B

BRYTER

The models are the engine of a core capability layered on a platform that would still function without them, which is the B band, and the vendor's own architecture says so. BRYTER's formula is stated on every page as the power of AI with the precision of workflows, and it names two halves. BRYTER Workflows is a rules-based no-code automation platform that was the entire company from 2018 and still sells on its own, covering intake and triage, approval workflows, document automation, client portals and compliance tools; strip the AI out and it remains a complete product.

BEAMON AI is the AI half and it is substantial rather than decorative: three named products in Assist, Extract and a Word add-in, its own domain, its own three-tier price list, its own free trial, and its own place in the master agreement's definition of the Software. What keeps this off A is that the record is BRYTER and BRYTER is both halves. The strongest argument the other way is recorded rather than suppressed: BEAMON AI is separately purchasable at published per-user rates and a buyer can acquire the AI without the workflow platform, which on a narrower record definition would read as the model being the product.

On the record as scoped, the honest reading is a rules engine with an AI suite bolted alongside it and increasingly through it, which is what the B band describes. Verified 12 September 2026.

Grounding is real and documented and no accuracy figure is published, which is the B band. What is documented: hallucination control and references are named as product capabilities; summaries are described as source-linked with citations kept attached and the level of detail selectable; Assist is described as returning verified answers based on the customer's own knowledge bank and playbooks; and the retrieval surfaces are named rather than implied, with a web search processor identified in the sub-processor schedule and a direct API integration to the Otto Schmidt legal database for research on the top tier.

The agreement adds a disclosure that is rare and worth quoting in substance: it states that the accuracy and reliability of Output may be affected where features such as web search or access to legal sources are not enabled, that in those cases the service is more likely to produce inaccuracies, and that the vendor therefore recommends enabling them. That is a vendor telling a buyer the conditions under which its own product gets worse.

What holds it off A is measurement. No accuracy figure, error rate, test set or evaluation is published anywhere, and the one comparative claim is unquantified marketing, that the proprietary AI outmatches generic tools like ChatGPT or Copilot on answer quality and accuracy. The agreement is markedly less confident than that page, stating that the vendor does not guarantee Output is accurate, complete, original, fair, unbiased or free from errors. R37 rule 1 gives it to the agreement and the gap is recorded here. Verified 12 September 2026.

A written commitment that the models work alongside a supervising lawyer, with real controls, short of the full structure, which is the B band. The commitment is contractual and stated three times over. The agreement provides that the customer is responsible for reviewing and validating all results before relying on them for business, legal or other decisions; that use of Output without proper human oversight, especially for decision-making or compliance, is at the customer's own risk; and that Output must be verified and validated prior to any use.

A fourth clause bars the customer from misleading any person that Output is human generated, which is an oversight provision pointed at the reader of the output rather than its author. The genuine control, and the best of them, is model governance rather than task governance: an administrator determines which language models are enabled for the tenant and may disable any of them at any time, and within that set the author selects the model for each individual input, with both selections expressly constituting instructions under the data processing addendum.

In the delivery state only the default model is enabled and no customer data reaches any other provider until an administrator turns it on. Access is role-restricted, actions are logged and changes traceable. R124(2) decides the ceiling. Nothing here attaches a boundary to a named tier and states what that tier's output may not be used for; what is published is a general assurance of human review, however carefully drafted, plus a model-enablement control. Verified 12 September 2026.

Named customers in unusual quality and no figures attached to any of them, which is the B band. The naming is specific and spans three buyer types that rarely appear together: global law firms in Linklaters, Ashurst and Hausfeld; professional services firms in Deloitte, PwC and KPMG; and corporate legal departments at McDonald's, ING Bank, Rakuten and Telefonica. A published customer story covers the Telefonica legal team, and one attributed testimonial carries a named individual with title and firm, a counsel and AI expert at REIUS.

For a vendor of this size that is a strong reference base and it is treated as attribution rather than outcome. What is absent is measurement. Not one figure is attached to a named customer on the surfaces read, nothing is dated to a deployment, and the benefit claims are generic, saving hundreds of hours reviewing contracts for key clauses and freeing teams from repetitive work, with no basis, population or period given.

Nothing published measures BEAMON AI specifically as against the workflow platform, which matters because the AI is what is being graded. Two artifacts are named and not opened under R25, each corroborating rather than carrying this grade: the customer stories library and the guides and reports section, and they are what would move this row if any study carries a dated figure with a stated method. Verified 12 September 2026.

