Clio vs CosmoLex: how they compare in 2026

An eighteen-year-old practice platform with an AI workspace built on top, and the vendor frames it that way itself: the marketing line is that the 18-year foundation now powers the legal AI. Clio Manage, Grow, Draft and Accounting are conventional software, and the entry tier is sold on practice management rather than models. Clio Work is genuinely AI-native and is sold as a separate product with its own trial, and Grow AI and Manage AI sit inside the older modules. Remove the models and the system 400,000 professionals run their firms on still works. B on the band. Pages read 1 September 2026.

A mature conventional platform carrying a recent and narrow AI layer. The substrate is a complete product on its own: matter management, calendaring and tasks, document storage, time capture, billing, a client portal, and the built-in trust and general accounting that is the reason most customers give for choosing it. None of that requires a model, and all of it predates the AI by years. The machine learning arrived in August 2026 as four named features: AI Intake with an accompanying form builder, document summarisation extracting key terms, obligations, deadlines and action items, and natural-language filtering across the customer's own records. Those are real and shipped rather than announced, so the membership bar is cleared comfortably, and the vendor's framing is honest about what they are, describing them as extending an existing philosophy to the front end and to everyday administrative work rather than as a reinvention. Strip the models out and a working practice management and accounting system remains, which is what places this in the middle of the band rather than lower. Verified 2 September 2026.

Grounding is real and doubly sourced, and no accuracy figure was located. The pricing page states the AI is grounded in the firm's matters and cites its sources; the home page describes it working from matters, clients, filings, communications and financial data combined with over a billion legal documents across 100-plus countries, which is the vLex library Clio acquired in June 2025. That is a described retrieval method over identifiable corpora, and the subprocessor list corroborates the architecture by naming a vector database provider. What holds it at B is the absence of published measurement, and the Terms cut against any implied one: clause 14.2 states Themis does not review Output for accuracy or completeness and that Output may be incomplete or inaccurate, and clause 17.2 says the same of the legal Materials, disclaiming accuracy, completeness and currentness. Clio publishes an accuracy resource at /resources/ai-for-lawyers/ai-accuracy-legal/ which was not opened and is the rebuttal route toward A.

Nothing addresses accuracy, and the feature where it matters most is a deadline extractor. Checked the home page, the practice management, document management, security, pricing and subscription agreement pages in full on 2 September 2026. No accuracy figure, benchmark or evaluation is published for any AI feature, nothing describes how a summary is grounded in the underlying document or how a user traces an extracted item back to the page it came from, and hallucination is not mentioned anywhere by name or description. The AI material is entirely about data handling rather than output quality: the published questions and answers cover encryption, training, processing location and matter isolation, and none asks whether the output is right. That gap is consequential on this product specifically, because the summarisation feature is marketed as extracting deadlines and action items from pleadings, correspondence and discovery, and a missed or misread deadline is the classic malpractice event. The 2019 subscription agreement disclaims warranties of data accuracy in terms. Verified 2 September 2026.

A blanket review commitment stated in marketing and made binding in the contract, without the finer structure the top band asks for. The pricing page says the firm reviews and approves everything before it reaches a client or the court, which is the right commitment for the highest-stakes outputs. Terms clause 14.2 converts that into an obligation running the other way: the subscriber agrees to use AI Services and Output only with human oversight, and is responsible for reviewing Output for accuracy, completeness, appropriateness and compliance with legal, regulatory and professional requirements before disclosing or using it. Clause 14.4 leaves the subscriber in full control of how AI Services are implemented, configured and presented. B rather than A because no modes are described, no thresholds are published, and nothing sets out what the product does on its own versus what it waits to be asked.

No oversight model is published for any AI feature. Checked the home page, the practice management, document management, security and pricing pages and the subscription agreement in full on 2 September 2026. Nothing states what a lawyer must review before relying on a document summary, whether extracted deadlines are written to the calendar automatically or proposed for confirmation, what happens to an item the system is unsure about, or where a person sits in the intake flow. The vendor's own launch language runs toward autonomy rather than away from it, describing AI Intake as capturing client data and moving every intake forward without anyone touching it, which is presented as the benefit. The subscription agreement predates the AI features by seven years and contains no review obligation, no acceptable use provision covering generated output and no allocation of responsibility for acting on it. The one adjacent statement located is a caution rather than a control, advising firms to review their own professional responsibility obligations in their jurisdiction. Verified 2 September 2026.

One named firm with a figure and a period, against portfolio numbers with no method. King Law is named on the home page with a 275% increase in revenue over four years, attributed to visibility and tooling across the firm. Named individuals appear with their firms: Angela Lennon at Koenig | Dunne, Danielle Harvey-Jacob at Harvey-Jacob Law. Platform figures are 400,000-plus legal professionals, 130-plus countries and 300-plus integrations. Nothing is dated, no method is described for the King Law figure, and the 4.7 from 12,000-plus reviews is directory material excluded under the ground rules along with the vendor's own comparison page. B: real deployment evidence with substance, short of the dating and method the top band requires.

