Avvoka vs Clarilis: how they compare in 2026

The models are the engine of a core capability and a complete platform sits underneath them that would work without any. Avvoka's own framing is precise about the split: its Harvey partnership announcement describes combining domain-specific legal AI with a deterministic drafting engine, and the features overview says the platform brings AI and automation together. The AI does real work, building templates from existing documents through AI-powered no-code automation, drafting inside firm rules, and driving MassDraft for high-volume exercises. Remove it and a firm still has questionnaire-driven document assembly, a clause library, live multi-party negotiation on a single controlled draft, clause locking, version history, e-signature and the negotiation analytics layer, which is the product's most distinctive feature and is a data property rather than a model output. That is the definition of a model layer on a working system rather than a product that disappears without it. Checked 4 September 2026.

The models are a feature layer on a product whose value plainly stands without them, and the vendor says so in its own words. Clarilis describes deterministic, template-based automation as remaining the gold standard for legal drafting, positions AI Draft as augmenting rather than replacing it, and quantifies the split: the automation takes a draft roughly ninety per cent of the way and AI Draft exists to assist with the remaining ten. The architecture confirms the framing rather than merely asserting it. AI Draft is confined so that it creates novel content without changing the logic-generated content, and the FAQ records that AI Draft can optionally be disabled in an automation for firms that do not permit generative AI, with the rest of the platform unaffected. What remains when it is switched off is the entire product: questionnaire-driven assembly of complete document suites, hundreds of maintained templates across four practice areas and five jurisdictions, house-style formatting and the managed service behind it. Graded on that evidence rather than on the quality of the vendor's AI disclosure, which is strong and belongs on other rows.

Grounding is real and described in architectural terms, and nothing about it is measured. The mechanism is published rather than asserted: drafting runs on a deterministic engine, generation is constrained to the firm's own approved templates and clause library, and the company describes the AI as running inside firm rules and keeping the firm's language, logic and standards consistent in every draft. The Harvey announcement makes the same point from the other side, a lawyer drafting from the firm's approved template and clauses so the firm's precision is carried into every draft. For a drafting product that constraint is the substantive answer to hallucination, and it is more than most in this lane publish. What is absent is any test of it. No accuracy figure, no evaluation, no test set, no error rate and no statement of failure modes appears anywhere, and there is no accuracy or benchmark page on the estate, while the marketing carries unmeasured superlatives including flawless, perfectly every time and full confidence in every draft. The limbs about linked primary sources and citation status do not apply to a drafting tool and are named here rather than counted against it.

Nothing is measured, and the vendor publishes its known failure modes with unusual candour. There is no accuracy figure, no evaluation, no test set and no error rate anywhere on the estate, and the ninety per cent figure that recurs throughout is a time-saving claim about the automation rather than an accuracy measure. The limbs about linked primary sources and citation status do not apply and are named rather than counted: AI Draft generates novel contractual language, so there is no authority to cite. What the vendor does publish instead is a specific list of what its own AI gets wrong, telling users that AI content must be reviewed for cross-references because the AI does not yet automatically match references to other clauses, for definitions used but not defined elsewhere, and for legal and commercial effectiveness, closing with the instruction to always treat AI-generated content as a first cut. Naming a concrete limitation of one's own product is rare in this corpus and it is why this sits at the top of the band rather than the bottom. Grounding is described in terms of matter context and codified practice-area knowledge from in-house lawyers, without any account a reader could test.

Controls are published with mechanisms rather than promises, short of the threshold at which the system acts alone. The strongest is clause locking: a firm decides in advance which clauses are open to negotiation, which are locked and what can be changed, so the boundary of permitted variation is set by the firm and enforced by the platform rather than reviewed after the fact. Around it sit routing of drafts to named people at defined points, tracking of every decision as the deal progresses, automatic saving of every version, and conversion of comments into assigned tasks with an owner. The company's own summary is that AI works inside firm rules with the customer in control end to end. What is missing is the account of the machine's own behaviour: nothing states what the AI does unattended, what happens when it produces something outside the firm's standards, whether generated text is marked as generated, or how a reviewer is alerted. Clause locking constrains the document; it does not describe the model.