All five limbs are met in instruments a buyer can read before signing, and the limb R33 makes decisive is met more substantively than on any record in this corpus. Privilege and professional secrecy: the data processing addendum carries a dedicated section on the processor's obligation to maintain professional secrecy, which applies wherever the customer is subject to such duties, and which provides that the processor shall treat all data as potentially subject to professional secrecy obligations unless clearly determined otherwise, shall obtain knowledge of professional secrets only to the extent strictly necessary, with obtaining knowledge expressly defined to include any technical or organisational access regardless of whether human review occurs, shall keep them strictly confidential, may disclose them to sub-processors only where those sub-processors are contractually barred from onward disclosure and must flow the obligation down, and shall ensure its own personnel undertake in writing not to disclose them.

Alongside it the vendor states compliance with the German professional rules and the criminal professional-secrecy provision, and states that it upholds attorney-client privilege. Training: prohibited contractually. Retention and deletion: a retention table by data category, deletion or return at the customer's choice on termination, and a bounded backup window. Third-party model provider: the providers are named with their processing locations and the default model's processing is stated to occur in the European Union.

Segregation: tenant separation and a role concept documented in the technical and organisational measures. Recorded against it, because a reader should weigh both: the master agreement separately makes the customer solely responsible for ensuring that any Input complies with privilege and professional secrecy obligations, so the vendor undertakes to protect the material and disclaims liability for the decision to submit it. Verified 12 September 2026.

A real position on advice versus tooling, published contractually and reaching into supervision, short of A on jurisdiction. What is published: the agreement states in terms that Output shall not be viewed as legal advice, that BRYTER is not a law firm, does not practise law and does not give legal advice, and that it bears no legal responsibility for Output or anything derived from it. The usage restrictions add two provisions most records lack, a bar on misleading any person that Output is human generated, and a bar on offering tailored financial advice without a qualified person reviewing the information, which is the supervision idea applied to an adjacent regulated activity.

Supervision of the legal output itself is addressed through the review and validation obligations graded on the autonomy row. Beyond the agreement the vendor engages professional conduct directly in at least two jurisdictions, mapping the AI suite to named Austrian provisions and stating compliance with the German professional rules and the criminal professional-secrecy provision, which is graded on the bar guidance signal.

What holds it off A is that jurisdiction limits are not stated. Nothing tells a buyer where the product may or may not be used, which is a live question for a German-contracting vendor selling in the United States through a separate entity, operating in at least three language markets, and publishing bespoke supplementary terms for Austrian and Swiss professional-secrecy holders without stating what that implies for anyone else.

Nor is who may operate the tool addressed beyond the licensing concepts of Authorized User and Author. Verified 12 September 2026.

A published governance framework with real substance, short of testing results and a named owner, which is the B band. What lifts it well clear of a principles page is that the framework is contractual. The master agreement carries a dedicated AI Terms section governing every feature enabled by machine learning or other AI functionality, and that section does governance work rather than risk-shifting alone: it prohibits training on customer content, it governs model selection and model change, it requires human oversight, and it sets out an acceptable-use regime barring generation of hateful, harassing or violent content, exploitation of children, self-harm encouragement, misuse of personal data, malware, high-risk economic harm, health advice, political campaigning and lobbying, and use that infringes privacy or confidentiality.

The bias limb is addressed expressly and in the agreement rather than in marketing, the customer acknowledging that Output may reflect or amplify biases inherent in training data and undertaking to verify and validate all Output before use. That is a vendor putting the bias risk on the face of the contract. Around it sits real organisational governance, an ISO 27001:2022 certification, a data protection management system with reporting lines to senior management, records of processing under Article 30, privacy-by-design training for product teams, and internal and external audits.

What A asks for is still absent. Nobody is named as accountable for AI, nothing describes what is evaluated before a model or feature ships, no testing results are disclosed, and the bias acknowledgement is a warning rather than a mitigation. R124(3) is not the binding constraint here, because this framework governs the vendor's own conduct as well as the customer's, but the vendor-side testing limb is what separates this from A. Verified 12 September 2026.

Every limb this band asks for is published, and the sub-processor limb, which most records fail, is answered in a table. Retention is stated by data category rather than in general terms: security and audit logs up to ninety days, contact, credential and customer content until termination and then deletion under the addendum's deletion clause, and backup copies for a bounded thirty days thereafter. Deletion is a commitment with a choice attached, the processor deleting or returning and deleting all personal data and processing results at the controller's election on termination, alongside a switching and porting right under the EU Data Act with a documented technical process.

Access control is documented control by control across twelve headings covering physical access, user and data access, separation, pseudonymisation and encryption, input control, order control, transmission, availability, resilience, security management, purpose limitation and data protection management. Sub-processors are named individually with service, corporate location and server location, and changes require prior written notice with a twenty business day objection right.

