Product and intended use
Define the configuration, people, information and tasks being assessed.
Product and plan covered
Public claim only
The vendor groups six features under an AI suite heading: Deep Dive, Coding Suggestions, Writing Assistant, Review Assistant, Predictive Coding and Clustering. The AI governance framework organises them into Review Assistant (batch document summaries, summaries by topic within a document, coding suggestions, sentiment analysis, people and organisation extraction, date and number extraction, page and line references) and Writing Assistant in Storybuilder (draft summaries, document references in drafts). The Customer Terms define 'Everlaw AI Assistant' far more widely than that list: 'any machine learning capability or functionality made available by Everlaw, including without limitation any capability or functionality labeled or otherwise identified by Everlaw as AI-powered'. The FAQ says an organisation obtains the suite through its Customer Success Manager.
Ask: List on our Order Form every feature that falls inside the contract definition of Everlaw AI Assistant today, and tell us how we learn when a new feature is added to that definition.
Instance choice and hosting regions
Public claim only
DPA section 10.1 lets the customer choose which of Everlaw's available AWS instances stores and processes its Customer Data, and the same clause says Everlaw may still need to transfer and process data in the US and anywhere else Everlaw or its sub-processors maintain data processing operations. DPA Schedule 3 lists AWS hosting in the United States (US East and West), Canada (Central), Australia (Asia Pacific, Sydney Region), the United Kingdom (London) and the EU (Frankfurt). The service is cloud only on the documents read; a separate FedRAMP-authorised version is referred to in the Customer Terms for controlled data.
Ask: Will our production database sit in the AWS Sydney instance, and which processing for the AI features, support, monitoring, backup and disaster recovery leaves that instance?
How the AI features are built
Public claim only
The AI governance framework (read 27 September 2026) states that Everlaw builds no models: 'Everlaw does not develop generative AI systems or LLMs independently. We also do not train or fine-tune LLMs developed by others.' Text generation calls third-party enterprise LLMs per request, all features use grounding so that task-relevant customer text is placed in the model's context window, and results cite the source documents. Deep Dive additionally stores vector embeddings of customer documents in a vector database, which the framework says does not have access to the readable contents of the data. Predictive Coding and Clustering are described in machine-learning rather than generative terms.
Ask: For each of the six features, state which are generative (a third-party LLM call) and which are not, and whether Predictive Coding and Clustering send any document text outside our instance.
Australian instance versus where AI processing happens
Public claim only
Australia is one of the AWS instances a customer may choose for storage and processing (DPA section 10.1 and Schedule 3; sub-processor list of 16 September 2026). The AI features are routed differently. The AI FAQ, last updated 12 August 2025, states that for clients in the EU and UK environments most AI tasks are processed in the EU region, that Deep Dive uses a third-party vector database that processes data in the US even for those clients, and that 'For all other clients using Everlaw AI Assistant, data is sent to servers in the US.' The sub-processor list places OpenAI, L.L.C. in the United States (EU only for UK or EU instances) and Pinecone Systems, Inc. in the United States with no regional alternative.
Ask: For an Australian instance with the AI suite enabled, confirm in writing that inference and vector storage occur in the United States, and state whether any Australian or EU routing option exists for Australian customers.
Check for vendor content directing how AI describes the product
Not confirmed
A deliberate check for content addressed to AI models, assistants or crawlers found none on this vendor, checked 27 September 2026. A site map returned 40 URLs and none was /llms.txt, /ai-info, /ai or /for-ai, and none of the six pages read for this profile (trust centre AI page, AI governance framework, sub-processor list, Customer Terms, AI FAQ, DPA) contains a block instructing AI models how to describe Everlaw. This is a negative result on the URLs returned and the pages read, not proof of absence.
Ask: Do you publish, anywhere on your sites or in your documentation, guidance addressed to AI models, assistants or crawlers about how the product should be described?
Data, controls and requirements
Compare the actual configuration with firm, client and jurisdiction requirements, supported by evidence.
Client authority and permitted content
Not confirmed
The obligation to have the right to send material to the AI features sits with the firm. Customer Terms section 3.4 requires the customer to obtain and maintain all necessary consents before uploading or processing Case Materials, and section 8.2 warrants that it has all necessary rights and authority. DPA section 4.2 repeats that the customer has provided all notices and obtained all consents necessary for Everlaw to process the data. Section 3.2 of the Terms bars classified government data, export-controlled data, criminal justice information and protected health information from the Service unless a signed Order Form puts it on the FedRAMP-authorised version or a HIPAA business associate agreement is executed. Administrators opt in to AI features; no published control excludes a single client or matter from the AI suite.
Ask: Can the AI suite be disabled for a named client or a named matter while other projects keep using it, and what written description of US processing can we give a client who asks?