Many customer voices, almost no attributable outcomes. Roughly seventeen individuals are quoted across the pages read, most with a full name and a state or province, spanning Virginia, Louisiana, Pennsylvania, Utah, Ontario, Alberta, British Columbia and Quebec, and one carries an employer, a paralegal at Ramos Law. That breadth is real but the quotes are experiential rather than measured, and only one offers anything numeric, a user saying they spend about a third of the time they did before. Scale is asserted without attribution as thousands of small to mid-sized firms across North America. One third-party endorsement is substantive and worth recording: the company states the Canadian Bar Association selected CosmoLex as its exclusive preferred practice management software. The headline efficiency claim is candidly labelled rather than passed off, with 18 or more hours returned each week and a 15.3 times return footnoted as modelled from third-party research for a five to ten attorney firm with results varying. Honest labelling, but a model is not a deployment result. Verified 2 September 2026.

Substantive on every limb but the one the band names last. Terms clause 4.2 binds Themis and its third-party vendors and hosting partners to hold Content in strict confidence and not use or disclose it outside the agreement. Clause 21.3 goes further than most: only Themis, with strict business reasons, may access and transfer Content, and only to provide the Service, with reasonable efforts to notify the subscriber before doing so. Clause 14.6 states use of AI Services grants no right to use confidential information to train generalised LLMs. The security page states AI data is encrypted and processed in the customer's region, that AI tools process in real time without storing or reusing, and that outputs are generated only for the authorised user requesting them. What is missing for A is express treatment of privilege and work product, absent for a platform holding the entire matter file, and any documented segregation between matters within a firm.

Four limbs are addressed and one of them is answered more directly than by almost any record in this corpus. On segregation, the vendor states that AI summaries and outputs are scoped to individual matters and that the system working on one matter cannot surface or reference information from another, and then says why: the isolation is designed specifically to preserve attorney-client privilege. Naming privilege as the design objective, rather than leaving a buyer to infer it from a confidentiality claim, is what this limb asks for. Underneath it the platform's own controls are described, with matter-level permissions and privilege settings, predefined roles including Administrator, Billing Only and Matter Owner, and matter owner restrictions limiting who can see documents. On training, the AI material states that documents are never used to train external AI models. Section 9 of the subscription agreement gives a mutual confidentiality regime in which the subscriber's Confidential Information expressly includes its client information, and section 9.5 removes personnel and subcontractor breaches from the liability limitation. Two gaps hold it below the top band: no model provider is identified anywhere, and no retention period is stated for anything during the term. Verified 2 September 2026.

Every limb is covered and all of it is contractual. What it is not is stated at the top of the Terms in bold: Themis is not a law firm and does not provide legal advice, through the Service or otherwise, with clause 12.1 adding that provision of the Services, Output or Materials creates no attorney-client relationship. Who may use it is scoped: the Service is intended only for use by legal professionals and those working under their supervision, with a named exception for academic access participants. Competence and supervision are addressed directly in clause 14.2, which requires human oversight and makes the subscriber responsible for reviewing Output against professional organisation requirements and applicable fiduciary rules. Jurisdiction limits are named three ways: the Service is intended for North America with separate EMEA terms, Materials are limited to the country associated with the account, and clause 2.16 warns that Themis runs one code-base for all jurisdictions and the subscriber must configure and verify settings for its own. The client-facing limb is met by clause 14.4, which requires the firm to give its end-users disclosures covering the use of AI, the nature and limitations of AI-generated responses, and any human oversight processes.

One general pointer exists and nothing else. Checked the home page, the practice management, document management, security and pricing pages, and the subscription agreement in full including both exhibits, on 2 September 2026. There is no statement that output is not legal advice, no account of what a user must verify, and no jurisdictional statement. Applying the rule on inapplicable limbs, the unauthorised practice question largely does not bite here, since practice management and accounting software produces no legal advice, and its absence is not counted against the record. The professional responsibility limb does bite, in two specific ways the vendor itself writes about at length in its own blog: trust accounting is governed by detailed rules whose breach draws bar discipline, and the AI now extracts deadlines from litigation documents. Neither is connected to any published guidance. The single located statement is in the AI questions and answers, advising that as with any technology firms should review their own professional responsibility obligations in their jurisdiction, which points at the duty without naming a rule, an authority or a jurisdiction. Verified 2 September 2026.

Principles are published, the mechanism behind them is not. A dedicated AI Principles page sets out five commitments, attributed to Chief Technology Officer Jonathan Watson, covering AI as collaborator, transparency of AI actions, high-quality data, relentless refinement and uncompromising security. Bias appears once, as a claim that diverse top-tier data minimises it, rather than as anything measured or found. Nothing names an individual or committee accountable for the AI, describes what is tested before a feature ships, or discloses a single evaluation result. The one auditable trace of a testing function is indirect and sits elsewhere: the subprocessor list names Braintrust Data as an AI analytics, evaluation and testing provider across five products. That is evidence a pipeline exists, not a published regime a buyer could examine. C on the band's own words.