All four limbs are published, and the controls are mechanisms rather than promises. What the system runs alone is bounded by design: AI Draft produces novel content only where a user invokes a named task, and it is confined so that it cannot change the logic-generated drafting the automation produced, so the boundary between machine and rules is architectural. How a lawyer checks it is built into the artifact rather than left to a dashboard, and this is the distinguishing feature: every AI-generated clause is highlighted in the delivered Word document, and any document containing AI content carries a notice at the top stating that the AI content requires review. The route back to human judgement is stated explicitly and with specificity, the vendor publishing that AI content is novel and has not been reviewed in advance, unlike automated content, and setting out what to check, being cross-references, use of definitions, and legal and commercial effectiveness, before concluding that AI-generated content is a first cut that still needs expert review. A firm that does not permit generative AI can disable the layer entirely and keep the automation. Marking machine-generated text inside the document that leaves the building is a stronger form of oversight than an in-platform review state, because it survives the export.

Attributed customer voices with no figures attached to any of them. Two testimonials carry a name, a role and an organisation: Karen Soan, a solicitor at CLS Holdings, on the platform adapting as requirements change, and Greg Snow of the same company, stating that Avvoka delivers measurable business value at scale. That is real attribution a reader can check, which is what separates this from a logo wall, and it is also the whole of it: both come from one customer, neither is dated, and the claim of measurable value carries no measurement. The only scale claim published is unattributed, that the product is trusted by one in five of the world's top firms and by corporate teams everywhere, with no basis given for the ranking or the count. No case study, no named deployment and no published figure for time saved, volume processed or negotiation outcome was located on the estate. Trade press names larger customers; it is third-party and was not used.

Named customers with attribution, and the figures live somewhere else. Two customers are quoted with organisations attached, Miri Stickland, Head of Knowledge at Forsters, on risk management benefits from updates being made quickly and consistently behind the scenes, and a Head of Knowledge Management at TLT, by role rather than name, saying the solution exceeded expectations on delivery. Both link to individual customer pages, and CMS is named in a first-party blog post about faster drafting for its Scottish real estate team. Separately the platform material publishes that the system reduces the time taken to produce suites of documents by around ninety per cent, and that firms draft suites ninety per cent faster, with no firm attached to either figure, no date and no method. So the named customers carry no measurement and the measurement carries no named customer, which is this band exactly. A customers section and a case studies section both exist in the navigation and neither was opened in this pass; they are named here as the limit and are the cheapest available upgrade on this record.

Substantive published security commitments, an unusually candid statement about privilege, and no agreement behind either. The security page is specific: documents encrypted at rest with AES-256, each with a unique initialisation vector and each key encrypted under a regularly rotated master key, connections including API access over HTTPS, passwords filtered from logs and hashed with BCrypt, two-factor authentication enforceable company-wide, SSO-only enforcement with automatic revocation for leavers, and ISO 27001 certification behind it. On privilege the vendor does something rare and states the position against itself: the legal advice disclaimer records that all information provided to Avvoka is not protected by legal advice privilege. That is an express treatment of the question, and it is a disclaimer rather than a protection, so a buyer knows where they stand. Three things hold the grade here. No position on training use appears anywhere. Nothing describes separation between customers or matter-level walls. And no customer agreement, master subscription agreement or data processing addendum is published, so every commitment above sits on a marketing surface rather than in a contract.

Substantive commitments published where a buyer can read them before signing, with two limbs missing. The AI-specific position is unusually complete: neither user inputs nor AI responses are used to train the models, processing runs through zero data retention endpoints, and data is processed in-region. Around it sits a detailed platform position, with encryption at rest and in transit, protected links time-limited and encrypted to 256-bit AES, all requests logged and verified including IP addresses, user-level authority checks restricting access to particular document suites and drafts, SAML 2.0 single sign-on, optional IP lockdown, separate development, testing and production environments, criminal record and eligibility checks on all staff, least-privilege access that is recorded and auditable, and no outsourcing of system development at all. Two things hold the grade. No privilege or work product treatment appears anywhere, on a platform holding transactional drafting for named clients. And no customer agreement, master services agreement or data processing addendum is published, so every commitment above sits on a product page rather than in a contract a firm could enforce.