Model and AI providers
Public claim only
The sub-processor list dated 16 September 2026 names Amazon Web Services, Inc. and Google LLC as 'Cloud and AI Service Provider', OpenAI, L.L.C. as 'AI Service Provider' processing in the United States (EU only for customers on the UK or EU instances), and Pinecone Systems, Inc. as the vector database for vector search features in the United States. The AI governance framework does not name a provider in its body, but its glossary states that 'Everlaw's zero data retention agreement with OpenAI means no user data is stored or used to train OpenAI's LLMs or for any other purpose' and gives Pinecone as an example of a vector database. The framework also reserves flexibility 'to shift between third-party service providers as needed'. No model name, model version or customer model-selection control appears on the pages read.
Ask: Name each model and version, and the provider entity and endpoint region, that will serve our instance for each AI feature, and confirm whether we can refuse a provider or pin a version.
Sub-processor list, notice and objection
Public claim only
A dated list (16 September 2026) with a downloadable PDF and eight archived versions names AWS, Google LLC, Zendesk, Inc. (only where the customer attaches data to support communications), OpenAI, L.L.C., Pinecone Systems, Inc. and Datadog, Inc., plus the affiliates Everlaw, Inc. and Everlaw UK Ltd. Security, endpoint detection and response sub-processors are expressly not disclosed on the website and must be requested by email. DPA section 11.1 requires notice at least 10 days before adding or replacing a sub-processor; section 11.3 gives the customer 15 calendar days to object in writing on reasonable data-protection grounds, and if no resolution is reached Everlaw will at its sole discretion either not appoint that sub-processor for that Customer Data only or permit the customer to suspend or terminate the affected Service without liability to either party. Silence is deemed approval.
Ask: Supply the full list including the undisclosed security and endpoint tools for our instance, and confirm the 10-day notice and 15-day objection rights are not narrowed by our Order Form.
Training on customer content
Public claim only
Customer Terms section 8.5 states: 'Everlaw will not use Input or Output to train or finetune a generative AI model for general use through the Service.' The customer retains intellectual property rights in Input and Output. Two qualifications sit beside that clause. Section 5.1 grants Everlaw a worldwide licence to use Case Materials to provide the Service and 'to support, manage, and improve the Service', subject to the section 8.5 limits. DPA section 3.2(D) permits Everlaw to de-identify or anonymise Customer Data in the course of providing the Service for its internal business purposes, provided it does not attempt re-identification. The AI FAQ adds that the data submitted and responses received are not used to fine-tune or improve any of the providers' models, are not stored by them, and are not used to train models across customers or shared between customers.
Ask: Confirm in the signed agreement that prompts, documents, outputs, embeddings and de-identified derivatives are not used to train or evaluate any model, Everlaw's or a provider's, including models that are not 'for general use'.
Feedback, telemetry and service improvement
Public claim only
Customer Terms section 5.3 says Everlaw and its affiliates may use Feedback 'without restriction and without obligation to Customer or Authorized Users'. Section 13.1 says beta services are made available to collect Feedback and that by using them the customer agrees Everlaw may make contact to collect it. Service-Generated Data, defined as all data generated or derived from interaction with the Service, is excluded from Case Materials and is handled under the Privacy Notice rather than the AI clause. The DPA describes processing to provide, maintain and improve the Service, including analytics native to the platform. No in-product rating control for AI outputs is described, and no opt-out from improvement use of telemetry was found.
Ask: Are reviewer corrections to Coding Suggestions, Deep Dive queries and AI usage telemetry used to tune prompts, retrieval or evaluation sets, and can that use be switched off for our tenancy?
Retention of prompts, outputs, embeddings and logs
Public claim only
The framework states that Everlaw keeps customer data only for as long as the customer stores it on the platform, and that its LLM providers commit to zero data retention so that data sent for processing is deleted once the response is returned; the FAQ repeats that none of the models store submitted data or responses. Deep Dive is the exception the vendor itself flags: a vector database stores embeddings created from customer documents, and the sub-processor list places that database in the United States. Everlaw also logs each AI request with the input, output and prompt where applicable, plus the user, time and documents affected, and surfaces this through a dashboard and for developer retrieval for auditing. No retention window is published for embeddings, AI request logs, Datadog monitoring data or backups, and the DPA states no numeric retention period.
Ask: Give a class-by-class retention schedule for prompts, outputs, vector embeddings, AI request logs, monitoring data and backups on our instance, including any provider-side abuse-monitoring window.
Deletion during the term and after termination
Public claim only
DPA section 6.1 says the customer may delete or export Customer Data during the subscription in a manner consistent with the functionality of the Service, and that deleting it is an instruction to Everlaw to delete it from Everlaw's systems or anything otherwise in its possession or control. Section 6.2 makes termination or expiry an instruction to delete all Customer Data including copies, with a 30-day grace period for export. Section 6.3 says Everlaw will comply within a commercially reasonable timeframe except where law requires retention. The DPA expires automatically on deletion of all Customer Data. No maximum number of days, no backup-purge lag and no certificate of deletion are specified, and deletion of vector-database indexes is not addressed separately.
Ask: State the maximum days to purge primary stores, vector indexes, logs, backups and each sub-processor after a deletion instruction, and will you certify completion in writing?
