Advocacy vs Opus 2: how they compare in 2026

Remove the models and nothing remains to sell. The product is defined by its model architecture: Axiom is described as a proprietary case memory layer that ingests and understands every document in a case, builds memory persisting across every interaction, and reasons across the entire matter rather than one document at a time, expressly in contrast to querying a generic legal database. All three surfaces on top of it are generative, being a conversational assistant answering questions about the matter, a drafting environment with generation streaming onto the page, and a structured extraction tool. The vendor's own positioning makes the point sharply, rejecting the sidebar plugin model on the ground that a plugin is bolted onto someone else's product while this is the product. There is no document management system, matter management layer or review platform underneath that a firm would license on its own. Graded on material recovered through search indexing on 5 September 2026, since the domain refuses automated fetching; the reasoning here does not depend on any unreachable surface.

The models drive several capabilities a buyer would miss, on a platform that ran without them for most of its life. Opus 2 Ingenuity is presented as an agentic framework and the feature list is substantial: summarisation and analysis of key documents, extraction of people, organisations, events and legal topics, natural-language querying scoped to a single matter or a selected document set, a prompt builder with reusable saved prompts, drafting and sense-checking of work product, speech recognition producing legal-grade transcripts, and live hearing analysis that compares testimony as it is transcribed against the documents in the bundle to surface inconsistencies. That last one is the most machine-dependent thing in this lane. Underneath it sits eighteen years of case management, shared document workspace, annotation, chronologies, electronic bundles, evidence presentation and hearing infrastructure, none of which needs a model to work, and the vendor's own framing places the AI on top of that rather than in place of it, describing the platform as the source of the matter intelligence the AI runs on. Graded level with Casepoint, Consilio and Nextpoint's peers on the same reasoning: the engine of real capabilities, layered on a workflow system that stands without it.

Accuracy is asserted and the mechanism behind it is named without being described. The reachable material claims outputs a buyer can trust, states that citations are verified inside the drafting surface, and makes an architectural argument that grounding comes from case memory built on the matter's own record rather than a generic database, which the company frames as the system learning about the case before it does the work. That is an assertion of accuracy and a claim about where grounding comes from. What is absent from everything reachable is any test or description: no accuracy figure, no evaluation, no test set, no error rate, no statement of failure modes, and nothing describing what verifying a citation actually consists of or whether a user can open the source behind an assertion. The floor band does not fit, because accuracy and grounding are both addressed rather than unmentioned. Graded on the home page and the vendor's December 2025 announcement, recovered through search indexing on 5 September 2026; the site refuses automated fetching and any accuracy or benchmark page it publishes is the route to correct this row.

Quality is asserted repeatedly and never measured. The published claims are confident: better data relationships produce more accurate insights, AI features only contribute information that is relevant and helpful to lawyers, and the design prevents errors and erroneous data from having a negative impact on a case. Grounding is claimed in general terms, with query answers described as based on the content inside the selected documents and the data surrounding the matter, and the whole system described as locked down to a single matter. What is absent, checked across the AI page, the home page, the security page and the arbitral institutions page on 31 Aug 2026: no accuracy figure of any kind, no test set, no benchmark, no evaluation, no error rate, and no hallucination disclosure. Nothing states whether a generated answer or summary cites the document it came from, or whether a reader can open the source from the output, which is the verification surface this axis turns on. The gap matters most on the hearing feature, where the system flags inconsistencies between live testimony and the bundle in real time and counsel may act on that within minutes. One limb does not apply and is neither credited nor penalised: a citator or good-law check is out of scope for a platform working on a matter's own evidence.

The division of labour is implied by the shape of the product and never described as a control. What is reachable presents a workspace the litigator works inside: research sits beside the draft, generation streams onto the page the user is writing, and the recorded activity types are queries, reviews and notes attributed to people against docket entries, which describes a lawyer doing the work with the system assisting. The company also states that the system learns about the case before it does the work, which is a statement about sequencing rather than supervision. Nothing reachable describes a review gate, an approval step, a threshold at which the system acts without a person, what happens when an output is wrong, or any abstention or confidence behaviour, and one line runs the other way, promising that what you draft is what you file. This is the middle band rather than the floor because oversight is implicit in the described workflow rather than absent from it. The product pages that would ordinarily carry a control description could not be fetched on 5 September 2026 and are the rebuttal route.

