Product and intended use
Define the configuration, people, information and tasks being assessed.
Product identity and current name
Public claim only
The product is carried in the vendor's own documentation at help.relativity.com, read 2026-09-27, under the heading 'Assist (aiR Assist)' with the page last modified 18 September 2026. The name changed in the vendor's September 2026 rebrand: 'aiR Assist is now Assist. References to aiR Assist may still appear in this documentation, in the product, and in other sources such as marketing materials, published content, and commercial or legal agreements. Both names refer to the same product/feature.' The platform documentation lists Assist as one of five aiR systems.
Ask: Which name does our current agreement use, and does the rebrand change anything in the schedule that licenses the feature?
Release stage
Public claim only
Documented as generally available rather than pre-release: the documentation describes the contracts under which Assist is available, the permissions to configure, the index limits and the audit events, and states 'It does not require installation in Relativity aiR.' A dated observation, recorded without inference: the platform page 'AI products and features in Relativity aiR', last modified 25 September 2026, lists five aiR systems (aiR Review, Privilege, Case Strategy, Data Breach Response and Assist) and does not mention Relativity claiR, which the vendor separately describes as being in an Advanced Access pre-release programme.
Ask: Is Assist generally available to Australian customers today on the same documentation and terms as elsewhere, and is any part of it still gated behind an early-access programme?
What the feature does and how it is documented to work
Public claim only
Documentation read 2026-09-27: 'Assist is a conversational search tool integrated within a Relativity aiR workspace, designed to empower legal teams to interact with their data using natural language.' Mechanically, 'It works by searching the extracted text of indexed documents. Users can create up to five indexes per workspace, each supporting up to 300,000 documents. When a query is submitted, Assist identifies the documents deemed most relevant and employs a large language model (LLM) to generate answers grounded by citations tied back to the documents.' The published architecture is retrieval-augmented generation in five steps — question, analysing, search, generation, answers — in which the LLM 'decides whether the question is an appropriate one to answer', formulates the queries run against the index, and then generates the cited answer. Documented limits: individual documents up to 5 MB of extracted text, larger files excluded during indexing; only documents with extracted text stored in Data Grid rather than SQL are indexed; a maximum of five built indexes per workspace.
Ask: Which of these documented limits are configurable for us, and what happens to a matter that needs more than five indexes or more than 300,000 documents in one index?
Deployment and where the work happens
Public claim only
Assist runs inside an existing Relativity aiR workspace on data already held there and needs no installation: 'It does not require installation in Relativity aiR.' The text it searches is the extracted text held in Data Grid rather than SQL. Inference is separate from storage: the platform documentation states that aiR systems 'use AI models and conduct AI processing using Microsoft Azure services unless otherwise specified in the terms of your Agreement with Relativity', and that such processing 'may take place outside the Microsoft Azure region where your Relativity aiR instance is hosted'. Only a cloud deployment is documented; the vendor separately publishes a Server Policy page, which was not read, so whether any self-managed deployment carries Assist is not established.
Ask: For our instance, in which Microsoft Azure region does the Assist index sit, and in which region does each of the retrieval and generation steps run?
Jurisdiction, legal content and language
Public claim only
Assist carries no legal-content library: everything documented points at the customer's own indexed workspace documents, and the FAQ states that 'Assist is designed to respond to prompts related to the documents in the current workspace and not general knowledge prompts.' Nothing read claims coverage of Australian legislation, case law or rules of court, or addresses Australian discovery practice. On language the vendor states 'The LLM used by Assist has been evaluated for 83 languages. Although Assist has primarily been tested on English-language documents, it is designed to support non-English datasets', and documents that retrieval defaults to English so a user must add an instruction such as 'Search in <Language>' for other languages, with the caveat that 'the LLM may not always apply the requested language consistently.'
Ask: Who performed the 83-language evaluation — Relativity or the model provider — and what testing has been done on Australian-English document sets?
Vendor content directing how AI describes the product
Not confirmed
A deliberate check was made on 2026-09-27 and nothing of this kind was found. Checked: the three help.relativity.com pages read for this profile, the vendor's Artificial Intelligence Principles page, the Calder7 product security and compliance pages, the privacy policy, and a cached 40-URL map of relativity.com. No page carried an 'Instructions for AI Assistants' heading or any equivalent block directing how AI models, assistants or crawlers should describe the product, and the map returned no /llms.txt, /ai-info, /for-ai or /ai.txt. No credit was spent guessing URLs, so the absence of such a file at an unmapped path is not established. A related check was also made and recorded: nothing read answers a comparative or superlative question about the vendor's own product; the only superlative-looking strings on the pages read are links titled 'Best practices' and the sentence 'The RAG process works best when key evidence is found in a few focused documents.'