The fullest advice-line statement located in this pull, on a dedicated page, and no engagement with the professional rules. The disclaimer states that Avvoka is not a firm of solicitors, does not practise law and does not give legal advice, that it is not the user's solicitor, is not a substitute for a solicitor's advice and has no solicitor-client relationship with the user, and that a user entering into a document created on the platform does so at their own risk and represents themselves. It then does what most disclaimers omit and enumerates what the product does not do: it does not review information for legal accuracy or sufficiency, does not draw legal conclusions or provide opinions about the documents a user selects or creates, and does not apply the law to the user's circumstances. It also states that information provided is not covered by legal advice privilege. What is absent is the professional layer: no Solicitors Regulation Authority guidance, bar rule or ethics opinion is named, and nothing addresses a firm's supervision or competence duties when generated drafting reaches a client. One scope point is recorded: the page opens by framing itself as terms in relation to use of the website, while its substance plainly addresses documents created on the platform.

A real but narrow position on where the product stops, published outside any agreement. The vendor states that AI Draft does not replace professional legal judgment, that AI-generated content should always be treated as a first cut that still needs expert review, and that automated content is carefully reviewed in advance while AI content is novel and has not been. That is a genuine statement about the relationship between the product and the lawyer's own responsibility, and it is more than the boilerplate this band usually describes. What is absent is the advice line itself. Nothing states that Clarilis is not a law firm or that its output does not constitute legal advice, no jurisdiction limit is drawn despite pre-automated content spanning England and Wales, Scotland, Ireland, the United States, Canada and Asia Pacific, and nothing addresses a firm's supervision or competence obligations. There is also no terms of service on the estate in which such a statement could sit, so the position rests entirely on a product FAQ and a blog post. Searched the AI Draft page, the security page and the full footer inventory on 5 September 2026.

No governance position of any kind was located. There is no responsible AI page, no principles statement, no accountable owner or function named, no pre-release testing regime, no AI management system and no certification such as ISO 42001. The company holds ISO 27001 and publishes its ISMS policy, which is an information security management system and a different subject; it is credited on the certification and stewardship rows rather than counted here. Nothing addresses uneven output, which on a drafting product would bear on how generated language performs across document types, jurisdictions and the languages a multinational user base drafts in. The site navigation and footer were inventoried in full on 4 September 2026 and the published policy set is the GDPR policy, the privacy policy, the legal advice disclaimer, the signing disclaimer and the security page.

A published commitment to an external framework, short of anything audited or owned. Clarilis states that it has signed the Litig AI Transparency Charter, links to it, and describes it as setting out commitments for organisations developing and providing legal AI solutions, promoting responsible innovation while safeguarding ethical and professional standards. That is a named, checkable external instrument rather than a self-authored principles page, which is why this sits above the floor. The security page adds a responsible AI section with four specific commitments: clear content indicators distinguishing AI from rules-based drafting, in-region processing through zero data retention endpoints, no use of data for model training, and adherence to third-party intellectual property rights. What is missing is everything the band above asks for. No management system such as ISO 42001 is claimed, no independent audit of AI governance is published, no accountable owner or function is named, no pre-release evaluation regime is described and no results are given, and nothing addresses uneven output across practice areas or jurisdictions. The ISO 27001 certification is information security and is credited on the certification row rather than here.

Most of the ground is covered in real detail, with retention the notable gap. Published: AES-256 encryption at rest with per-document initialisation vectors and a rotating master key, HTTPS for all connections including the API, BCrypt password hashing with passwords filtered from logs and definable rotation rules, enforceable two-factor authentication, SSO-only enforcement with automatic revocation when a user leaves, real-time replication of all production databases with a hot-standby failover arrangement, and encrypted backups. Testing and response are addressed rather than asserted: annual black-box penetration testing restricted to CREST-approved providers with the latest scorecard available to clients on request, a published vulnerability reporting process with a PGP key, a pipeline continuously tested for CSRF, XSS and SQL injection, and a breach procedure committing to identify compromised information and notify affected individuals and the bodies required by law. What is missing is a retention period, the GDPR policy saying only that data is retained when necessary and deleted on a data subject's request, and any subprocessor list beyond the named infrastructure providers.

