Omnilex

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 engine of what a buyer uses and a substantial content asset sits beneath them. The corpus is counted openly at 1,057,541 sources across five categories, dominated by 797,321 Swiss court decisions and 22,181 laws and articles, and that collection would remain a usable Swiss law database if every model were removed, which is the product several incumbents in this market already sell. What the AI adds is the reasoning layer over it, plus one thing that is genuinely model-created: the Omnilex Commentary, 222,525 annotations the company writes itself article by article on Swiss law, which exists only because the models produced it. Agents built without code from a firm's own playbooks sit on top. The pricing page frames the whole offering as one licence covering corpus, models and agents together, which is an accurate description of a product where the machine learning is the mechanism and the collection is the substrate. Verified 2 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 well labelled. The stated design is that every claim is cited and ready for the file, and the Omnilex Commentary carries the line that every statement is citable, with the worked example on the home page showing an annotation to article 336c of the Code of Obligations footnoted back to its authorities. The user also selects which source types the research draws on, choosing between court rulings, commentary or internal data, which is a described retrieval control rather than an opaque one. One practice deserves specific credit and is rare in this corpus: the commentary the vendor generates is expressly labelled as AI-drafted, so a lawyer can tell machine-written analysis from published doctrine. What is absent is measurement. No accuracy figure, evaluation, benchmark or test set appears anywhere. Clause 6 of the terms goes the other way and disclaims it directly, stating that the provider accepts no liability for the accuracy or legal usability of generated content, which is candour about the limitation rather than evidence about the rate. Verified 2 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 review obligation is contractual and the agent layer is user-triggered. Clause 6 of the terms states that the user remains responsible for reviewing, adapting and using all content, which places the checkpoint on the lawyer in the operative document rather than in marketing. The product framing is consistent with it, describing AI-assisted suggestions at clause 2 and, on the product page, a system that analyses documents, laws and judgments, structures arguments and highlights relevant content. Agents are described as built from a firm's own playbooks and review steps and run on demand, so they execute when a person starts them rather than autonomously, and the worked example sets out its four steps openly. What is not published is the boundary. Nothing states what an agent completes without a human once started, what happens when a step fails or returns nothing, or what threshold separates a suggestion from an action. The firm authors the workflow, so the operating envelope is customer-set and undescribed. Verified 2 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.

Ten customers are named by logo across private practice, insurance, corporates and the state: the Swiss Confederation, Bayer, the legal-protection insurer Dextra, the pension fund Asga, the disability charity Pro Infirmis, and the firms SwissLegal, Ruoss Voegele, Neuwerk, Domenig and Seeger Frick. Two carry dedicated case pages with named individuals speaking on video, Florian Probala, Chief Claims Officer at Dextra, and Dr. iur. Marco S. Marty, Founding Partner at MV Legal Partners. A named public authority among the references is worth noting on its own, because a federal administration adopting an AI research tool is a procurement signal a buyer can weigh. One figure is published, that over 6,000 lawyers use the platform, described as the most-used AI workspace in Swiss law. What holds this below A is measurement: the 6,000 figure is an adoption count rather than an outcome, no figure is attached to any named customer on the pages read, and nothing is dated. The two case pages were not opened this pass. Verified 2 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.

The privilege limb is met squarely, which most records in this pull fail, and the training limb is not addressed at all. Omnilex engages the professional-secrecy question in its own legal system's terms rather than gesturing at confidentiality: the home page states that models are processed on Swiss infrastructure so that attorney-client privilege stays intact, that the platform was built for professional secrecy from day one, and that the licence agreement with its data processing agreement is drafted to respect professional secrecy and auxiliary-person liability. That last concept is the Swiss doctrine governing a lawyer's responsibility for the third parties they engage, and naming it is a substantive engagement with the buyer's own obligations. Retention and deletion are addressed at clause 7 of the terms, with data deleted after the contract ends subject to statutory retention. The position on model providers is partly answered by keeping processing on Swiss infrastructure. Two limbs fail. Nothing published describes segregation between customers, users or matters, despite the product organising work into named projects and mandates. And nothing located addresses whether customer content is used for training, in either direction, as set out on that signal. Verified 2 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.