The clearest published statement of an autonomy limit in this pull, and it comes with a worked example. The Control principle commits that lawyers must control the process, using their own knowledge, experience and ingenuity, before any data is saved permanently, which is a persistence gate rather than a slogan: AI output is a suggestion until a person commits it. The chronology feature shows what that means in practice, and the vendor explains the reasoning rather than just the behaviour. It states that automatically generating a chronology fills a timeline with irrelevant placeholders a team must then review and remove, and that its approach instead surfaces events the AI identifies for the lawyer to assemble, deciding relevance themselves. The entity workflow has the same shape, presenting extracted people, events and organisations with a summary of each so the user can remove irrelevant ones and eliminate duplicates before anything is added to the case. Against that, checked on the AI page, the home page and the security page on 31 Aug 2026: no threshold at which a feature declines to answer, no confidence or uncertainty signal exposed, no route to report or correct a wrong output, and nothing describing what happens after an error reaches a work product.

No production evidence was located in any reachable material. The buyer segment is stated, mid-sized and Am Law firms, and no firm within it is named. No customer, case study, logo, deployment or figure appears in the home page material recovered through indexing, and none appears in the vendor's own December 2025 announcement, which is the surface where a company ordinarily names its first references and which instead quotes only its own chief executive and chief operating officer. The interface material published uses a fictional matter, Estrada v. Kincade, with invented docket entries, so nothing there is a deployment either. The grade rests on the announcement rather than on the block: a press release naming no customer is evidence in its own right. The customer-facing pages of the site could not be fetched on 5 September 2026 and are the rebuttal route; a named reference with a date or a figure would move this row.

Deep, named and current deployment evidence, with no measured outcome attached to any of it. Five customer quotes carry a name, a title and a firm: a contentious legal technology portfolio manager at Linklaters, the chief knowledge and innovation officer at Dorsey & Whitney, the director of practice technology at Vorys, a litigation partner at Pillsbury, and a paralegal at Shepherd and Wedderburn, with two more from barristers' chambers clerks. Around nineteen firm logos appear including Baker Botts, Dentons, Herbert Smith Freehills Kramer, Holland & Knight, Mishcon de Reya, O'Melveny, Troutman Pepper Locke and Wilson Sonsini. The institutional deployments are the strongest element because they are independently checkable: the Singapore International Arbitration Centre's SIAC Gateway is stated to be powered by Opus 2, the Dubai International Arbitration Centre is named as a partner, and the transcript and hearing services are stated to have supported the Manchester Arena Inquiry, the Grenfell Tower Inquiry phase 2 and the Hillsborough Inquiry. Two adoption figures are published, 98 per cent of Am Law 50 firms and 92 per cent of the Financial Times top 25 most innovative firms. What is missing is measurement: those figures describe reach rather than result, no quote carries a number or a date, and no case study with a stated method was located.

The commitment that matters cannot be read before signing, and the vendor's own policy is what establishes it. The published privacy policy states that the company does not use customer-uploaded content to train its AI systems, which is a real and specific undertaking; the same policy opens by scoping itself to the Website and provides that the handling of data uploaded to the AI platform is covered separately in individual Customer Agreements. A commitment stated in a document that disclaims coverage of the platform has not been made about the platform, and the instrument that governs it is negotiated per customer and therefore unpublished by design. That is the band's second limb exactly: the commitment lives in a sales conversation and cannot be read in advance. Nothing reachable addresses privilege or work product treatment, separation between matters or between firms, or what any model provider retains, on a platform the vendor describes as holding strategy notes, evidentiary rulings and work product, and which it intends to make shareable across a team. Any security or terms page the site publishes could not be fetched on 5 September 2026 and is the rebuttal route.

One precise commitment on the question that matters most for this product, and nothing else. The Secure principle states that clients can be certain their data will never be shared outside their instance or between two or more cases inside an instance, which is an explicit cross-matter and cross-tenant undertaking rather than a general assurance, and it is reinforced by the query feature being described as locked down to a single matter or selected documents, and by named infrastructure controls including firewalls, encrypted volumes, logical identifiers and ownership tags. The arbitral institutions material acknowledges the confidential nature of arbitration directly. That is the whole of it. No customer agreement of any kind is published, so none of it is contractual: the footer carries a privacy policy, a cookie policy, a modern slavery statement, a supplier code of conduct and a code of business conduct, and no terms of service, master agreement or data processing addendum. In consequence nothing addresses whether client material is used to train or refine models, no retention position exists for documents, prompts or generated output, and privilege and work product are nowhere treated in writing despite the platform holding the working papers of some of the most sensitive disputes and public inquiries in the country. The privacy policy was not opened on 31 Aug 2026 and is the rebuttal route.