Incident practice is specific: notification without undue delay and where feasible within seventy-two hours, with the four contents of that notice enumerated, plus tested business continuity, incident response and disaster recovery plans. Two disclosures are unusually candid and belong on the record rather than being smoothed over: the addendum discloses monitoring of AI-powered processing described as LLM observability, and discloses that the provider of the selected language model performs automated screening of inputs and outputs against its own usage terms with case-by-case human review of flagged content. Verified 12 September 2026.

A real published position on liability, short of the full picture, which is the B band, and it is better than most records reach because it contains actual warranties rather than only exclusions. The vendor warrants that it has power to grant the rights, that the Software will be provided with reasonable skill and care and in compliance with applicable laws, that it will materially conform to the Documentation and the Order, and that it will not materially decrease functionality during the term, with a remedy of prompt correction or an alternative means of performing the task.

There is an uptime commitment with a support and maintenance appendix behind it, and published support response targets of forty-eight hours standard and twenty-four hours on the top tier. The liability regime is German-law shaped and therefore more favourable at the top than an American cap: unlimited liability for intent, gross negligence and injury to life, body or health, with slight negligence limited to breaches of material contractual obligations and capped at the fees payable for the current subscription period.

What holds it off A, and the note names it plainly because it is the crux, is that the AI is carved out of the strongest protections. The agreement states that the support and maintenance services apply to the AI Service only to the extent that it may suffer downtime, interruptions and errors without any guaranteed response or resolution time, and separately that the vendor does not warrant the availability, accuracy or completeness of results delivered through credit-based features.

So the uptime commitment covers the platform and expressly not the models. No indemnity running to the customer was located, no insurance position is published, and the only indemnities run the other way. Verified 12 September 2026.

Real integrations, named individually across several classes, with conditions published, short of documentation an implementer could work from. The catalogue is broad and specific. Document and knowledge systems: iManage and SharePoint, described as connecting repositories so the right precedents appear in context. Business systems: Salesforce and DocuSign. Collaboration: Microsoft Teams, Slack and email as intake channels into the legal front door.

Authoring: a Microsoft Word add-in for drafting, editing and redlining, sold as a named product. Publishers and data: a direct API integration to the Otto Schmidt legal database, Haufe Lexware, the Winsolvenz insolvency product via Septeo, and Handelsregister company data via Company.info. Language: DeepL, which also appears in the sub-processor schedule with its processing locations. Platform: BRYTER Connect and an open API, with published API terms of use, a licence to build against the API, software development kits and code samples, and a stated ability to trigger modules from a third-party system and export data to an internal dashboard.

Commercial conditions are published per connector on the pricing table, which is better than most records manage, stating which integrations require the customer's own subscription to the third party and which are purchasable add-ons. What holds it off A is documentation depth: no public API reference or connector register was located, the help centre was not opened, and for most connectors nothing states what objects move in which direction or what configuration a firm must perform. Verified 12 September 2026.

Real residency detail is published, in more depth than almost any record in this corpus, and the vendor's own documents contradict each other twice on the central fact, which is what holds this at B rather than A. What is published: hosting in the EU stated on the pricing table for every tier; a sub-processor schedule giving a server location for each named party, with cloud hosting in Frankfurt, the default language model hosted by Microsoft Azure in the European Union, and other locations named individually; and a clause stating that in the delivery state only the default model is enabled, whose processing takes place within the European Union, and that until an administrator enables another model no customer data is transferred to that model provider.

Tenancy is addressed through the Tenant concept used throughout the agreement and through documented separation of customer data into separate systems. That is a buyer being told both where the data sits and where inference happens, which is the distinction most records never draw. The two contradictions are recorded under R37 rule 2 and neither is resolved in the vendor's favour. First, within the data processing addendum, the sub-processor schedule gives server locations in Frankfurt, the EU, Sweden and the Netherlands, while the technical and organisational measures schedule in the same document states that all personal data is stored and processed in United States data centres of AWS and Microsoft Azure.

Second, the security page states that customer data, prompts and outputs are never shared outside the EU and EEA, while the same schedule lists an optional language model provider and a web search processor with server locations in the USA. A buyer cannot reconcile these from what is published. Verified 12 September 2026.

Certification is real, current and named to the standard's own version, and the route to the evidence is broken, which holds this at B. The claims are made first-party in two places, on the pricing comparison table against every tier and again on the security page, and they are specific: SOC 2 Type II and ISO/IEC 27001:2022, the latter named with its version year rather than as a bare standard number. The data processing addendum corroborates from a different angle, recording internal and external ISO 27001 audits under its security management measures and setting out twelve headings of technical and organisational measures in detail, and it separately names the certifications held by the two cloud sub-processors.