Most of the ground is covered in real detail, with retention and incident practice the gaps. Published: encryption at rest and in transit, protected links time-limited and encrypted to 256-bit AES, servers behind a firewall with all access over encrypted communication, separate development, testing and production environments, all data stored on encrypted file systems and backed up daily at minimum, horizontal scaling with active resource monitoring, all requests logged and verified including IP addresses, user-level authority checks limiting access to particular document suites and drafts, SAML 2.0 single sign-on with optional IP lockdown and configurable lockout, pre-employment criminal record and eligibility checks on all staff, least-privilege access that is recorded and auditable, and no outsourcing of development. Hosting is named as AWS and Azure, and the AI layer runs zero data retention endpoints in-region. Regular penetration testing is stated. What is absent is a retention period, since the only statement is that the customer has complete control over platform storage and deletion of data without any period or process described, and any incident notification commitment, which appears nowhere on the readable estate and has no agreement to sit in.

Nothing published addresses who bears the loss when the product is wrong. The full navigation and footer were inventoried on 4 September 2026 and the published legal set comprises a GDPR policy, a privacy policy, a legal advice disclaimer, a signing disclaimer and a security page. There is no terms of service, no customer agreement, no master subscription agreement and no service level agreement anywhere on the estate, so no indemnity, no liability cap, no warranty position, no disclaimer of warranties and no insurance statement exists to read. The legal advice disclaimer allocates responsibility for the decision to enter into a document to the user, stating they do so at their own risk, but that is a statement about advice rather than about loss, and it does not address a defective template, a clause assembled incorrectly, a negotiation control that failed to lock, or a generated draft that departed from the firm's standards. The band above does not fit, because it requires a standard limitation clause disclaiming the exposure, and no such clause is published.

Nothing published addresses who bears the loss when the product is wrong. The footer was inventoried across two pages on 5 September 2026 and carries exactly two entries, a privacy policy and a sitemap. There is no terms of service, no master services agreement, no customer agreement and no data processing addendum anywhere on the estate, so no indemnity, liability cap, warranty position, disclaimer of warranties, service level commitment or insurance statement can be read before entering a sales process. The gap is more visible here than on most records because of what surrounds it: this vendor publishes a certificate number for its ISO certification and a specific list of what its own AI gets wrong, so the absence of any published commercial terms is a choice rather than an oversight. The exposure is real on a product that produces complete suites of transactional documents through a managed service, where the question of who carries a defect in a maintained template is a live commercial issue. The band above requires a standard limitation clause and there is no clause of any kind.

Named connections with the object of the connection described, short of implementer documentation. The most substantial is the Harvey partnership, announced first-party and specified rather than badged: a lawyer starting work in Harvey can draft from the firm's own approved Avvoka template with matter details in place, draw on the clause library the firm has approved, and carry the document through negotiation, with the announcement describing the combination as domain-specific legal AI over a deterministic drafting engine. Execution integrations are named for Docusign and Adobe Sign, with completion tracked inside the drafting flow. API access is confirmed on the security page, which describes API connections as encrypted over HTTPS. A dedicated extend page in the navigation promises connection to existing systems across teams and the wider business. What is absent is depth a buyer could plan against: no document management, matter management or CRM counterparty is named on any page read, and no field mapping, sync direction, trigger or developer documentation was located. The extend page was not opened in this pass and is named here as the limit.