A real position on tooling versus advice is published in the operative terms rather than a footer. Clause 2 characterises the service as software that supports users in working on legal questions with AI-assisted suggestions, which frames output as input to a lawyer's work. Clause 6 states that the provider accepts no liability for the accuracy or legal usability of generated content and that the user remains responsible for reviewing, adapting and using all content, which is the supervision dimension stated as an obligation. Coverage is bounded in practice, since the product, the corpus and the commentary are Swiss law throughout and the terms are governed by Swiss law with jurisdiction in Zurich. The professional frame is engaged more directly than usual through the references to professional secrecy and auxiliary-person liability. Two things are missing. No express statement that output is not legal advice, and no statement about the professional relationship, appears anywhere on the surfaces read. And no bar or professional body guidance is named, with neither the BGFA nor Swiss Bar Association material referenced, which is recorded on the bar guidance signal. Verified 2 September 2026.

AI Governance and Bias Disclosure

Published governance over model behaviour: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.

A governance surface exists and its contents could not be read. The SafeBase trust centre carries a dedicated AI section listing AI Security, AI Governance and AI Monitoring as published items, which is more structure than most records in this corpus offer and is evidence that a position exists rather than that none does. The item pages themselves are driven by query parameters that did not render, so what those headings contain is unestablished, and this grade rests on the existence of the surface rather than on any substance credited from it. Around it sit an ISO/IEC 27001:2022 certification audited across the whole organisation, a dated penetration test report, and an ESG section covering anti-bribery, anti-competitive practices and modern slavery, all of which are adjacent subjects this axis treats separately. What was located nowhere, on any surface, is anything about bias or uneven output: no statement addresses whether the system performs differently across cantons, languages, court levels or party types, which matters for a product whose corpus spans a federal system with three official languages. Verified 2 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 across most of the ground, short of specificity on two elements. Access control is the strongest part and is stated as the default set legal teams expect: SAML single sign-on, audit logs, IP allow-listing and data lifecycle management, with the trust centre adding credential management, brute-force protection and account hardening, endpoint detection and response and disk encryption. Encryption is stated at rest and in transit. Subprocessors are named, with the trust centre legal section identifying Amazon Web Services and Microsoft Azure. Incident practice is stated, with a dedicated response team and an incident response item in the trust centre, and backups, business continuity and disaster recovery and anti-DDoS all appear as published items. What keeps this off the top band is that two of the five elements are stated without being specific enough to hold the vendor to. Retention is given only as data deleted after the end of the contract unless statutory retention applies, with no period for anything during the term. And the incident commitment carries no notification deadline to the customer, the trust centre saying instead that more detail is available on request. The named subprocessors are also infrastructure only, which is examined on the model supply chain axis. Verified 2 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.

A real position is published and it is short of the full picture. Clause 6 of the terms sets out the allocation in four sentences: use is at the user's own risk; the provider accepts no liability for the accuracy or legal usability of generated content; the user remains responsible for reviewing, adapting and using all content; and the provider's liability is limited to gross negligence and intent. That last limb is more informative than it looks, because it is the maximum exclusion Swiss law permits, so a reader knows precisely where they stand, with nothing recoverable for ordinary negligence and no monetary cap at all on the residual liability for gross negligence and intent. Cyber insurance is separately published as a document in the trust centre, which is one of the four things the top band asks for. What is absent is the rest. No indemnity runs in either direction, no warranty is given on output and clause 6 expressly disclaims it, and no cap figure appears. The availability commitment at clause 5 is weak and honest about it, targeting at least 90 per cent outside maintenance windows and stating it is not guaranteed. One scope limit belongs on the record: clause 9 states that use is based on a separate licence agreement, and that agreement is not published, sitting instead as a template behind the trust centre's access flow, so what a signed customer actually receives is not established. Verified 2 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.

Integration is claimed and never specified. The product page states that a firm can connect its document management and legal tools without media breaks and keep its proven workflows, which is a direct claim to integrate with the systems legal work already lives in. Not one of those systems is named. No document management product is identified, no practice management or e-filing tool, no connector list, and no integrations page exists. Enterprise plans are stated to include custom integrations and APIs, but that is a scoping conversation rather than a published capability, and no API or developer documentation was located. What is concrete is authentication rather than workflow: SAML single sign-on across plans and company SSO at enterprise level. A claim to integrate, with the systems unnamed and nothing an implementer could use, is what this band describes. Checked the home page, product page, pricing page, terms, privacy policy and trust centre on 2 September 2026. Verified 2 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.