Nothing in the reachable material addresses the advice line, on a product that drafts filings. No statement that the company is not a law firm, no disclaimer that output is not legal advice, no description of the professional judgement required before a draft is filed, no jurisdiction limit and nothing on supervision or competence appears in the home page material, the privacy policy or the vendor's own announcement. The exposure is specific rather than formal, since the product is sold to surface critical facts and draft winning motions, its drafting surface promises that what you draft is what you file, and its own framing is that lawyers who control the context win. The terms of service, which is where such a statement would ordinarily sit, did not surface in any search; that is weak evidence of absence rather than proof, and this note does not assert the document does not exist. The home page and privacy policy were attempted directly on 5 September 2026 and returned bot detection, and the site's own terms and product pages are the rebuttal route.

Nothing published addresses the advice line, and no document on the property would carry it. The footer legal inventory was read in full on 31 Aug 2026 and consists of a privacy policy, a cookie policy, a modern slavery and human trafficking statement, a supplier code of conduct and a code of business conduct. There is no terms of service, no disclaimer page and no professional responsibility statement. Nothing states that generated output is not legal advice, nothing addresses where responsibility sits when a drafted document or a flagged inconsistency proves wrong, and no bar or regulatory authority is named anywhere, including ABA Formal Opinion 512 and, for a vendor headquartered in London and selling to English firms and chambers, any Solicitors Regulation Authority or Bar Council guidance. What does exist is recorded rather than credited, because it is the inverse of what the band above describes: a supervision posture is published and a disclaimer is not. The vendor states that it is not trying to replace a lawyer's intelligence, that the AI is designed to enhance human ingenuity, and that lawyers must apply their own judgement before anything is saved. That is a statement about how the work is done, not about who answers for it.

No governance position appears in any reachable material. Nothing names an accountable owner or function for model behaviour, describes pre-release evaluation, sets out principles or a policy, or claims a management system or certification, and nothing addresses uneven output. The gap has a specific shape on this architecture and is worth stating: a case memory layer that accumulates contested issues, strategy notes and evidentiary rulings across a matter, and that the company intends to make shareable across a team, is making judgements about what counts as case-critical context, and nothing reachable describes how that selection is governed or tested. The privacy policy, which is the one policy instrument reachable, addresses website data handling and does not touch model governance. The site refuses automated fetching, attempted on the home page and privacy policy on 5 September 2026; any responsible AI or governance page it publishes is the rebuttal route.

Five published principles that are better written than most and still not a governance framework. Empower, Enhance, Control, Secure and Trust each get a paragraph, and unusually the paragraphs explain design decisions and rejected alternatives rather than restating the principle: the vendor describes why it does not auto-generate chronologies, why it preserves existing legal workflows rather than replacing them, and what it will not let the AI write into a case without a lawyer's action. Two of the five carry real commitments, on human control before persistence and on data never crossing an instance or a case boundary. What is absent is the apparatus. Searched the AI page, the home page, the security page and the arbitral institutions page on 31 Aug 2026 and located no accountable owner for model behaviour, no described pre-release testing or evaluation regime for the AI specifically, no published result of any kind, no model documentation, and no external standard or certification covering AI as distinct from information security. STRIDE threat modelling and CREST-accredited penetration testing are published but cover software security rather than model behaviour. Nothing anywhere addresses bias or uneven output, which has a specific edge here: the hearing feature analyses live testimony, and how it performs across accents, languages and interpreted evidence in international arbitration is exactly the question, unasked.