Ask: Does Relativity publish any file or embedded block that tells AI assistants or crawlers how to describe Assist or the aiR platform, and if so where?
Data, controls and requirements
Compare the actual configuration with firm, client and jurisdiction requirements, supported by evidence.
Who can interrogate which documents
Not confirmed
Two documented design points a firm must reconcile with its undertakings. Answers are bounded by the asking user's document permissions — 'Assist responses are based only on the documents the requesting user is allowed to access at the time they ask the question' — and the vendor adds that 'every interaction is evaluated against the requesting user's document security permissions at the time of the prompt'. But the scope of what can be asked about is set by someone else: an index is built by a user with Index Management permission from public saved searches or the Case Home document set, and the person with Prompting permission then queries whatever that index contains. On a matter holding another party's produced documents or material under a confidentiality regime, deciding which saved searches may become an Assist index is a client-authority question rather than a technical one.
Ask: Who in our firm approves which saved searches may become an Assist index on a matter holding another party's produced documents, and can an index be restricted to exclude documents subject to a confidentiality undertaking?
Foundation-model provider
Public claim only
Named at platform level, not per system. The platform documentation, last modified 25 September 2026, states: 'Relativity's aiR systems currently leverage a combination of generative AI LLMs including GPT-4o, GPT-4.1 mini, GPT-5 mini, GPT-5.1, and GPT-5.2 through Microsoft Azure services. The specific LLM or combination of LLMs used by an aiR system depends on the aiR system objective, and the capability and availability of an LLM that meets Relativity's minimum product requirements.' Assist uses a model at two separate steps, which the vendor distinguishes: 'Retrieval—the LLM generates search queries to find relevant documents in the index. Generation—the LLM summarizes the retrieved documents, picks relevant ones, and synthesizes them into an answer.' Nothing read states which model serves either step on a given day, and no customer model choice is described anywhere read, so none is recorded.
Ask: Which named model serves the Assist retrieval step and which serves the generation step for our instance today, and what written notice will we get before either changes?
Subprocessors
Not confirmed
No subprocessor register was found. The cached 40-URL map of relativity.com shows no subprocessor page, and neither Calder7 page nor the privacy policy carries a register. Microsoft Azure is the one named infrastructure and model-hosting provider, from the statement that aiR systems 'use AI models and conduct AI processing using Microsoft Azure services'. The privacy policy addresses third parties only in general terms and, critically, excludes the data a firm cares about: 'Data uploaded by our customers into the software is not covered by this policy and you should contact your software administrator if you have questions about how Relativity processes such data.'
Ask: Provide the subprocessor register that applies to Assist, naming each entity, its country and its function, including the model-hosting entity for retrieval and for generation, with our notice and objection rights.
Training on customer content
Public claim only
The vendor publishes one statement on the point, in the platform documentation last modified 25 September 2026: 'Customer data processed using aiR systems is not used to train or improve third party models or services.' Its scope is third-party models and services. Nothing read states whether Assist prompts, answers, conversations or indexed content are used to train, fine-tune or evaluate Relativity's own models. The vendor's Artificial Intelligence Principles page, read the same day, refers to models Relativity itself trains and states 'We utilize data minimization in the models we train, using personal data only to the extent necessary to fit the purpose of the model and help users efficiently organize information', without saying whose data those models are trained on. The privacy policy lists 'improving our software, websites and related services' among its purposes but expressly excludes customer-uploaded data.
Ask: Will you warrant in our agreement that no Assist prompt, answer, conversation, index or document chunk from our workspace is used to train, fine-tune or evaluate any model, including Relativity's own models, and not only third-party models?
Feedback and quality monitoring
Public claim only
The FAQ, last modified 17 September 2026, documents two feedback routes: 'use the thumbs up or down icons below each response to rate it and add comments', and an occasional 'Assist Feedback dialog asking how helpful were Assist's responses for completing your task'. The vendor links feedback to product improvement: 'Using the feedback buttons in the response box can help improve the tool over time.' Nothing read states where that content goes, whether the free-text comment and the associated prompt and answer leave the instance, who can read them, how long they are kept, or whether the control can be turned off for a workspace.