Residency is stated with a named city and a named cloud, and the tenancy limb is silent. Omnilex publishes that it is hosted on Microsoft Azure in Zurich, that data is encrypted at rest and in transit and hosted entirely on its Swiss Azure cloud, and that models run on Swiss infrastructure, which is the point on which the privilege claim rests. Processing is separated from hosting and bounded, with the pricing FAQ stating that data is processed in Switzerland and in selected EU countries recognised as adequate by the Swiss Federal Data Protection and Information Commissioner. Naming the supervisory authority rather than asserting adequacy is a good habit. Two things hold this below the top band. No tenancy or isolation model is published anywhere: nothing states whether the platform is multi-tenant, single-tenant or how one firm's data is separated from another's, which is the limb this axis pairs with region. And there is a small tension a buyer should notice between hosting stated as entirely on the Swiss cloud and processing stated as extending to selected EU countries; both are published, neither is reconciled. No on-premises or customer-cloud option is offered. Verified 2 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.

A real trust centre exists, it renders, and it is inventoried. The SafeBase portal at trust.omnilex.ai lists ISO/IEC 27001:2022 under compliance, and among its documents a penetration test report dated 5 September 2025, the Omnilex Licence Agreement including a DPA template, technical documentation, cyber insurance, a data security policy and a network diagram, alongside sections for product, data, application and endpoint security, infrastructure and legal. The certification is described on the home page as audited across the whole organisation, which is a scope statement rather than a bare badge, and a document carrying a date is visible without any request. What is not established is the access tier for the documents themselves. The portal offers a public and private split with a get-access flow, and it does not state whether access is granted instantly on an email or an NDA click-through, or only after review, so the lower tier is graded and the reason recorded. No auditor or certification body is named, and no certificate number, issue date or validity period is published. No request was submitted. Verified 2 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 models are named to the version on the home page, which nothing else in this pull does, and the disclosure stops short of the rest. Five are listed as selectable, GPT-5.6, Claude Opus 5, Mistral Large 3, Kimi K2.6 and Grok 4.6, with the user choosing which answers a given question, and the accompanying claim is that every frontier model is available and processed on Swiss infrastructure so that attorney-client privilege stays intact. Where inference runs is therefore stated at estate level, on Microsoft Azure in Zurich. The trust centre names Amazon Web Services and Microsoft Azure as subprocessors. Two limbs are missing. No model house is identified as a legal entity and none appears in the subprocessor list, so a reader can see that Claude Opus 5 is available without learning which entity contracts for it or what it may retain, and the inference that the models run inside the Swiss cloud estate rather than through the vendors' own APIs is left to be drawn rather than stated. And no commitment to notify customers when the model set changes was located, which matters on a product that markets access to whichever frontier model is current. Verified 2 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 strongest commercial disclosure located in this pull. Rates are published with the unit and the term: CHF 2,000 per user billed annually for a single user, CHF 1,600 per user billed annually for teams from two users, and a custom tailored agreement for enterprise from twenty users, all stated excluding VAT, with an annual and monthly toggle so a buyer can compare billing cadences. The unit is explicit and the vendor makes a point of it, stating that the price scales only with the people who use it and that corpus, models, agents and security are in every plan. The limb this axis asks for that almost no vendor answers is answered directly: implementation is not an extra. Onboarding, set-up, training and support are listed as included in every plan without asterisks, and the FAQ states that nothing essential is held back in a higher tier, with Team adding collaboration and know-how integration and Enterprise adding SSO, custom integrations and APIs. A fourteen day free trial of the full product requires no card and no commitment. Usage limits are published too, at up to 1,000 pages and 500MB per document with otherwise unlimited use. The only withheld figure is the enterprise tier, which the page explains is a scoping conversation about integrations and support rather than a different product. Verified 2 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.

Four buyer segments each have their own page and each is described in terms of the work rather than as a label: law firms handling research, drafting and review across mandates; in-house teams moving from contract review to regulatory questions; public administrations applying federal and cantonal sources; and insurers assessing claims against statute and precedent. Covering government and insurance explicitly is unusual and is borne out by the named references, which include the Swiss Confederation and a legal-protection insurer. Jurisdictional coverage is stated precisely through the corpus itself, with federal and cantonal decisions, statutes, official documents and literature counted by category. What is missing is the boundary. No practice areas are enumerated, and while the worked examples on the home page range across employment, tenancy and accident insurance law, that is illustration rather than a coverage statement. No firm size band is given beyond the seat thresholds in pricing, nothing addresses whether the product serves any jurisdiction other than Switzerland, and nothing states what it does not support. Verified 2 September 2026.