Nothing reachable addresses retention, deletion or access for the material the platform holds, and the reason is established from a readable document rather than inferred from the block. The privacy policy describes secure storage with a database provider, industry standard security protocols and regular security reviews, retains personal data for as long as reasonably necessary with no period stated, and names service providers by category only as a database provider, an analytics provider and payment processing. It then scopes all of that to the Website and refers the handling of data uploaded to the AI platform to individual Customer Agreements. So the stewardship material that exists covers website visitors, and the material governing a firm's case files is unpublished by design. No retention period, deletion commitment, access control model, encryption specification, subprocessor list or incident practice for platform data was located, and no AI or model provider is named anywhere. A security or trust page did not surface in any search, which is weak evidence of absence rather than proof; the site was attempted directly on 5 September 2026 and refused automated fetching, and any such page is the rebuttal route.

Strong on how the system is built and silent on what happens to the data inside it. What is published is specific and technical: AES-256 encryption at rest, TLS 1.2 in transit with internal component traffic also over TLS and email encrypted end to end, logical data segregation using firewalls, encrypted volumes, logical identifiers and ownership tags, client-defined password policy, multiple multi-factor authentication options, single sign-on linking to the client's own provider, and full client control over creating, modifying and removing users. The development side is better documented than most vendors manage, with STRIDE threat modelling and an assigned risk owner for each identified threat, regular CREST-accredited external penetration testing, automated and manual code assessment covering third-party dependencies, a continuous integration pipeline, and a formal change management protocol with documented risk and impact assessment. Against that, the data lifecycle is absent. Searched the security page, the AI page and the footer inventory on 31 Aug 2026 and located no retention period for anything, no deletion or return commitment, no data export provision, no subprocessor list, no named cloud provider, and no incident response or breach notification practice at all. Four of the five elements this axis asks for are missing. The privacy policy was not opened and is the rebuttal route.

Nothing published addresses who bears the loss when the system is wrong, and on this record that is a design consequence rather than an omission. The privacy policy states that the handling of data uploaded to the AI platform is covered separately in individual Customer Agreements, so the governing instrument is negotiated per customer and is not published anywhere by construction. No indemnity, liability cap, warranty position, disclaimer of warranties, service level commitment or insurance statement is therefore readable before a buyer enters a sales process, and none appears in any reachable material. The band above does not fit, because it requires a standard limitation clause disclaiming the exposure the product creates and no such clause can be read. The exposure is not theoretical on a product whose drafting surface promises that what you draft is what you file. This row rests on the vendor's own statement about where platform terms live rather than on the fetch failure; publication of a standard-form agreement is the rebuttal route.

Nothing is published on who bears the loss when the system is wrong, because no customer agreement is published. The footer legal inventory, read in full on 31 Aug 2026, is a privacy policy, a cookie policy, a modern slavery and human trafficking statement, a supplier code of conduct and a code of business conduct in PDF. There is no terms of service, no master services agreement, no subscription terms, no service level agreement and no data processing addendum. Nothing therefore states a liability cap, an indemnity in either direction, a warranty on the software or on generated output, an insurance position, a limitation period, a governing law or a forum. The security page refers in passing to Opus 2 Services and Agreements meeting global legal and regulatory requirements, and names six regimes they are said to satisfy, which confirms agreements exist without publishing any of them. A buyer cannot read the allocation of risk in advance at any level. This is the second vendor in this category to publish no customer agreement at all, and the pattern in both cases is the same: an enterprise vendor whose customers negotiate individually and never require a public one.

Connection is claimed by category and no counterparty is named. The reachable material states that the platform synthesises case-level context including from e-discovery and legal research platforms, which identifies the two classes of system a litigation team runs and names neither a product nor a vendor within them. The drafting environment is described as Microsoft Word-native and as the only native editor in legal AI rendering .docx with full fidelity, which is a file-format and editing claim rather than an integration into a document or matter management system, and the company positions it explicitly against Word plugins. Nothing reachable describes what data moves between the platform and any external system, in which direction, on what trigger, or what a firm must configure, and no API or developer documentation surfaced. This sits in the middle band because real connection classes are named rather than left as logos. The integrations material on the site could not be fetched on 5 September 2026 and is the rebuttal route.

