Product and intended use
Define the configuration, people, information and tasks being assessed.
What is sold, and what is not yet available
Public claim only
The vendor publishes a technology menu of Kanon, the Blackstone Graph (marked 'Soon') and semchunk, and a capability menu of Embedding, Reranking, Enrichment, Extraction, Classification and Research (marked 'Coming soon'), read 2026-09-27. The pricing page lists five chargeable models: Kanon 2 Embedder, Kanon 2 Reranker, Kanon Universal Classifier, Kanon Answer Extractor and Kanon 2 Enricher. Two plans are published: pay-as-you-go API/Cloud, and Advanced (Enterprise) for private deployments, volume discounts or an alternative pricing model. The API is a paid service requiring a monthly billed, zero flat fee, usage-based subscription.
Ask: When will the Research capability and the Blackstone Graph become generally available, and will they be priced per token on the same subscription?
What the product does
Public claim only
The vendor's home page (read 2026-09-27) publishes capabilities of Enrichment ('Turn messy documents into data your systems can digest. Kanon 2 Enricher maps sections, entities, citations, and cross-references into a structured knowledge graph'), Retrieval, Extraction, Classification and Research (marked coming soon). The documented API exposes endpoints for embeddings, rerankings, enrichments, extractions described as extractive question answering, and universal classification, together with a query language the vendor calls IQL with its own specification and templates, and a documented chunking algorithm. semchunk is published as an open-source chunking library. There is no user-facing legal workflow, matter file or document editor: what is sold is model inference consumed by other software.
Ask: For an Australian firm that is not building software, is there any supported way to use these models other than through a vendor or platform that has embedded them?
Who it is sold to, and named users
Public claim only
The home page states products are sold to legal-technology builders and publishes, under the heading 'Our products are used by', a logo wall naming Cleary Gottlieb, Webb Henderson, Nous Group, AccuFind, Microsoft, IBM, Australian Government and Singaporean Government. Four named-attributed testimonials on the same page are from AccuFind, Jurimesh, Lovguiden and Esheria, and a fifth from Webb Henderson; the vendor publishes these as customer statements. The terms state that no attorney-client relationship is created and that outputs 'do not constitute and are not a substitute for legal advice and legal counsel'.
Ask: Which Australian Government agency is the named user on the logo wall, and on what basis is each organisation on that wall described as using the products?
Delivery modes
Public claim only
The home page publishes two modes: 'Online platform - Access our models and documented APIs via our online platform, with continuous availability, minimal latency, and zero permanent data retention'; and 'Private deployments - Host our models and APIs in your own AWS tenancy or on your own hardware with private, air-gapped containers-ensuring maximal SOC, ISO, and IRAP compliance.' The pricing page lists private air-gapped deployments, finetuning on the customer's own data and 'Azure on request' under the Advanced plan, and states that pricing for the air-gapped Amazon SageMaker models is published on AWS Marketplace. No region is named for any private deployment.
Ask: For a private deployment in our own tenancy, can the container run in AWS ap-southeast-2 or on hardware in Australia, what licence fee applies, and who patches the model containers?
Beta programme
Public claim only
Terms section 11 establishes a Beta Program under which beta services 'are experimental and may not be documented, supported, reliable, stable, or secure', are provided as is without warranty of any kind, and may change or be withdrawn at any time, with use entirely at the member's own risk. Beta membership may be granted, denied or withdrawn at any time without notice. Beta members must keep non-public information about the programme confidential, and grant the vendor a licence to use their organisation's name, logo and trade marks publicly in marketing to identify them as a beta member, with removal only from new materials on request.
Ask: Is any capability we would rely on currently part of the Beta Program, and can the marketing-use licence over our firm's name in section 11 be excluded before we join?
Check for vendor content directing how AI describes the product
Not confirmed
A deliberate check was made on 2026-09-27. A site map of isaacus.com returning 76 URLs shows no /llms.txt, /ai-info, /ai or /for-ai path; the site does publish /media-kit and /press pages, which were not read. The home page, terms of service, privacy policy and pricing page, all read in full, contain no block addressed to AI models, assistants, LLMs or crawlers and no 'Instructions for AI Assistants' heading. No question-and-answer page on the site was read, and no comparative or superlative self-answer in machine-readable question-and-answer form was found on the pages read; the home page does carry the sentence 'We lead the pack in frontier legal AI' as marketing prose, recorded under qualityEvidence.
Ask: Does the vendor publish any page, file or embedded block, including in the media kit, that tells AI models, assistants or crawlers how to describe Isaacus?
