Product and intended use
Define the configuration, people, information and tasks being assessed.
Product and plan in scope
Public claim only
As at September 2026 the SaaS platform is branded Relativity aiR (formerly RelativityOne). Pricing materials include Assist, aiR Review, Privilege, Case Strategy and claiR (Advanced Access) with committed data volume. Relativity Server, Relativity aiR Government, Translate, Data Breach Response, FOIA, Legal Hold, Gavel and MCP-connected external models are separate configurations.
Ask: Confirm in writing the exact Relativity aiR instance region, included AI modules, add-ons, Relativity Server versus SaaS, Government versus commercial tenancy, and whether Gavel and MCP will be enabled for our matters.
Deployment model
Public claim only
Relativity aiR is a Microsoft Azure SaaS service with named regional instances, including Australia East. Help documentation still publishes Relativity Server versions, which is a different deployment and must not be assumed from the cloud profile.
Ask: Will we take Relativity aiR SaaS in Azure Australia East only, and if any Relativity Server, private-cloud or Government tenancy is proposed, what data and models sit in each?
AI features in the platform
Public claim only
Documented aiR systems include Review (relevance, issues, citations), Privilege (privilege prediction and draft logs), Case Strategy (fact extraction, timelines, witness summaries, deposition outlines), Data Breach Response, Assist conversational search, and claiR. Other AI-integrated features include AV transcription, conceptual analytics, Contracts extraction, FOIA, sentiment analysis, Translate, Review Center classifiers and MCP.
Ask: Which of these features will be enabled in our tenancy, which are generally available versus Advanced Access, and which consume aiR Units or separate licences?
Australian hosting and legal-content coverage
Public claim only
Help documentation (modified 18 September 2026) lists Australia as both Relativity aiR instance region and aiR system inference location. The platform reviews the firm's hosted documents; it does not supply Australian primary-law databases. Relativity has marketed the product to Australian government agencies with an IRAP Protected claim.
Ask: Contractually pin Australian storage, backups, logs and inference for our instance, and list every exception (support, telemetry, MCP, Translate, disaster recovery).
Data, controls and requirements
Compare the actual configuration with firm, client and jurisdiction requirements, supported by evidence.
Client and outside-counsel authority
Not confirmed
Relativity aiR is an enterprise discovery system that holds confidential, privileged and often personal matter data. Relativity names Australian users and partners but does not publish client-consent or outside-counsel-guideline tooling. Matter-level exclusion of clients who prohibit generative AI is a firm configuration issue.
Ask: Can administrators exclude nominated clients or matters from aiR processing, and what plain-language description of hosting, models and retention can we give clients and include in engagement terms?
Foundation models in the inference path
Public claim only
Help documentation (18 September 2026) states aiR systems currently use GPT-4o, GPT-4.1 mini, GPT-5 mini, GPT-5.1 and GPT-5.2 through Microsoft Azure services, with the specific model chosen by Relativity according to capability and availability. Customers do not appear to pin a model. MCP separately lets external assistants (Claude, Gemini, Copilot, OpenAI) act on the platform.
Ask: Name the Azure OpenAI deployment type and model versions that will process our tenancy, whether we can pin or refuse a model, and which data MCP-connected assistants receive.
Subprocessors and change notice
Not confirmed
Microsoft Azure and Azure OpenAI are named in product and security pages. No dated public subprocessor schedule with purpose, location, flow-down terms, notice period or objection right was found. MCP and App Hub can add further processors if enabled.
Ask: Provide the current subprocessor list (hosting, model, transcription, translation, support, analytics), processing countries, flow-down of no-training and deletion, and 30-day notice with a right to object or terminate.
Training on customer content
Public claim only
aiR Review help states Azure OpenAI does not retain analysed documents beyond the organisation's instance and does not use them to train Relativity, Microsoft or other third-party models. The AI-products page repeats that customer data processed using aiR systems is not used to train or improve third-party models or services. These statements sit in documentation, not in a public DPA or MSA reviewed here.
Ask: Point to the signed clause covering prompts, uploads, outputs, embeddings, logs and feedback, binding Relativity and every model provider, and confirm no 'improve the services' licence remains.