No integration into the systems a firm runs was located, on a product whose entire output is documents that must then be filed somewhere. The platform is web-based and delivers finished drafts in Word in the firm's house style, which is a format and a delivery convention rather than a connection. Two named third parties appear and both are identity rather than practice systems: SAML 2.0 single sign-on for authentication, and Certivox M-PIN offered as an optional additional security element. No document management system is named anywhere, so nothing addresses iManage or NetDocuments, and no matter management, e-billing, CRM or e-signature counterparty appears. No API, developer documentation or integrations page exists in the navigation, which carries platform, AI Draft, security, managed service, partners and working globally. The absence is notable rather than neutral for a drafting product used at volume in large firms, where the return path into a matter file is the workflow question a buyer asks first. Searched the AI Draft page, the security page, the platform page and the full navigation on 5 September 2026.

The residency limb is answered with genuine specificity and the tenancy limb is not addressed at all. The security page states that servers are located in clients' core business regions, naming the United Kingdom, Europe, the United States and Australia, and identifies the datacentre providers as AWS, OVH and Azure, with physical access by Avvoka staff restricted. Naming both the regions on offer and the providers behind them is more than most records in this corpus publish, and it lets a buyer with a data residency requirement start a real conversation. Against it, nothing states whether the platform is single or multi-tenant, no dedicated or isolated instance is described at any tier, and no on-premises option is mentioned. One conflict is recorded rather than resolved, because a buyer will hit it: the GDPR policy states that the main application server is in the United Kingdom with the backup in France and that the application is hosted by OVH, which describes a narrower and older arrangement than the security page, and the same policy dates its most recent penetration test to Q3 2020 where the security page describes annual testing.

Residency is committed rather than described, and tenancy is not addressed. The AI Draft page states that data is hosted in the customer's region via Azure, and the security page makes the same commitment for the AI layer specifically, that data is processed in-region using zero data retention endpoints. That is a regional commitment tied to a named provider, which is more than most records in this lane offer, and it matters on a platform serving firms across England and Wales, Scotland, Ireland, the United States, Canada and Asia Pacific. Infrastructure is named for the platform as a whole, hosted on Amazon Web Services and Microsoft Azure. Against that, nothing states whether the platform is single or multi-tenant, no dedicated or isolated instance is described at any tier, and no on-premises or self-hosted option is mentioned. The nearest thing to a separation statement is a permission control, that user-level authority checks restrict access to particular document suites and drafts, which governs users within a customer rather than the boundary between customers.

A named standard with a stated history and a published artifact behind it, short of naming the certifier or the scope. Avvoka states that it is ISO 27001 certified, continuously certified since October 2017, and subject to annual external audits, and it publishes its executed ISMS policy as an ungated PDF a buyer can download without an account or a conversation. That combination, a period, a recurring audit cycle and a readable governing document, is materially more than a badge. Penetration testing is addressed with the same specificity, annually, black-box, and restricted to CREST-approved providers, with the latest scorecard available to clients on request, which is a gated tier and is credited as such rather than as open publication. What is missing is what an assessor would need: the certification body is not named, no certificate number is given, and no scope or statement of applicability is published, so a buyer cannot establish which systems and which locations the certificate covers. There is no trust portal.

The most complete certification statement located in this pull, missing only the period. The security page states that Clarilis is ISO/IEC 27001:2022 certified by BSI across the entire company, and gives the certificate number as IS 677941. That is four of the five things a buyer needs: the standard, the current revision rather than a superseded one, the certifying body, and a scope stated as company-wide, plus an identifier that can be checked against the certifier's own register. Regular penetration testing is stated separately on the AI Draft page. What is absent is the coverage period: no certification date, expiry or surveillance cycle is published, so a reader cannot tell how current the certificate is without looking it up. No report, statement of applicability or summary is published, and there is no trust portal on the estate, so nothing is gated and this is an absence of publication rather than a retrieval limit. No other certification is claimed, and no badge stands unsupported anywhere on the site.

The vendor refers to its AI throughout and identifies nothing underneath it. The estate speaks of AI-powered automation, drafting at AI speed and AI running inside firm rules, and describes a deterministic drafting engine combined with legal AI, without naming a model, a version, a provider or a processing location anywhere, and without committing to notify customers when any of it changes. Three companies are named on the estate and none of them answers this question: AWS, OVH and Azure are datacentre providers, which is infrastructure rather than a model and is graded on the deployment row, and Harvey is an integration partner whose own models sit on its side of the connection. No subprocessor list exists. The question has weight here because the material passing through the system is the firm's own precedent language and live negotiation positions, and a buyer cannot learn from any published source whose model reads them. This is the middle band because a distinct AI capability is described rather than merely labelled.