Integration is referenced by category and never by name. The AI page states that core documents flow seamlessly from eDiscovery and document management systems into a single workspace, which tells a buyer the direction of travel and identifies neither system. The security page describes linking to a client's own single sign-on provider to centralise account control, which is a real and useful enterprise integration but an identity one rather than a practice systems one. Beyond those, checked the AI page, the home page, the security page, the arbitral institutions page and the full navigation on 31 Aug 2026: no document management system is named, with nothing located for iManage or NetDocuments, and nothing for matter management, e-billing, court filing, Word or Outlook. No application programming interface, developer documentation or integration index was located. There is a real physical integration the record should not miss, since it is unusual: the vendor operates hearing infrastructure and has partnered with the International Dispute Resolution Centre in London to build a dedicated suite, which is an integration into a venue rather than a system. The platform page was not opened and is the most likely home for named connectors.

Cloud delivery is implied throughout and neither co-equal limb is stated in any reachable material. The product is described as a shared workspace a team logs into, with case memory persisting across interactions and intended to become shareable across a firm, and the privacy policy refers to secure storage with a database provider, so delivery over the internet through infrastructure the vendor arranges is not in doubt. Beyond that nothing is available. No statement says whether the platform is single or multi-tenant, no dedicated or isolated option is described, no region is named for storage or processing, no cloud provider is identified, and no residency commitment was located. The only geographic fact reachable is corporate, a San Francisco base, which locates the company rather than the data. No on-premises or self-hosted option is mentioned. Any security or enterprise page the site publishes could not be fetched on 5 September 2026 and is the rebuttal route.

Residency is offered explicitly, and where processing happens is not addressed. The security page states that the vendor hosts the client's solution and data in the geographical location that best fits their needs, which is a residency option stated as a customer choice rather than a fixed footprint, and it is corroborated by the shape of the business: three regional operations with separate EMEA, North America and Asia Pacific sites, and a published list of the data protection regimes the services are said to meet across those markets, naming GDPR, UK GDPR, Canadian PIPEDA, the California Consumer Privacy Act, the Singapore Personal Data Protection Act and the Australian Privacy Act. Tenancy is described in substance if not in name, with data said to be self-contained through firewalls, encrypted volumes, logical identifiers and ownership tags, which is a logical rather than physical separation model. What is not published: no list of the available regions or data centre locations, no named cloud provider or statement of whether the infrastructure is the vendor's own, and no distinction between where data is stored and where processing or model inference happens, which matters for a customer choosing a jurisdiction precisely to keep material inside it.

No independent security attestation appears in any reachable material. No SOC 2 of either type, no ISO certification, no penetration test summary, no named auditor, no examination period and no certification claim of any kind was located, and no trust centre or portal surfaced in search. What the privacy policy does offer is internal rather than independent: secure storage with a database provider, industry standard security protocols and regular security reviews, with no assessor, scope, date or report attached to any of it. The band above does not fit, because it describes badges displayed without scope or report, and no badge or certification mark was located to display. This row is graded on a documented absence in the material that is reachable, in the same way the AscentAI governance row was graded against an absence with an unrenderable trust portal named as the route to correct it. The home page and privacy policy were attempted directly on 5 September 2026 and returned bot detection; any certification or trust page the site publishes is the rebuttal route.

Real, current, named certification with nothing attached to let a buyer verify it. The security page, last modified 19 Aug 2026, names ISO/IEC 27001:2022 and Cyber Essentials Plus, and adds a genuine third-party testing element in regular CREST-accredited penetration testing by external groups, alongside STRIDE threat modelling with assigned risk owners and automated and manual code assessment. Naming the 2022 revision of ISO 27001 rather than the superseded 2013 version is worth crediting, since the transition deadline has passed. What is missing is everything around it: no certification body or auditor is named, no certificate number or scope statement is given, no validity period or issue date appears, no route is published for obtaining the certificate or any report, there is no trust portal, and no SOC 2 of any type is claimed, which a US buyer will look for first. One inconsistency belongs on the record: the arbitral institutions page states ISO 27001:2013 while the security page states 2022, so the same site publishes two different versions of the same certification and one of them is out of date.

The vendor names its own architecture and identifies nothing underneath it. Axiom is presented as a proprietary case memory architecture and the platform as context-first and AI-native, with three named generative surfaces built on it, so a distinct architecture is described rather than merely labelled. No model is named, no version, no provider and no hosting arrangement, and no commitment to notify customers of a change was located. The privacy policy comes closest and stops short in a way worth recording: it names service providers by category only, as a database provider for storage, an analytics provider and payment processing, and identifies no artificial intelligence or model provider at all, on a platform whose entire value proposition is model-driven. This is the middle band because the architecture is described; a buyer simply cannot learn from any reachable source whose model reads a firm's matter file. The site refuses automated fetching, attempted 5 September 2026, and any model or subprocessor page is the rebuttal route.