Ask: When a user types a comment into the Assist feedback box, does that text and the associated prompt, answer and citations leave our instance, who reads it, how long is it kept, and can we disable the control?
Retention of prompts, answers, conversations and indexes
Public claim only
The FAQ states the retention rule for conversations: 'Your conversation history is saved across your user sessions and retained as long as the workspace exists, unless the conversation is deleted.' Conversations are private to the user who created them and cannot be shared inside the product. Indexes persist until rebuilt or deleted, to a maximum of five built indexes per workspace. The audit trail records that a question and an answer occurred but not their content: 'Audit does not record the content of the question' and 'Audit does not record the content of the answer'. No retention period is published for the index itself, for the document chunks it holds, for metadata mapping, or for any prompt or completion data held at the model endpoint. The only published retention sentence outside the product documentation is in the privacy policy, which excludes customer-uploaded data from its scope.
Ask: State the retention of Assist indexes, document chunks, metadata mapping and any prompt or completion data held outside the workspace, and confirm whether we can set a shorter retention for conversations than the life of the workspace.
Deletion and certified erasure
Public claim only
Deletion is documented as a user action rather than a vendor commitment. A user 'can delete an entire conversation and its history using Conversation Manager', a user with Index Management permission can delete an index, and both deletions appear in the audit trail as Assist Conversation Delete and Assist Index Delete events. What is not published: whether deleting a conversation erases the underlying prompt and completion records wherever they are held, what is deleted when a workspace is closed or a subscription ends, over what period, and whether any erasure is certified. One documented limit matters on exit and on transfer: 'ARM is not supported for Assist. Assist permissions, indexes, conversations, and metadata mapping cannot be archived, restored, or moved using ARM.'
Ask: When we delete an Assist conversation or index, what is erased, where, and over what period — and what is deleted on workspace closure, with written certification?
Vendor human access to content
Public claim only
The Calder7 product security page publishes a control aimed at this: 'Customer Lockbox — Prevent administrators – including those from Relativity – from seeing your data unless you say so', with Security Center able to 'Perform real-time audits of Lockbox to enable and revoke access to individuals as needed'. That is a security page rather than a contractual document; no standing approval process, access-log format or retention of access records is described; and nothing read states whether Lockbox covers Assist conversations, indexes, metadata mapping or prompt data as well as the underlying documents.
Ask: Is Customer Lockbox available on our plan, does it cover Assist conversations, prompts, answers and indexes as well as matter documents, and what audit record do we receive of any Relativity access?
Access control, permissions and audit
Not confirmed
Assist has its own named permissions set per group per workspace — Access Management, Index Management and Prompting — and the FAQ confirms these can be separated, so a user may be given Prompting without Index Management or the reverse. One default requires a decision: 'By default, System Administrators are granted access to all Assist permissions across all workspaces.' Platform-level controls on the security page are 'OpenID Connect and SAML2 protocols, two-factor authentication, and single sign-on', with 'discrete permissions to give you total control'. Auditing records that an event occurred rather than its content, and the vendor describes it as temporary: the Audit application records Assist Question, Answer, Conversation, Index, Indexing Error List and Metadata Mapping events without the content of questions or answers, and the vendor calls this 'a temporary solution' that it does 'not recommend building any integrations' on, with the data 'planned to be moved to Custom Reports by the end of 2026'.
Ask: How do we restrict the default under which every system administrator holds all Assist permissions in every workspace, and how do we evidence who asked what before the audit data moves to Custom Reports?
Encryption and customer-managed keys
Public claim only
The Calder7 product security page states 'By default, Relativity encrypts data at rest and in transit' and offers 'Customer-managed keys' which 'add an extra layer of protection, giving you total control over your environment's encryption key'. No algorithm, protocol version or key-management detail is published; the statement sits on a security page rather than in a contract; and nothing read states whether customer-managed keys extend to the Assist index, the document chunks it holds, stored conversations, or data in transit to the model endpoint.
Ask: Name the encryption algorithms and key management, and confirm in writing that customer-managed keys cover the Assist index, its document chunks and stored conversations, and that traffic to the model endpoint is encrypted.