No governance material was located on any surface. Checked the home page, the practice management, document management, security and pricing pages and the subscription agreement in full on 2 September 2026. There is no responsible AI page, no framework or set of principles, no individual or function named as accountable for model behaviour, no account of what is evaluated before an AI feature or model change ships, and no certification such as ISO 42001. A dedicated AI practice management page exists in the site footer and was not opened on this pass, which is recorded as a limit on this reading. What the published AI material does cover is entirely data handling, addressing encryption, training, processing location and matter isolation, and none of it concerns how the models behave or how that behaviour is checked. Nothing anywhere addresses uneven output across document types, practice areas or drafting styles, which is the live question for a summariser that will meet everything from a two-page retainer to a discovery production. Verified 2 September 2026.

One limb short of a full set, and the limbs it clears are clearer than anything else in this pull. Incident practice is contractual and specific: clause 5.5 requires Themis to report any event it reasonably believes represents unauthorised access to, disclosure of, use of or damage to Content within 72 hours, with clause 5.6 setting out cooperation, investigation and mitigation duties. Deletion is stated with a period: 90 days to retrieve Content after termination, after which all Content is irrevocably deleted, with escrowed data held six months under a separate escrow agreement the customer controls. Subprocessors are published in full and dated. Access control is covered by clause 21.3, limiting Themis access to strict business reasons with advance notice. Encryption, geo-redundancy and at least annual third-party penetration testing are stated. The gap is retention during the term: no period is published for AI inputs and outputs, and clause 14.6 permits de-identified aggregated Content and Output to be kept for quality improvement without limit.

The security page carries real specifics and the agreement backs several of them. Encryption is stated on both sides, with 256-bit SSL in transit and encryption at rest, and backups run automatically every few hours, stored encrypted across multiple locations. Access control is described at two levels, infrastructure controls plus assignable security levels, roles and access privileges at user level, with two-factor authentication by SMS. Incident practice is contractual rather than aspirational: section 8 of the subscription agreement commits the vendor to promptly report any unauthorised access upon discovery and to use diligent efforts to remedy the breach, while making clear that the cost of notifying affected individuals falls on the firm. Deletion is bounded by section 13.3, giving thirty days after termination to retrieve everything before the vendor may irrevocably delete it. Data centres are stated to be in North America, in the United States or Canada depending on the firm's location, and core support personnel are in the United States. Two elements are missing: no subprocessor or vendor is named anywhere, and no retention period is stated for anything during the term. Verified 2 September 2026.

A real and unusually two-sided position that stops short of the output itself. Clause 11.1 excludes liability generally but carves out breaches of confidentiality, security and managed backup, so those obligations remain live. Clause 11.2 caps liability at the total paid in the six months before the claim arose, and expressly disapplies that cap to the indemnity. Clause 13.2 gives the subscriber a genuine indemnity on two grounds: third-party intellectual property claims against the Service, and, more unusually, claims arising from a violation by Themis of its own confidentiality or security obligations. Carve-outs are itemised in 13.3. What keeps this at B is that none of it reaches AI output. Clause 13.2(a) expressly excludes Output from the IP indemnity, clause 14.2 provides Output as is with no representations, and clause 12.2 disclaims any warranty as to results. No insurance position is published.

A complete and reasonably balanced allocation is published, in an agreement that predates every AI feature by seven years. Section 20 caps aggregate liability at twelve months of fees but carves out the vendor's indemnity and its confidentiality obligations entirely, and section 19 excludes consequential damages with the same two carve-outs, so confidentiality exposure is uncapped rather than merely enhanced. Section 17 gives the customer an intellectual property indemnity with a defined remedy ladder ending in termination and a refund of prepaid fees plus return of data in a database format. Section 15 is a real warranty rather than a bare disclaimer, covering reasonable efforts on uptime, freedom from material defects and conformity to the descriptions published on the site, with prompt correction as the remedy. Section 9.5 removes breaches caused by the vendor's own personnel or subcontractors from the security liability limitation. The limitation worth naming is currency rather than content: the agreement carries an effective date of 19 June 2019, names CosmoLex Cloud, LLC as the contracting party against a ProfitSolv Purchaser, Inc. copyright line, and says nothing whatever about artificial intelligence, generated output or model providers. Verified 2 September 2026.