Feedback, ratings and evaluation data
Not confirmed
Public help pages address document inference and third-party model training. They do not state whether thumbs-up/down, corrected coding, prompt criteria or support tickets are used to evaluate or improve Relativity's own models or prompts.
Ask: Are reviewer corrections, ratings, prompt criteria and support content used to evaluate or improve any Relativity or provider model, and can that use be switched off for our tenancy?
Retention of prompts, documents and logs
Public claim only
Azure OpenAI is described as not retaining analysed document data beyond the organisation instance. Relativity still stores the discovery corpus, aiR outputs, audit logs and backups in the customer's Azure region. No public numbered retention schedule for prompts, conversation history, embeddings or abuse-monitoring copies was found. Privacy policy retention language applies to Relativity's own website personal data, not Customer Data in the software.
Ask: Provide a retention schedule for each class (documents, prompts, aiR outputs, embeddings, audit logs, backups, provider abuse stores), administrator controls, and written confirmation of zero or modified retention at Azure OpenAI for our tenancy.
Deletion on instruction and on exit
Not confirmed
Help documentation covers workspace archive/restore and customer-managed keys. A public contractual deletion period, matter-level certified deletion, backup-purge lag, or flow-down to Azure OpenAI was not found. Website privacy deletion rights do not govern Customer Data in the product.
Ask: On termination and on a matter-level instruction, what is deleted from primary stores, indexes, embeddings, caches, logs, backups and subprocessors, within how many days, and will you issue a deletion certificate?
Vendor and support access to content
Public claim only
Technical overview describes Relativity Lockbox: support staff can access workspaces only with explicit customer permission; Security Center lets customers audit and revoke that access and export historical support-access events. Privacy policy says Relativity is a processor of Customer Data and does not independently access it, which is a general statement rather than a staff-access procedure.
Ask: Confirm per-ticket consent, logging visible to us, whether Azure or OpenAI staff can read prompts, and the legal-process notification clause before producing our content.
SSO, MFA, permissions and ethical walls
Public claim only
Authentication supports OpenID Connect, SAML 2.0, password MFA (authenticator app or email) and trusted IP ranges. Product security describes Security Center, two-factor tracking, inactive-user disablement, Lockbox and discrete permissions. Relativity workspaces and client-domains isolate matters. Native SCIM was not documented on the pages reviewed. Permission-aware retrieval for aiR Assist/claiR at query time is asserted by platform design (AI comes to data in place) but not evidenced by a published permission-leakage test.
Ask: Demonstrate SSO-only with our Entra conditional access, immediate session revocation, matter/ethical-wall enforcement on every aiR retrieval (including caches and Assist), and any SCIM option.
Encryption and keys
Public claim only
Product security states data is encrypted at rest and in transit by default, with an optional customer-managed key. Help documentation includes a customer-managed keys topic. Algorithms, TLS version, coverage of embeddings/logs/backups, and HSM details are not specified on the marketing security pages.
Ask: Confirm TLS 1.2+ and AES-256 (or equivalent) across databases, object storage, indexes, embeddings, logs, backups and the Azure OpenAI link, key rotation, and whether CMK is available in Australia East.
Independent security assurance
Public claim only
The compliance page lists ISO/IEC 27001:2022, ISO/IEC 27018:2019, FedRAMP Class C, HIPAA, IRAP Assessed at Protected, SOC 2 Type II, SOC 3 and CSA CAIQ, with certificates available on request to securityquestions@relativity.com. An Australian government landing page still cites ISO/IEC 27001:2013. Report dates, exceptions and whether aiR generative features sit in scope were not inspected. ISO/IEC 42001 is not listed.
Ask: Provide the current SOC 2 Type II report and ISO 27001 certificate under NDA, with scope covering Relativity aiR generative features and Azure OpenAI, plus IRAP letter of completion and any ISO 42001 status.
DPA, confidentiality and liability
Not confirmed
The public Terms of Use govern the marketing website, not the product. Privacy policy states Relativity is a processor of Customer Data uploaded to the Software. No master services agreement, DPA, output-ownership clause, IP indemnity, liability cap, governing law or audit-rights schedule was found on the public site.