Independent security assurance
Public claim only
The Calder7 compliance page publishes a certification list, read 2026-09-27: 'ISO/IEC 27001:2022 certification, ISO/IEC 27018:2019 certification, FedRAMP Class C Certified, HIPAA Compliance, IRAP Assessed at Protected, SOC 2 Type II Report, SOC 3 Report, Cloud Security Alliance's (CSA) CAIQ', with certificates available on request to securityquestions@relativity.com. The IRAP entry is the Australian-relevant one. Four limits: no certificate or report was seen, so none is verified here; the page states no scope, date, assessor or exceptions; the FedRAMP authorisation is attributed elsewhere on the same site to Relativity aiR Government rather than the commercial platform; and nothing read states that Assist, or generative-AI processing at a model endpoint, sits inside the assessed scope of any of them.
Ask: Provide the current SOC 2 Type II report, the ISO/IEC 27001:2022 certificate and the IRAP assessment letter with their scope statements, and state in writing whether Assist and its model inference are inside those scopes.
Contract, DPA, entity and liability
Not confirmed
No contractual document was read and none is published on the pages read. There is no subscription agreement, data processing addendum, AI schedule or subprocessor annex on the vendor's public site as mapped; the Terms of Use link in the documentation footer governs use of the website rather than the platform and was not read. The only entity name published on any page read is the site copyright line, '© 2026 Relativity ODA LLC', which is not a statement about who would contract with an Australian customer, and the published address on every page read is 231 South LaSalle Street, 20th Floor, Chicago, IL 60604, with a United States telephone number. No governing law, forum or liability position was found. The documentation does confirm that the agreement is where several Assist facts actually live, including model and processing geography 'unless otherwise specified in the terms of your Agreement with Relativity' and aiR Unit consumption 'as explained in a customer's contract'.
Ask: Provide the subscription agreement and data processing addendum that would cover Assist, name the entity that will contract with us in Australia with its ACN or ABN, and state the governing law, forum and liability cap.
Hosting region and cross-border processing
Public claim only
The vendor publishes a table of aiR instance regions against inference locations, on the platform documentation page last modified 25 September 2026. Australia appears in both columns: instance region Australia, aiR system inference location Australia. The same section sets the general rule and its limits: aiR systems 'use AI models and conduct AI processing using Microsoft Azure services unless otherwise specified in the terms of your Agreement with Relativity'; inference 'may take place outside the Microsoft Azure region where your Relativity aiR instance is hosted'; 'Your data remains stored in the region you selected, only the aiR system AI inferencing may occur elsewhere'; and the table 'reflects aiR System processing locations at the time of publication. These locations may change based on AI model availability and the evolving capabilities of our infrastructure providers.' So an Australian inference location is published, but as documentation that the vendor reserves the right to change, not as a commitment. The privacy policy's own transfer statement — data 'may be processed by us in the United States, by our affiliates elsewhere in the world' with Standard Contractual Clauses where needed — expressly does not cover customer-uploaded data.
Ask: Will our agreement warrant that both the retrieval and the generation steps of Assist run in the Australian region for our instance, and what written notice will we get if the published processing-location table changes?
Incident notification and history
Not confirmed
No incident-notification obligation, period or history was found on any page read on 2026-09-27. The Calder7 pages describe detection and prevention — around-the-clock monitoring, external penetration tests 'multiple ... a year', a bug bounty, 'AI-driven threat detection and response' and Security Center alerts to the customer for unusual events such as a change in login method — but say nothing about notifying a customer of a breach, the period from discovery, incidents at a subprocessor or model provider, or the Australian Notifiable Data Breaches scheme. The privacy policy contains no breach-notification clause.
Ask: State the contractual notification period from discovery for our instance, confirm it covers an incident at Microsoft Azure or at the model endpoint, and set out how you will support our assessment under the Notifiable Data Breaches scheme.
Change to the feature, the model and the audit trail
Not confirmed
The vendor publishes that aiR systems change without a fixed notice period and that outputs change with them: 'Relativity regularly updates aiR systems by modifying Relativity inputs/instructions, AI technology, and other features ... As a result, customers may notice changes in the product experience or differences in outputs after an update, even when the same prompt criteria or workflow is used.' Communication is 'through product release notes, knowledge base articles, and/or official announcements'. Three changes are already documented: the rebrand on 17 September 2026, which renamed every Assist audit object type; the audit records themselves, described as 'a temporary solution' with the data 'planned to be moved to Custom Reports by the end of 2026' and an express recommendation not to build integrations on them; and throughput, which the vendor may vary at will — 'Relativity may, in its sole discretion, rate-limit, queue, prioritize, or restrict usage of these features', which 'Relativity determines and applies ... independently of any job size limits your administrators configure'.