Breadth is stated and documented, depth is not. Clio publishes 300-plus integrations through an app directory, runs a developer hub and partner programme, and the Terms treat API access as a governed feature in clause 3.9, setting out that API use is bound by the agreement, that excessive use may lead to suspension with a reasonable attempt to warn first, and that access may be modified or discontinued. Third-party services are addressed squarely in clause 18, which states they are not Services under the agreement, carry no warranties, indemnities or service commitments from Themis, and may be replaced, disabled or restricted at any time without notice. That is honest about where the boundary sits. B rather than A because no page read on 1 September 2026 describes what any specific integration moves, in which direction, or what a firm must configure; the app directory was not opened and is the rebuttal route.

The named systems are numerous and include the ones that matter for this product class. Document storage integrates with NetDocuments, Box, Dropbox, OneDrive and Google Drive, which means a firm with an established legal document management system is not forced to abandon it. Microsoft 365 and Outlook are named, with LexShare and LexSign operating inside Outlook for secure sending and signature collection, and QuickBooks Online and Zapier appear on the pricing page as standard inclusions. Exhibit B of the subscription agreement names further third parties in the e-signature path, Groupdocs Signature and HelloSign, alongside Appointlet, Adobe and DocuSign as separately licensed. An application programming interface is expressly authorised at section 7.5 with a published quota of 1,000 calls per user licence per month, which is more concrete than most vendors state. What keeps this below the top band is documentation: no connector list or developer reference was located, nothing describes what data flows in which direction for any named system, and the depth of each integration is left unstated. Verified 2 September 2026.

The most detailed residency disclosure located in this pull, held off the top band by one absence. Regions are published and real: hosting options in the European Union, Australia, the United States and Canada, five regional site variants, a separate EMEA instance at eu.app.clio.com with its own terms, and separate North American and EMEA agreements. Processing is distinguished from storage rather than conflated: the security page states all data used by the AI is encrypted and processed in the customer's own region, naming US, Canada, EMEA and APAC, and the subprocessor list gives a data location for every provider individually, including each AI processor. Infrastructure is named by provider and product. What is absent is the tenancy model, which is stated nowhere, and any statement of whether residency options vary by plan, which is what the top band asks for alongside the regions.

Residency is addressed on two pages that do not agree with each other. The security page states flatly that all data is stored in United States data centres and that core support teams are in the United States and subject to United States jurisdiction. The trust accounting page states that data is stored in North America and that, depending on the location of the firm, it may be held in a data centre in either the United States or Canada. The second is the more informative statement and is consistent with the separate Canadian site the company operates, but a reader cannot tell from the pages which governs, and no page sets out the placement rule in one place. Beyond that the picture is thin: no cloud provider is named, with Exhibit B referring only to a United States-based cloud provider, no tenancy or isolation model is described at infrastructure level, and no region choice is offered to the customer. One AI-specific statement is useful and is credited here, that AI processing occurs within the customer's CosmoLex environment rather than being sent to an outside service. Verified 2 September 2026.

The only attestation in this pull that states its coverage period. The trust centre at trust.clio.com announces that the 2025 SOC 2 Type II report has been issued for the period 1 June 2024 to 31 May 2025, and the security page states Clio completes annual SOC 2 Type II and SOC 1 Type II examinations with both reports available through that trust centre. The route to the report is not only a portal but a contractual right: clause 5.4(d) obliges Themis to provide a SOC 2 or SOC 3 report, or a comparable description of security measures, within thirty days of a subscriber request. At least annual third-party penetration testing is stated, and Clio's own help centre names Deloitte as the independent auditor, albeit for the earlier Type I engagement. Weaknesses a buyer should still note: the trust services criteria are not enumerated, the auditor for the current Type II is not named, the penetration testing firm is described only as a leading cybersecurity firm, and clause 5.4(d) requires entering an agreement with the report's third-party provider.

A named standard, a stated cadence and a named assessor, with no artifact behind any of them. SOC 2 Type 2 certification is claimed on the security page and repeated on the pricing page as standard across all plans, and the cadence is stated rather than left vague, with the infrastructure examined and passing inspection every twelve months. More unusually, an independent security assessment is attributed to a named firm: a threat assessment of the code and architecture perimeter conducted by Phalanx Secure Solutions, Inc. Naming the assessor is rare in this corpus and is the strongest element here. What is absent is the evidence. No auditor is named for the SOC 2 work itself, no report period, issue date or scope statement is published, there is no trust centre or portal, and no request route exists by which a prospect could obtain any report. One gap deserves particular note because the vendor created it: section 8 of the subscription agreement commits the vendor to publish to its site, for subscriber review, any third-party audit reports regarding data security. No such report appears anywhere on the site. Verified 2 September 2026.

By a distance the best evidence on this axis in the pull, and still short of the top band on two limbs. The subprocessor list, last modified 20 August 2026, carries a dedicated AI Service Providers table naming five AI processors with the products each serves and the data location for each: AWS Bedrock, Anthropic, OpenAI, Google Vertex AI and Microsoft Azure Foundry, spanning Clio Manage, Grow, Draft, Work, Vincent AI and Clio Operate. It also names Turbopuffer as the vector database and Braintrust Data for AI analytics, evaluation and testing, and Vapi for the Grow voice agent. A buyer can therefore answer whose model sees their content, per product, and where it runs. Two things keep it at B. The specific models are never named, only the providers. And no change-notification commitment was located on the page; the Themis Data Protection Addendum, cited as the clause under which the list is published, was not opened and is the rebuttal route toward A.