Data, controls and requirements
Compare the actual configuration with firm, client and jurisdiction requirements, supported by evidence.
Authority to put client material through the API
Public claim only
Terms section 9 requires the user to represent and warrant that it has 'the rights, licenses and permissions that are necessary for us to process your Inputs under these Terms and to provide our Services to you, including for example, to integrate with third-party services and to share Materials with others at your direction', and that inputs will not violate the terms or any applicable law. Section 13 states that no service constitutes or substitutes for legal advice and that no attorney-client relationship is created. The privacy policy records that personal data may be received indirectly through inputs and outputs sent to and from the AI services.
Ask: Where the API is embedded in another vendor's platform, who is the controller of our client material for Privacy Act purposes, and does Isaacus require that vendor to obtain our authority?
Which models run, and whose they are
Public claim only
The vendor's own models are the product. Five are published with per-token prices: Kanon 2 Embedder, Kanon 2 Reranker, Kanon Universal Classifier, Kanon Answer Extractor and Kanon 2 Enricher; the caller selects which model an API call invokes. No third-party model or model provider is named on the pages read, and the Advanced plan offers 'finetuning on your own data'. The documentation pages describing model internals, tokenisation and chunking were not read within the credit budget. The terms allow the vendor to update its AI models without advance notice.
Ask: Which model version answers a given API call, how are versions pinned and deprecated, and is any third-party model or third-party inference provider in the path for the hosted API?
Subprocessors and their locations
Public claim only
The privacy policy (16 April 2026) names Sentry for error monitoring, Stripe for payment and billing, Azure for hosting the platform, API and user database in the United States, AWS for hosting and serving models in the United States with AWS Global Accelerator proxying between them, Hetzner in Hillsboro, Oregon with WordPress and RunCloud for the public website, and Microsoft 365, SendGrid and AWS SES in the United States for email. It states 'We provide a comprehensive, up-to-date list of our subprocessors and their locations in the Isaacus Trust Center' at trust.isaacus.com, which was not read within the credit budget. No notification period or objection right for subprocessor changes was found, and no data processing addendum is published at any URL on the site map.
Ask: Do you offer a data processing addendum with a subprocessor change-notification period and an objection right, and can Sentry error logging be disabled for our traffic?
Training on customer content
Public claim only
Terms section 9 states 'we will not train our models on any Materials that are not publicly available or otherwise use them to improve the Services' except in two circumstances: where the user provides feedback regarding any materials, which may then be used as feedback under the terms; and where the user otherwise expressly consents. The privacy policy section 8 states 'We will not train our models on Inputs and Outputs except where you expressly consent to doing so', giving as examples opting in 'if and when that ever becomes possible' or providing feedback regarding those inputs and outputs. Both documents were last updated 16 April 2026.
Ask: If a user of our firm rates an output with a thumbs up or thumbs down, does the material in that request become available for model training, and can rating be disabled for our account?
Feedback and ratings
Public claim only
Terms section 10 states that when a user shares comments, ideas or feedback, 'including rating Outputs, for example, by using a thumbs up or thumbs down button', the vendor is 'free to use such Feedback without any restriction or compensation to you'. Section 9 makes feedback one of only two routes by which materials may be used for training or service improvement. The privacy policy states inputs and outputs may be persisted where the user provides feedback on them. No limit on the retention of feedback content, and no way to redact material from feedback, is published.
Ask: What exactly is captured when a rating is submitted: the full input and output text, or the rating alone, and for how long is it kept?
Retention of inputs and outputs
Public claim only
The privacy policy section 7, headed 'No persistent retention of Inputs and Outputs', states 'By default, we do not retain Inputs and Outputs under normal operations', with an exception where an error is raised, in which case inputs or outputs or fragments may be sent to Sentry solely to identify and resolve the cause, and a further case where the user expressly opts in or provides feedback. The terms say the same and add that the vendor may use materials only as reasonably necessary to maintain the Services. Section 9 of the policy sets out retention of other data by category, including account, contract and billing records, usage metadata, support records and security, audit and incident records, each described by purpose rather than by period. Terms section 12 reserves the right to retain information for legal or regulatory compliance and under standard archiving, recovery and backup processes.
Ask: How long do Sentry error logs containing input fragments persist, in which region is that Sentry project hosted, and what is the maximum retention period for usage metadata?