Three limbs answered plainly in a customer-facing FAQ, which is where a buyer would actually look. The provider is named, the vendor stating that it prefers OpenAI's model, and the hosting arrangement is named with it, hosted in Azure, with the AI Draft page adding that data is hosted in the customer's region via Azure and processed through zero data retention endpoints. The vendor also discloses that it uses a range of different models depending on the drafting task and configures them with practice-area knowledge from its in-house lawyers, which tells a buyer that the supply chain is plural rather than single. What fails is model naming and change notice, and the same sentence causes both: the answer says OpenAI's model without naming which model or version, and adds that Clarilis is always exploring the latest models to ensure the best outputs, which is a statement that the set will change and the opposite of a commitment to notify when it does. No subprocessor register is published, though AWS and Azure are named as platform infrastructure on the security page.

No pricing information is published at any level, including the unit of charge. There is no pricing page in the site navigation, which carries features, solutions, resources and company sections only, and no pricing entry appears in the footer. Every commercial route on every page read is the same single call to action, book a demo, and no self-service signup exists. Nothing published states whether the platform is licensed per user, per template, per document, per matter or per firm, no tier names appear, no minimum commitment or term is mentioned, and no implementation or template-build cost is described even though building automated templates from precedents is plainly a delivery exercise. With no customer agreement published either, the payment terms that would ordinarily disclose the charging structure are also unavailable. A buyer can learn nothing about cost without entering a sales process. No pricing row is owed on this record.

No pricing information is published at any level, including the unit of charge. There is no pricing page in the navigation, which carries what we do, solutions, customers, about us and resources, and none in the footer, which carries only a privacy policy and a sitemap. Every commercial route on the estate is the same call to action, book a demo. Nothing states whether the platform is licensed per user, per template, per document suite, per matter or per firm, and nothing addresses how the managed service is charged, which is a material question on this product because the automation is built and maintained for the customer by the vendor's own professional support lawyers rather than configured in-house, so implementation is a service engagement rather than a setup task. No tier names, minimum commitment or term length appears, and because no terms of service or master agreement is published either, the payment provisions that would ordinarily disclose a structure are unavailable. No pricing row is owed on this record.

Coverage is described with substance on two axes and the boundary is left open. The vendor publishes solutions by sector, separating law firms from in-house teams with a dedicated page each, and unusually also publishes solutions by role, addressing automators who build the templates, operators who run them and leaders who buy them, which is a more precise account of who inside an organisation the product is for than most records offer. Document coverage is evidenced through the drafting examples and feature material, reaching complex negotiated instruments including merger and financing documents and confidentiality provisions, and MassDraft is positioned for high-volume repapering. Geographic presence is stated through offices in London and Singapore. What is absent is any limit: no practice area is identified as unsuitable, no firm size band is given, nothing states which jurisdictions the templates or the AI have been built against, and nothing addresses languages, which matters for a vendor with an Asian office and a European base.

Coverage is documented with more specificity than any other record in this lane, and the boundary is never drawn. Four practice areas carry their own product lines with counts and named content: real estate with over seventy automated templates for commercial investment and letting in England and Wales plus more than fifty Property Standardisation Group documents for Scotland; corporate with nearly two hundred documents covering share purchase agreements, early-stage investments, share reorganisations and business purchase agreements; banking and finance covering leveraged and real estate finance facilities, simplified lending, security documents, legal opinions and APLMA-based agreements for Asia Pacific; and private client covering wills and estate planning. Engagement letters and in-house legal are addressed separately. Jurisdictional reach is stated product by product rather than as a claim, with pre-automated suites for Ireland, the United States and Canada, the Canadian content based on CVCA model forms. What is absent is the limit: nothing states which practice areas or jurisdictions are not served, no firm size band is given, and nothing identifies a document type the platform handles poorly.