Where processing happens is stated clearly and what does the processing is not identified at all. The vendor answers the location question directly and usefully: all AI processing occurs within the customer's secure CosmoLex environment, and data does not leave that environment to be processed by an outside service. Read alongside the statement that documents are never used to train external AI models, that implies no third-party model provider sits in the path for these features, which is more than most records here disclose about inference location. But no model is named, no version, no provider entity, and nothing states whether the underlying model is built in-house, licensed and self-hosted, or reached through a cloud provider's own service, which are three materially different answers to the question a firm's client would ask. No subprocessor list exists in the subscription agreement or anywhere else, and no notification commitment covers a change in the model set. A dedicated AI practice management page exists and was not opened on this pass. Verified 2 September 2026.

Part of the range is priced and the rest is a conversation. The pricing page publishes a starting figure of $49 USD per user per month and names four tiers, Starter, Core, Signature and Elite, with the feature split described tier by tier and a bundle saving of over 15% advertised. Above the entry tier a buyer is routed to a custom quote, and large firms to a bespoke process. The Terms make clear that the published number is not the whole cost: clause 9.5 refers to setup fees, implementation charges and Metered Features fees as separate non-refundable items, and clause 9.6 gives purchasers of setup or professional services a sixty-day window to start them or lose them. None of those amounts is published. B on the band's first limb, real pricing published for part of the range with the upper tiers withheld. Note the tier names have changed generation again and third-party trackers still report the previous set.

Unusually detailed unit pricing is published in the agreement, and the subscription figures could not be read. Exhibit B of the subscription agreement is a genuine published rate card covering usage-based charges: local call tracking at three cents a minute and toll-free at five, SMS at three cents and MMS at four, additional local tracking numbers at three dollars a month and toll-free at five, task routing at five cents a route, e-signature at a dollar a request, an allowance of 2,500 emails per user licence per month with overage at two cents, a 1,000 call per user monthly API quota and a fifty dollar minimum prepaid balance for voice and messaging. Very few vendors publish that layer at all. The trial terms are equally specific, at ten days with no credit card, full functionality except a single bank feed connection, and the CRM and signature add-ons excluded. What could not be established is the core subscription price: the pricing page is headed as plans and pricing and lists what every plan includes plus four named add-ons, but the plan tier cards did not render on this pass, which is recorded as a retrieval limit rather than an absence. Verified 2 September 2026.

The widest published coverage in the pull, with the limits drawn jurisdictionally rather than functionally. Sixteen practice areas carry their own pages, from bankruptcy and criminal to intellectual property and real estate, and the segmentation runs across four firm sizes from solo to enterprise and Big Law, plus seven role-specific pages. Both buyer types the band asks about are addressed by name: in-house counsel and government law each have a page. Real limits are stated, though they concern where rather than what: the Terms scope the Service to North America with separate EMEA terms, restrict Materials to the country associated with the account, and warn in clause 2.16 that a single code-base serves all jurisdictions so the subscriber must configure and verify settings itself. B rather than A because the in-house, government and firm-size pages were not opened on 1 September 2026, and nothing states which practice types the product handles poorly.

The buyer segment is stated consistently and little else is bounded. Small and mid-sized law firms are named repeatedly as the intended customer, and the product is differentiated within that band, with material addressing solo practitioners moving from larger firms alongside multi-user support, role-based access and advanced financial reporting for mid-sized teams and higher case volumes. Geographic coverage is real and evidenced rather than asserted, with separate United States and Canadian offerings, data residency in either country depending on firm location, an exclusive preferred software designation from the Canadian Bar Association, and quoted users across four Canadian provinces and four American states. What is missing is practice and boundary detail. A practice type section exists in the site navigation and was not opened on this pass, recorded as a limit on this reading; no practice area is named on any page read, no firm size band is given in numbers, and nothing states what the product does not cover, including whether the trust accounting rules it enforces are configured per jurisdiction. Verified 2 September 2026.

The agreement permits a qualified form of training and the marketing says the opposite. Terms clause 14.6 lets Themis de-identify and aggregate the Content submitted to, and Output received from, AI Services and use it to improve and ensure the quality of those services, and clause 2.15 adds perpetual aggregate anonymised reporting on usage and content trends. The same clause withholds any right to use confidential information to train generalised LLMs, so the permission is bounded rather than open. The qualifier that matters to a buyer is that it runs on de-identified and aggregated material, not identifiable client content. Against that, the pricing FAQ states a firm's data is never used for AI training or any other external purpose. Where marketing and the agreement disagree the agreement governs, and it is recorded here.