Deletion on request
Public claim only
The privacy policy states that when personal data is no longer reasonably required 'we will delete it, anonymize it, or de-identify it, unless we are required or permitted by law to retain it', and that users may 'Request deletion of personal data in certain circumstances'. Complaints may be made to the vendor's legal mailbox and, by implication of the stated Australian Privacy Principles alignment, to the Australian regulator. No deletion timeframe, no self-service deletion control and no certificate or evidence of deletion is published, and the trust centre was not read.
Ask: On written request, within how many days is account data deleted, what evidence of deletion do you provide, and how are backups and Sentry logs covered?
Vendor personnel access to content
Not confirmed
Nothing on the pages read states whether Isaacus personnel can access inputs or outputs, under what conditions, or with what logging. The privacy policy section 11 says only that measures 'may include encrypting Inputs and Outputs in transit, hashing secrets, using firewalls, and implementing authentication, monitoring, and administrative safeguards', and section 5 states that service providers processing personal data on the vendor's behalf are required to handle it under appropriate confidentiality, security and data protection obligations. The stated default of no retention of inputs and outputs limits what could be accessed, but no access policy for the Sentry error path or for opted-in data is published. The trust centre at trust.isaacus.com was not read within the credit budget.
Ask: Which roles at Isaacus can read inputs or outputs captured in an error log, is that access logged, and are staff subject to background checks and confidentiality undertakings?
Authentication and account administration
Not confirmed
The terms require an account for paid services and allow authentication using Google and GitHub accounts as third-party services. Nothing on the pages read describes single sign-on with a firm identity provider, enforced multi-factor authentication, role-based access, API key scoping or rotation, organisation or workspace administration, or an audit log. The privacy policy refers generally to 'implementing authentication, monitoring, and administrative safeguards' without detail. The platform console and documentation were not read within the credit budget.
Ask: Does the Isaacus Platform support SAML or OIDC single sign-on, enforced multi-factor authentication, scoped API keys and an administrator audit log of key creation and use?
Encryption in transit and at rest
Public claim only
The privacy policy section 11 states the vendor takes reasonable technical and organisational measures and that 'Such measures may include encrypting Inputs and Outputs in transit, hashing secrets, using firewalls, and implementing authentication, monitoring, and administrative safeguards.' The wording is permissive rather than a commitment, names no protocol version or cipher, and says nothing about encryption at rest or key management. No security page exists on the site map; the trust centre was not read.
Ask: Which TLS version is enforced on api.isaacus.com, is data at rest encrypted and with whose keys, and will you commit to those measures contractually?
Certifications and independent assurance
Not confirmed
No certification held by Isaacus is named on any page read. The home page states that private, air-gapped deployments in the customer's own tenancy or hardware ensure 'maximal SOC, ISO, and IRAP compliance', which is a statement about the deployment mode rather than a certification the vendor holds; no certificate number, scope, auditor or date is published. The privacy policy points to the Isaacus Trust Center for 'our controls, subprocessors and security, privacy, and compliance posture'; that site was not read within the five-operation credit budget for this product, so any certification listed there is unverified here.
Ask: Does Isaacus itself hold SOC 2 Type II, ISO/IEC 27001 or ISO/IEC 42001 certification or an IRAP assessment for the hosted API, and can you provide the report and its scope?
Contracting entity, governing law and consumer law
Public claim only
The terms (16 April 2026) are a binding agreement with Isaacus Pty Ltd, ACN 684 344 134, ABN 15 684 344 134, an Australian proprietary company limited by shares registered in Victoria with a registered office in Surry Hills, New South Wales. They are governed by the laws of Victoria, Australia, with exclusive venue in the courts of Victoria, New South Wales or Queensland, and they state that refunds are offered in accordance with the relevant consumer law in the user's location, 'for example, the Australian Consumer Law', and that rights under the terms are in addition to consumer rights. The terms may be changed by posting revised terms, which take effect 30 days after posting for existing users and immediately for new users. No liability cap, indemnity or warranty section appears in the terms read.
Ask: Is there a separate enterprise agreement for the Advanced plan with a liability cap, warranties and an indemnity, and would it keep Victorian governing law?
Where processing happens and cross-border flows
Public claim only
The privacy policy section 10 states 'Isaacus is based in Australia but operates internationally. The vast majority of our critical infrastructure is located in the United States, with our own internal services operating from Australia.' It then names specifics: Azure hosts the platform, API and user database in the United States; AWS hosts and serves the models in the United States, with AWS Global Accelerator proxying between the Azure-based API and the AWS-based inference servers; the public website runs on Hetzner in Hillsboro, Oregon; and transactional email uses AWS SES in the United States. For an Australian firm this means inputs to the hosted API, including any client material, are processed in the United States unless a private deployment is used. No Australian region is offered for the hosted service on any page read, and the pricing page lists 'Azure on request' under the Advanced plan without naming a region.