Ask: Provide the MSA and DPA for Relativity aiR, including APP-aligned processing, confidentiality of prompts and outputs, output ownership and IP indemnity, liability caps and Australian law/forum options.
Cross-border processing
Public claim only
For an Australia instance, the 18 September 2026 geography table lists Australia as the aiR inference location, with a warning that inference 'may take place outside' the instance region depending on model availability and that locations may change. Storage remains in the selected region. Relativity is US-headquartered; website personal data may be processed in the United States. MCP-connected assistants may send data to other providers' regions.
Ask: For our Australia East instance, contractually require in-region storage and inference, list every category that may leave Australia, and bind subprocessors including any MCP provider.
Breach notification and incident history
Not confirmed
Website privacy policy says Relativity may notify of a security breach consistent with applicable law, without a numbered hour limit or NDB cooperation clause. No public status page, three-year incident register, or Relativity-attributed customer data breach was located in this review. Certificates are requested via securityquestions@relativity.com.
Ask: What contractual notice period (from suspected incident) applies to all customer data, including subprocessors, what will the notice contain, and will you cooperate at your cost with our Notifiable Data Breaches assessment?
Model and feature change notice
Public claim only
Help documentation says Relativity regularly updates aiR systems (instructions, AI technology and features) and will communicate changes through release notes, knowledge-base articles and/or official announcements. Inference locations 'may change based on AI model availability'. Rate-limiting of AI features is reserved in Relativity's discretion. No contractual pin, testing window or per-output model version identifier was found.
Ask: Will you give at least 30 days' written notice before a material model, prompt or region change, record the model version on each output, and let us re-run acceptance tests before the change becomes default?
Export, termination and portability
Not confirmed
The platform is a system of record with native productions, APIs and workspace archive/restore. Public pages do not state a post-termination export window, open-format bulk export of prompts, aiR outputs and audit logs, certified deletion, or escrow. Website Terms of Use do not cover product exit.
Ask: On termination, how long can we bulk-export documents, productions, prompts, aiR outputs, coding, audit logs and configuration, in what formats, and when is remaining data deleted and certified?
Collection, APIs and MCP connectors
Not confirmed
The platform collects from 'the most widely used enterprise platforms, including chat and enterprise AI tools', exposes REST APIs, an App Hub, and MCP so external assistants can stand up workspaces and manage operations. Each connector can move matter data to another processor. Permissions and data maps for each connector are not published as a single list.
Ask: List every connector and MCP tool we would enable, the identity and scopes used, what data leaves Relativity, and how we disable a connector without vendor involvement.
Open-web retrieval
Public claim only
Relativity's positioning is that AI comes to data already in the platform rather than copying it to a separate web-search tool. aiR Review is documented as using Relativity instructions, prompt criteria and extracted document text (and, for some analyses, images). No product-level open-web search toggle was found. MCP-connected assistants may have their own web tools outside Relativity.
Ask: Confirm that aiR Review, Assist, Case Strategy and claiR cannot call the open web for our tenancy, and how MCP assistants are prevented from sending matter text to web-search tools.
Agentic actions and kill switch
Not confirmed
MCP is described as allowing external AI applications to stand up workspaces, organise case data and manage day-to-day operations in natural language. Pricing says Relativity aiR remains the system of record. Security Center can disable users and revoke Lockbox access. A published per-action approval matrix for MCP tools (send, delete, produce, edit) and a tenant-level kill switch with timed effect were not found.
Ask: Which MCP and aiR actions can run without a lawyer confirming each one, how do we disable MCP and agentic features immediately, and what happens to in-flight jobs and caches?
Human oversight of AI output
Public claim only
aiR Review returns citations, rationales and 'considerations' on possible prediction errors, and verifies citations per document (up to five for relevance). Case Strategy documentation requires a human to oversee work, warns of hallucinations and non-determinism, and says attorneys or administrators must verify outputs. Relativity does not claim outputs may be filed or relied on without review. Mandatory approval gates before production or client-facing use are a workflow design choice, not a published hard control.
Ask: Can we require named reviewer sign-off before aiR coding, privilege logs, chronologies or Gavel drafts leave the workspace, and can we export a per-document record of who verified what?