The commitment sits on a product page rather than in the agreement, and it carries a qualifier a reader should notice. The published answer to whether data trains AI models gives the quoted statement and adds that CosmoLex AI operates on the customer's data to serve the firm, not to improve models for third parties. Both halves are directed outward: external models, and third parties. Nothing states whether customer content may be used to improve CosmoLex's own models, and the distinction is the same one this index has recorded on other records where a training denial is narrowed by a single word. The agreement was checked before this value was written. The subscription agreement dates from 19 June 2019, predates every AI feature and contains no training clause in either direction. The nearest provision, section 10.2, grants a perpetual and irrevocable licence to use, reproduce, modify and create derivative works for purposes of marketing, promoting or improving the services, but Content is defined at section 1.4 as files and data posted to public areas of the site such as public forums, so that licence does not reach matter material.

End-of-life is specific, the live term is not. Clause 10.7 gives the subscriber no less than ninety days after cancellation or termination to retrieve Content, after which all Content is irrevocably deleted from the Service, with escrowed data held a further six months under a separate escrow agreement the customer arranges directly. What is not published is any period for AI inputs and outputs while the subscription is running, and clause 14.6 permits de-identified aggregated Content and Output to be retained for quality improvement with no stated limit. The security page states AI tools process data in real time and do not store or reuse it, which points toward minimal model-layer retention but is a marketing statement rather than a term.

The exit window is fixed and contractual; everything during the term is unstated. Section 13.3 of the subscription agreement gives the quoted period, in the form and format provided by the software, and then reserves the right to irrevocably delete all subscriber Confidential Information and data beginning thirty days after termination or cancellation. One asymmetry belongs on the record because it can catch a firm out: that thirty-day window applies where CosmoLex terminates or declines renewal, whereas if the subscriber cancels, the agreement makes retrieval its sole responsibility before cancelling. Section 7.7 separately commits to redundant storage and to recovering lost or corrupted information without charge. What is absent is any period during the term, and nothing addresses AI material specifically: no statement covers how long uploaded documents, generated summaries, extracted deadlines or intake records are retained while the subscription runs, and no retention setting is described.

Per-user scoping is asserted without published enforcement detail. The security page states sensitive client information never leaves Clio's secure environment and that outputs are generated only for the authorised user requesting them. The Terms establish an administrator role able to designate users and deactivate access, but describe no matter-level partitioning, no conflict wall mechanism, and nothing about how retrieval is bounded when the AI reaches across a firm's matters, clients, filings and communications. Worth noting for comparison that the separately indexed Vincent AI record describes inheriting a firm's existing permissions and ethical walls when run inside Clio Operate; nothing equivalent was located for Clio's own products on any page read on 1 September 2026.

The AI is scoped to the boundaries the platform already enforces, and the vendor states the purpose rather than leaving it to inference. Asked whether the AI has access to matters it should not, the answer is no, with outputs scoped to individual matters and the quoted consequence, described as matter-level isolation designed specifically to preserve attorney-client privilege. The underlying access structure is the platform's own: matter-level permissions and privilege settings assign access by team member, predefined roles include Administrator, Billing Only and Matter Owner, and matter owner restrictions limit who can view or manage documents on a given matter. Because CosmoLex is itself the document system rather than a layer over a separate one, the AI inherits that permission structure rather than implementing a parallel one, which is what this value describes. No separate model per customer or per matter is claimed and none should be read in, and no separation between customers is described anywhere.

Notice is committed in the agreement and the process is published separately. Clause 7.2 provides that where Themis is legally required to disclose confidential information in a way the agreement otherwise prohibits, it will give the subscriber prompt written notice, to the extent permitted by law, before disclosing, so the subscriber may seek a protective order or other relief, and will then furnish only the portion legally compelled. Clause 7.1 states the primary duty is to protect Content to the extent the law allows. Clause 7.3 adds something few vendors publish: Themis will only accept legal requests for production of Content through the procedures set out at clio.com/legal-service, a dedicated public page for service of legal process. No transparency report was located on 1 September 2026, which is what keeps this below the top tier.

This is among the strongest formulations of the commitment in the corpus, and it has three parts rather than one. Section 9.4 of the subscription agreement addresses requirements from law enforcement authorities, the investigative process of a criminal or civil matter, or any applicable law, regulation, legal process or enforceable governmental request such as a subpoena or court order. To the extent permitted by law, the vendor gives the quoted notice, which is notice before disclosure rather than after it and is expressly tied to the customer's opportunity to seek protective relief. It then adds a minimisation undertaking, disclosing only that portion of the requested Confidential Information which is strictly compelled or otherwise required. The definition of Confidential Information at section 9.1 expressly includes the subscriber's client information and non-public personal information, so the notice right attaches to matter material rather than only to commercial terms. No transparency report was located, and the clause was drafted in 2017 per the agreement's own modification history.