The vendor names its AI and never says what runs it. Opus 2 Ingenuity is described as an agentic AI framework running on the matter intelligence generated in the platform, and the capability set is set out at length across summarisation, entity extraction, natural-language query, prompt building, drafting and real-time speech recognition. Nothing identifies what sits underneath any of it. Searched the AI page, the home page, the security page and the arbitral institutions page on 31 Aug 2026: no model is named, no provider is named, no distinction is drawn between first-party and third-party models, no version appears, no inference location is stated as distinct from the general offer to host in a chosen geography, and no commitment exists to notify customers when a model or its configuration changes. The one adjacent commitment is a boundary rather than an identification: client data is stated never to be shared outside the client's instance, which tells a buyer where data does not go without telling them what processes it inside. For a platform whose AI reads live testimony in a courtroom, the absence of a named model is the disclosure a tribunal would most likely ask about.

No pricing information appears in any reachable material, at any level, including the unit of charge. No rate, band, tier, term or implementation cost was located on the home page material recovered through indexing, in the vendor's own announcement, or in the privacy policy, and no pricing page surfaced in search. Nothing states whether the platform is charged per seat, per matter, per firm or on usage. One structural fact points the same way and is evidenced rather than assumed: the privacy policy records that platform data is governed by individual Customer Agreements, which indicates a negotiated commercial relationship rather than published rates. That is consistent with the stated buyer segment of mid-sized and Am Law firms. No pricing row is owed on this record. Any pricing page the site publishes could not be fetched on 5 September 2026 and is the rebuttal route.

No pricing information is published at any level, including the unit of charge. The full primary navigation covering who we help, platform, AI, solutions, resources and company, and the complete footer across all three regional sites, were read on 31 Aug 2026: there is no pricing or plans page in either, and no pricing entry anywhere in the site structure. No rate, no unit, no tier or package names, no volume band, no term length and no statement of what implementation, hearing services or transcription add, which matters here because the offering mixes software subscription with delivered services including court reporters, transcript production and hearing room setup, each of which would ordinarily be priced differently. Every call to action across the home page, the AI page and the solutions pages is to watch a demo or contact the company. The only cost-adjacent claim located is a general assertion that connected data and workflows drive more efficient AI usage so a customer spends less.

Coverage is claimed at a level of generality that does not tell a buyer whether their practice is served. The segment is stated with reasonable precision, mid-sized and Am Law firms running high-stakes litigation, and the scope within a matter is stated as complaint through trial, spanning parties, claims, timelines, procedural history, evidentiary rulings, strategy notes and work product. That describes a stage of work rather than a practice. No practice area is named anywhere reachable: nothing says whether the product suits commercial, employment, securities, product liability or class work, and the only signals in that direction are indirect, being a fictional demo matter involving price-fixing and market allocation and a leadership team drawn from securities and antitrust practice. No jurisdiction is stated, no firm size band beyond the segment, no matter type identified as unsuitable, and nothing on languages. Any solutions or industries pages the site publishes could not be fetched on 5 September 2026 and are the rebuttal route.

Coverage is documented along two axes with real substance, and no boundary is stated. By role, which is an unusual and useful cut: partners and lead counsel, legal innovation teams, associates, and paralegals and litigation support, each with a dedicated page and its own stated benefits. By solution: Cases, Hearings, Transcripts, Arbitral Institutions, and Chambers sold as LEX, which together identify the buyer types as law firms, barristers' chambers and arbitral institutions. Practice focus is clear from the material rather than asserted, covering commercial disputes, international arbitration, public inquiries and inquests, with named inquiry work and named arbitral institution deployments, and the chambers product reaching criminal clerking. Geographic reach is real, with three regional operations and a published list of six data protection regimes across those markets. What is absent is any limit: the vendor states it modernises the litigation and arbitration lifecycle in any jurisdiction worldwide, and nothing identifies a practice area, matter type, jurisdiction or scale that the platform does not serve. Checked the navigation, the home page, the AI page and the arbitral institutions page on 31 Aug 2026.