from Legartis.
Do Legartis and ThoughtRiver train AI on customer contracts?
ThoughtRiver's data processing addendum lists machine learning training and feature enhancement among the purposes for which uploaded contract data is processed, and no opt out was located. Legartis offers an opt out of model training only on its Team and Enterprise plans, so Free and Professional customers have none; nothing published states what is trained or on whose models. 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 accurate is ThoughtRiver?
ThoughtRiver states 97.3 percent accuracy across 2,500 pretrained questions with an average review time of 8.2 minutes, and quotes a Shoosmiths partner on reviews above 90 percent accuracy in under three minutes, against 86 percent for qualified lawyers over four hours. No test set, sample or method accompanies either figure. Legartis publishes a per requirement quality score inside the product rather than an overall figure. 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 Legartis and ThoughtRiver cost?
Legartis publishes every plan: free for two NDAs a month, Professional at CHF 250 or 250 euro per user a month for 120 contracts a year, Team at CHF 300 for 250 contracts from five seats, and custom Enterprise from ten seats. ThoughtRiver offers a 28 day free trial, and its paid pricing was not located on the pages read. 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 Legartis and ThoughtRiver both leave unpublished?
Whose models read the contracts and what the vendor stands behind. Neither names its language model provider, and the index located no liability, indemnity or warranty position for either. Neither addresses privilege or professional secrecy, names bar or regulator guidance on AI, or describes what its engine does when it cannot answer a question with confidence. 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. ThoughtRiver says its engine runs on several large language models, its subprocessor list names none, and its security page says no external processors handle customer data; those statements need reconciling. Its accuracy figures come without a described test set. Legartis publishes no customer agreement, and its only data document is a 2019 privacy policy that does not cover uploaded contracts, so its low grade on liability records what could be read. Legartis was verified on 4 September 2026 and ThoughtRiver on 6 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.