The corpus is identifiable and the rights basis is described only in the abstract. Clause 17.2 states Themis may make available primary and secondary legal materials such as case law, legislation and articles, described as licensed or publicly available legal content, with additional third-party terms possibly applying. The source behind it is traceable through Clio's own disclosures rather than a provenance statement: the subprocessor list names Vlex Library and Clio Library as products and lists vLex entities and Fastcase, Inc. among Clio affiliates, and the home page puts the corpus at over a billion legal documents across 100-plus countries. What is absent is which licence covers which jurisdiction, any named publisher, and any update cadence. Clause 17.2 runs the other way on currency, expressly disclaiming that Materials are reviewed for accuracy, completeness or currentness, and clause 17.5 reserves the right to add, modify or remove Materials at any time without notice.

Checked the home page, the practice management, document management, security and pricing pages and the subscription agreement in full on 2 September 2026. CosmoLex ships no corpus and none is claimed. The AI features operate on material the firm already holds in the platform, its own uploaded documents, matter records and intake submissions, so there is no licensed database, published collection or third-party content behind any output and the provenance risks this signal tracks do not arise in their usual form. The honest record is that the question is not addressed rather than that a corpus was withheld. Nothing published describes the training data behind the underlying model either, which is a different question and one the model supply chain axis carries.

No located public material describes a citator or subsequent-history check in Clio's own products. The Terms cut the other way: clause 17.2 states Themis does not review the Materials for accuracy, completeness or currentness and provides them as is, which is the opposite of a good-law guarantee, and clause 17.5 permits Materials to be added, modified or removed without notice. Checked across the home page, pricing page, AI Principles page, security page, subprocessor list and the full North American Terms of Service on 1 September 2026. Note for the reader that the separately indexed Vincent AI record does describe a Cert citator; that capability is graded on that record and is not credited here.

Checked the home page, the practice management, document management, security and pricing pages and the subscription agreement on 2 September 2026. The product retrieves no legal authority and cites none, so this signal has nothing to operate on and its absence is not a criticism. The nearest analogue is worth recording because it is live: the summarisation feature extracts deadlines from pleadings, correspondence and discovery, and a deadline is only correct against the rules and calendar in force in that court at that time. Nothing published describes whether extracted dates are validated against court rules, jurisdiction-specific computation, or the firm's own docketing conventions, or whether a previously extracted date is revisited if the underlying schedule changes.

No located public material describes what the product does when it cannot ground an answer. The published position places the burden entirely on the reader instead: clause 14.2 states Themis does not review Output for accuracy or completeness, that Output may be incomplete or inaccurate, and that the subscriber must review it before disclosing or using it. No confidence signal, no abstention path and no described no-answer behaviour was found across the home page, pricing page, AI Principles page, security page or the Terms on 1 September 2026. The Clio Work product page and the published accuracy resource were not opened and are the rebuttal route.

Checked the home page, the practice management, document management, security and pricing pages and the subscription agreement on 2 September 2026. Nothing describes what the system does when it cannot produce a reliable result, and no confidence indicator is described as shown to the user for any AI output. The published AI material is entirely about data handling rather than behaviour, covering encryption, training, processing location and matter isolation. Nothing states what a summary looks like when a document is poorly scanned, handwritten or in an unexpected format, whether an extracted deadline the system is unsure of is flagged differently from one it is confident about, or whether the intake flow escalates a submission it cannot classify. A dedicated AI practice management page exists and was not opened on this pass.

No court order, opinion or disciplinary record naming any Clio product has been located. The AI Hallucination Cases database maintained by Damien Charlotin was searched on 1 September 2026 across Clio, Clio Work and Clio Duo alongside general sanctions coverage, and nothing naming a Clio product as the tool involved was found. Clio itself publishes extensive material on the sanctions record, including analysis of the Morgan and Morgan matter, where the tool involved was that firm's own in-house platform rather than a Clio product. This is a statement about the public record rather than a finding about the product.

Searched the AI Hallucination Cases database maintained by Damien Charlotin, and reporting drawing on it, on 2 September 2026 on the product and corporate names CosmoLex, CosmoLex Cloud, LLC and ProfitSolv, together with a separate search for security incidents, regulatory action or disciplinary matters connected to the product. No court order, opinion, disciplinary record or enforcement action naming it was located. This is a statement about the public record rather than a finding about the product. The signal fits this product class poorly: the platform produces summaries, intake records and accounting entries rather than legal citations, so its characteristic failures would be a wrongly extracted deadline or a trust accounting error. The second of those does draw bar discipline, but against the firm rather than the vendor, and it surfaces in state disciplinary records rather than in any tracker comparable to the hallucination database.