Ask: What written notice will we get before the model behind Assist, its system prompt or the audit record changes, and can an answer given on a matter last year be reproduced after such a change?
Export and exit
Public claim only
One documented export route: 'Conversations can be exported to an MS Word (.docx) file without losing the context, citations, or source links', and the export is audited as an Assist Conversation Export event. Its limits are documented too: individual prompts cannot be shared or exported, conversations cannot be shared with other users inside the product, and nothing read describes a bulk export of every conversation in a workspace, an export of indexes, metadata mapping or the question-and-answer history in a structured format. ARM cannot move Assist artefacts: 'Assist permissions, indexes, conversations, and metadata mapping cannot be archived, restored, or moved using ARM.' Because a conversation may inform advice on a matter, the absence of a bulk export matters for a firm's own file.
Ask: On exit or on a matter transfer, how do we obtain every Assist conversation in a workspace with its citations in one operation, and what happens to indexes and metadata mapping given that ARM does not carry them?
Connectors and whether external tools can drive Assist
Not confirmed
At platform level the vendor publishes Relativity MCP, 'an interface that allows AI assistants access to Relativity aiR capabilities through Model Context Protocol, an open standard for connecting AI tools to external data and services through MCP clients'. Nothing read states whether Assist is among the capabilities exposed, whether an external assistant could submit prompts or build indexes, who may enable the interface, or whether activity arriving that way is recorded as an Assist audit event. The vendor also publishes Gavel and the Relativity App Hub at platform level; neither page was read.
Ask: Can Relativity MCP submit Assist prompts or create indexes on our workspaces, who can enable it, and is MCP-driven activity recorded in the same Assist audit events?
Open-web retrieval
Public claim only
Everything documented points inward. Retrieval searches 'the extracted text of indexed documents'; generation is given 'the selected passages, along with the original question and system prompt'; and the FAQ states 'Assist is designed to respond to prompts related to the documents in the current workspace and not general knowledge prompts.' But no page read states that Assist makes no request to any source outside the workspace at any step, and the model itself is a third-party large language model reached through Microsoft Azure services, so the absence of a statement is not a statement. A user can inspect what the retrieval step attempted: 'You can inspect which searches were attempted using the Search Completed summary in the response panel.'
Ask: Does Assist make any request to a web index, a model-provider tool or any source outside our workspace at any step, and can that be evidenced from the audit trail or the Search Completed summary?
What the model decides and what it writes
Not confirmed
Two decisions are taken by the model without a documented confirmation step. The LLM 'decides whether the question is an appropriate one to answer', and the FAQ confirms rejection happens — 'Assist uses advanced intention capabilities to reject prompts that cannot be answered or are outside its scope. It also requests clarification for vague or overly broad queries.' The LLM also composes the searches actually run against the index rather than executing a query the user wrote. What Assist writes into the workspace is limited to conversation and index objects; it does not code, tag or classify documents. The heavier action is administrative: building or rebuilding an index over as many as 300,000 documents, which is audited but for which no approval gate is documented.
Ask: Is there any approval gate before an Assist index is built or rebuilt over a document set, and where can we see the searches the model chose to run and the prompts it declined?
Human review before reliance
Public claim only
The published control is verification by the user, not an enforced step. The FAQ states: 'Assist relies on indexed data to generate responses leveraging retrieval-augmented generation (RAG). While it aims to offer correct responses, always verify the offered information and citations for accuracy.' Step 6 of the vendor's own getting-started workflow is 'Review and verify responses', and answers are 'grounded by citations tied back to the documents'. Nothing read describes a setting that requires a cited document to be opened before an answer is accepted, records that it was opened, marks an exported conversation as AI-generated, or restricts who may circulate one. The vendor's own conditional framing is recorded as published: outputs may lead to 'defensible workflows when validated by users'.
Ask: Can we require and evidence that a reviewer opened each cited document before an Assist answer is relied on, and is an exported .docx conversation marked as AI-generated?