Ask: Is an Australian region available for the hosted API or, in a private deployment, for AWS ap-southeast-2, and if so under which plan and at what price?
Security incident and breach notification
Not confirmed
Neither the terms nor the privacy policy read on 2026-09-27 contains a security incident or data breach notification commitment, a notification period, or a reference to the notifiable data breaches scheme. The privacy policy section 9 mentions that 'Security, audit and incident records' are retained to detect, investigate, remediate and document security or operational incidents, and section 14 sets out a complaints process to the vendor's legal mailbox, but neither creates a notification obligation to the customer. No data processing addendum is published.
Ask: Within how many hours of becoming aware of a security incident affecting our data will you notify us, who is contacted, and will you commit to that in writing?
How terms and the service change
Public claim only
The terms state they may be changed by posting revised terms on the website, taking effect 30 days after posting for existing users and immediately for new users, and that a user who does not agree must stop using the Services before they take effect. Terms section 7 allows manual or automatic updates to the API, API client, AI models, applications and other software 'without advance notice'. Section 4.5 states the vendor may change the information, features and other services offered through a subscription and makes no guarantee that anything currently available will remain available. The pricing page states prices may change with notice before new pricing takes effect.
Ask: How are customers told when a model is updated or retired under section 7, and is there a deprecation window or a pinned-version option for production use?
Cancellation and getting out
Public claim only
The terms state a subscription may be cancelled through the billing portal or the support form provided all incurred fees are paid and no payments are outstanding, that cancellation must be at least one hour before the scheduled end date to avoid renewal, and that subscriptions renew automatically at the full price after any discount or trial. The API subscription is monthly billed with a zero flat fee and usage-based charges. Because the vendor states it does not retain inputs and outputs by default, there is no published export or data-return process; no migration assistance, escrow or model-availability commitment appears in the terms read.
Ask: If we depend on a Kanon model in production and either party ends the relationship, what notice and what continuity options exist, including any container licence or escrow?
Integrations and third-party services
Public claim only
The vendor publishes an Amazon SageMaker integration and a LangChain integration in its documentation, an integrations page, cookbooks in a public code repository, and AWS Marketplace listings for its air-gapped models. Terms section 6 states the Services may be used with third-party information, products or services, gives Google and GitHub authentication as examples, and states the vendor does not control them, accepts no responsibility, and that their use is at the user's own risk and subject to their own terms and privacy policies.
Ask: Which integrations are supported by Isaacus rather than community-maintained, and for a SageMaker deployment who is responsible for model updates and security patching?
Whether the service reaches out to live web content
Not confirmed
Nothing on the pages read states that any Isaacus model or endpoint retrieves live web content. The published capabilities operate on text supplied by the caller: embedding, reranking, enrichment, extraction and classification. The vendor's forthcoming Research capability and Blackstone Graph are described as legal research and legal knowledge graph products, but neither is available and no source or retrieval mechanism for them is published. The documentation pages that would settle this were not read within the credit budget.
Ask: Do any current endpoints fetch external content at inference time, and when the Research capability ships, what sources will it retrieve from and will retrieval happen in Australia?
Autonomous actions
Not confirmed
No page read describes agentic behaviour, tool use or any ability to act in external systems. The published surface is a set of inference endpoints called by the customer's own software, so any action sits in the calling application rather than in the product. The documentation and the IQL specification were not read within the credit budget, so nothing is asserted about what those describe.
Ask: Do any endpoints or planned features initiate actions or make outbound calls of their own, or is every call initiated by the caller's application?
Verification of output before use
Public claim only
Terms section 9 states that outputs 'may not always be accurate and may contain material inaccuracies even if they appear to be accurate because of their level of detail or specificity', that the user 'should not rely on any Outputs without independently confirming their accuracy', that reliance is strictly at the user's own risk and the vendor disclaims all liability for decisions made based on outputs, that outputs may not be correct, current or complete, and that outputs 'do not constitute and are not a substitute for legal advice and legal counsel'. Section 13 states no attorney-client relationship is created. These statements sit alongside the home page headline 'Hallucination-free legal AI models. Really.', which is not repeated in the terms.
Ask: How should the 'hallucination-free' description on the home page be read against section 9 of your terms, which says outputs may contain material inaccuracies and must be independently confirmed?