Professional responsibility is engaged repeatedly and no specific guidance is named. Clause 14.2 makes the subscriber responsible for reviewing Output to ensure compliance with legal, regulatory and professional organization requirements and applicable fiduciary rules, and clause 14.4 enumerates the disclosures a firm must give its own end-users about AI use, limitations and human oversight, which is closer to operationalising a professional obligation than most vendors get. Clio also publishes a substantial body of writing on sanctions, the duty of competence and verification practice. But ABA Formal Opinion 512 is not named on any surface read on 1 September 2026, and no state bar opinion is cited; the 100-plus bar association partnerships are commercial relationships rather than engagement with guidance.

Professional obligations are referred to without any authority being named. The published AI material closes its confidentiality answer by saying that, as with any technology, firms should review their own professional responsibility obligations in their jurisdiction. That is a real pointer, directed at the buyer's duties rather than the vendor's, and it is why the bottom value does not fit. But it names no rule, no ethics opinion, no bar association and no jurisdiction, and it appears once. The gap is more visible here than on most records because the company writes about professional conduct constantly in its own marketing content, publishing material on trust accounting mistakes, common ethics violations and the American Bar Association's expectations on data security, none of which is connected to the product's AI features or carried into any product surface. Nothing addresses the guidance a lawyer should follow when relying on a machine-extracted deadline or a generated intake record.

Efficiency and revenue claims are published and the treatment of AI-assisted work on a bill is not. King Law's 275% revenue growth is the headline, and the platform is itself a billing and time-capture system. One clause does address passing a technology cost to a client, and it is worth noting for precision rather than credit: clause 17.3 states the cost of Purchased Materials may be recovered from a client as a disbursement or similar. That is research cost recovery, not a position on how time compressed by AI should be recorded or disclosed, which is the question this signal asks. No audit record identifying which work product was AI-assisted is described. Checked 1 September 2026.

Time savings are quantified and labelled honestly, and nothing addresses the bill. The pricing page states that CosmoLex returns 18 or more hours a week to a firm, broken into five components covering billing and invoicing, payment chasing, accounting and operations admin, matter and document admin, and automated tracking, alongside a 15.3 times return figure. All of it is footnoted as modelled from third-party research for a five to ten attorney firm with results varying, which is more candid labelling than most vendors offer. What is absent is anything connecting that compression to what a client is charged. Nothing addresses whether time captured or recovered through automation is identified as such, whether a summary generated in minutes rather than read over an hour changes what appears on an invoice, or how a firm would disclose AI-assisted work. No per matter record of AI-assisted work is described, and the platform's own time and billing module is not linked to the AI features in any published material.

A firm could answer a client's AI questionnaire from published material alone. The subprocessor list, last modified 20 August 2026 and reachable without a form or agreement, names every infrastructure, AI, feature and payment provider, maps each to the products it serves, and gives a data location for each, including a dedicated AI Service Providers table naming five model processors. Alongside it Clio publishes the Data Protection Addendum, the service level commitments exhibit, the AI Principles page, a public trust centre and a dedicated legal process page at clio.com/legal-service. Together that is a forwardable pack covering who processes client content, where, and on what contractual basis. The specific models are not named, which is the one thing a demanding client questionnaire might still ask for.

Checked the home page, the practice management, document management, security and pricing pages and the subscription agreement in full including both exhibits on 2 September 2026. No subprocessor list exists in any form and no model or AI provider is named anywhere, so the AI-specific limb this signal turns on is entirely unanswered. Third parties are named in other contexts, with Groupdocs Signature and HelloSign identified in the e-signature path at Exhibit B and integration partners named across the product pages, but none of those is presented as a processor of client content and none is an AI provider. A GDPR page and a data request route are published in the footer and neither was opened on this pass. What a firm could forward today is the subscription agreement, which is public and contains the confidentiality and security commitments, and the security page. Neither answers which entity processes client documents when an AI summary is generated.

The pieces of a record exist and none is assembled for disclosure. Output is stated to carry citations traceable to source, the subprocessor list would let a firm identify which providers could have processed a given product's content, and the review-and-approve commitment describes a human step. But nothing identifies which model produced a specific output, no verification event is recorded against a document, and no export is offered or described for the purpose of answering a court. Clause 14.4 pushes the disclosure obligation onto the firm, requiring it to tell its end-users about AI use and human oversight, without giving it an artifact to do so from. Checked 1 September 2026.

Checked the home page, the practice management, document management, security and pricing pages and the subscription agreement on 2 September 2026. Nothing addresses disclosure of AI use to a court, regulator or bar, and no record identifying which outputs a model produced is described as available or exportable. Two adjacent capabilities exist and neither is presented as serving this purpose. The platform maintains audit trails as part of its trust accounting and document controls, which is the material a bar auditor would examine, but nothing states that AI involvement is recorded within them. And section 13.3 provides a data export in the form and format the software provides, which is a portability right rather than an audit record. Nothing describes whether a summary a lawyer relied on, or a deadline the system extracted, is retained with an indication that it was machine-generated.