C does not fire: these are explained claims in the vendor's own voice sitting on top of a documented control set, not badges. What is missing for A is the evidentiary apparatus and, concretely, a working route to it. No auditor is named, no report period or observation window is published, no certificate or report number appears, and the trust centre was not reached. The access flow could not be graded under R5 because of a published-surface defect recorded here as a finding about the vendor rather than a limit on this index: the pricing table links a trust centre at trust.bryter.io, which returns a 404, while the security page links trust.bryter.com.

The working address was identified and not opened, and it is named as the artifact that would move this row. Verified 12 September 2026.

The supply chain is partly disclosed, and the part that is disclosed is stronger than any record so far, which is B by design on this band. Providers are named, in the agreement and in the sub-processor schedule: Microsoft Azure is identified as the cloud server and host of the default language model used within the AI suite, Anthropic Ireland is identified as an optional language model engaged only upon enablement by a customer administrator, a web search processor is named for the AI suite, and the usage restrictions separately name OpenAI and Azure OpenAI among the third parties involved in providing the AI Service.

Each is given a corporate location and a server location. Change notification exists and is a real mechanism rather than a promise: prior written notice before any addition or replacement of a sub-processor, with a twenty business day objection right on materially important grounds and a defined consequence if the objection stands. Customer control over which provider processes what is documented at two levels, administrator enablement per tenant and author selection per input.

R34 decides the ceiling and it is applied strictly here because the temptation to round up is real. The A band requires that the models underneath are named and their providers identified as separate limbs, and no model is named anywhere: the pricing table says only that available models are updated continuously, and the agreement states that the vendor may add, replace, upgrade or discontinue underlying models at any time and that the customer has no claim to the availability of any specific model or model version.

A buyer knows whose infrastructure runs the inference and does not know which model is answering. Verified 12 September 2026.

A buyer can price the AI suite to the euro without speaking to anyone, and can then read the commercial mechanics that will govern the contract, which is the A band met twice over. Rates, per user per month: Single at 99 euros annual or 118 euros monthly, for solo practitioners without team collaboration; Team at 149 euros annual or 176 euros monthly, adding shared projects, documents and grids with a central prompt library; Pro at 199 euros annual or 234 euros monthly, adding workflow automation, single sign-on, integrations and priority support.

Both billing cadences are shown side by side so the annual discount is calculable. Structure sits underneath in a comparison table of roughly eighteen rows across core AI features, enterprise features, integrations and publishers, and compliance and support, and it is unusually honest at the edges, marking which integrations require the customer's own third-party subscription and which are purchasable add-ons. The purchase path is genuinely self-serve, with a free trial and sign-up rather than a demo request, and a published set of purchase terms for buying through Stripe.

What is rarer still is that the agreement publishes the mechanics a buyer normally discovers later: annual invoicing in advance, an uplift capped at ten per cent a year, ninety days' notice and a termination right for renewal-term fee changes, thirty days' notice for credit rate changes, a monthly credit cap with non-rollover and forfeiture, a stated token consumption limit with notification at eighty per cent, thirty-day payment terms and the interest rate on late payment.

Two limits are recorded and neither displaces the grade: BRYTER Workflows is not priced publicly and is quoted, and premium model access may carry additional fees set in the Order. Verified 12 September 2026.

Coverage is documented with real substance across segments and tasks, short of the stated limits the A band asks for. Segments are addressed on surfaces of their own rather than in a list: a Law Firms solution area and a Corporate Legal solution area, each with its own landing page and its own named use cases, and the customer roster demonstrates all three of the buyer types claimed, global law firms, professional services firms and corporate legal departments.

Firm size is addressed through the pricing tiers, which are described by the shape of the practice rather than by seat count, the entry tier stated to be for solo practitioners and expressly without team collaboration features, the middle tier for teams needing shared projects and a prompt library, and the top tier for firms and legal departments needing a full enterprise environment. Task coverage is the strongest part and is enumerated rather than asserted, a published use-case library of around twenty items running from clause extraction and contract review to due diligence, discovery summarisation, litigation and matter summary, horizon scanning, invoice summarisation, redlining and NDA generation.

Jurisdictional reach is evidenced indirectly but concretely, through English, German and Spanish estates, German and Austrian legal database integrations, and bespoke supplementary terms for Austrian and Swiss professional-secrecy holders. What holds it off A is that the limits are not stated. Nothing identifies a practice area or a firm type the product is not for, nothing addresses government use, and the jurisdictional position is left to be inferred from which language sites and which national integrations exist rather than being written down. Verified 12 September 2026.