Vendor personnel access to content
Public claim only
DPA section 7.1 requires anyone Everlaw authorises to process Customer Data to be under an obligation of confidentiality. Schedule 3 adds that Everlaw maintains a record of personnel authorised to access systems containing Customer Data, that privileged access requires a formal account management and access control procedure with review and approval from a line manager or other executives, that credentials are deactivated promptly on termination or a role transfer that no longer requires access, that personnel are legally obliged to maintain confidentiality after their engagement ends, and that employees complete annual training covering data protection and information security. Support data reaches Zendesk only where the customer attaches or uploads it. The documents read do not say that support access is granted per ticket with customer approval, or that access logs are available to the customer.
Ask: Is support or engineering access to our case materials approved per ticket by us, time-boxed, and available to us as an access log, and from which countries can support staff read our data?
Access control, single sign-on and sessions
Public claim only
DPA Schedule 3 states that Everlaw uses role-based access control, multi-factor authentication and account management procedures to control access to Customer Data, that customers may set their own password complexity when they enable single sign-on via the SAML 2.0 protocol, that web application passwords are hashed using bcrypt, that sessions expire after 30 minutes of inactivity and hold a session lock until the user re-establishes access. Customer Terms section 3.1 gives tiered Administrator Access at Database and Project level, including the right to restrict access and to opt in to AI features. SCIM provisioning and enforced SSO-only login are not mentioned in the documents read.
Ask: Demonstrate enforced SSO-only login, immediate session revocation on deprovisioning, any SCIM support, and that project permissions are enforced at query time for Deep Dive.
Encryption in transit and at rest
Public claim only
DPA Schedule 3 states that Customer Data at rest is encrypted using AES-256, that Everlaw uses TLS 1.2 or higher for HTTPS communications, that it requires full-disk AES-256 encryption on all employee computers, that data in AWS is encrypted at rest as set out in AWS documentation, that encryption keys are rotated at least annually, and that email uses opportunistic TLS by default, with hybrid techniques aligned to NIST Special Publication 800-53. The AI FAQ says each AI data request is sent to a sub-processor individually over an SSL encrypted service. Customer-managed or customer-held keys are not offered in the documents read.
Ask: Confirm the encryption applied to vector embeddings, AI request logs and backups, and whether customer-managed keys are available on the Sydney instance.
Independent assurance and audit rights
Public claim only
The DPA defines 'Report' as Everlaw's Service Organizational Control (SOC) 2 Report or comparable reports, and section 8.1 requires Everlaw to supply a summary copy on written request on a confidential basis, and to complete a customer-provided information security questionnaire once per calendar year on 90 days' notice. Section 8.2 allows an audit or inspection once per calendar year at the customer's expense during normal business hours. Schedule 3 adds annual penetration testing, vulnerability scanning by an independent testing team, and annual auditing of the incident tracking system. The SOC 2 type, the report's scope and whether it covers the AI features are not stated, and no ISO or AI-management-system certification appears in the documents read; the separate security page was not opened inside the credit budget.
Ask: Provide the current SOC 2 report and any ISO 27001 or ISO/IEC 42001 certificates, with scope that names the AI features and the AI sub-processors, plus the latest penetration-test summary.
Contracting entity, governing law and liability
Public claim only
The Customer Terms state that the agreement is between 'Everlaw, Inc., a Delaware corporation' and the party named as Customer in the Order Form. Section 11.1 applies the law of the State of California; section 11.3 sends disputes to binding arbitration administered by the American Arbitration Association after informal resolution, section 11.4 reserves injunctive relief to the federal or state courts of Alameda County, California, and section 11.5 waives class and representative actions. Section 10.3 caps each party's aggregate liability at the fees paid or payable in the 12 months preceding the date the liability first arises, and at US$100 for an unpaid subscription; section 10.2 excludes indirect, consequential and similar damages including loss of data. Section 9.1(D) removes Everlaw's intellectual-property indemnity where the claim alleges that Output from the AI suite infringes a third party's rights. No Australian entity, law or forum is published.
Ask: Will you contract through an Australian entity under Australian law and forum, lift the 12-month cap for a confidentiality or data breach, and indemnify us for third-party rights claims about AI output?
Cross-border transfers and Australian Privacy Principle 8
Public claim only
DPA section 10.1 lets the customer pick the AWS instance but says Everlaw may transfer and process data in the US and anywhere else it or its sub-processors operate. Section 10.2 relies on the Data Privacy Framework for transfers from the EEA, UK and Switzerland to the US, and section 10.3 substitutes the Standard Contractual Clauses if that framework is invalidated or does not apply; no equivalent mechanism is named for Australia, and Australian privacy law is not mentioned in the DPA read. The AI FAQ says data is sent to servers in the US for all clients outside the EU and UK environments. The sub-processor list places OpenAI, Pinecone, Datadog and Zendesk processing in the United States.
Ask: For an Australian customer, list every overseas recipient of our case materials with the AI suite enabled, and what contractual steps you will take for Australian Privacy Principle 8 purposes.
Security incident notification
Public claim only
DPA section 7.2 says that upon becoming aware of a verified Security Incident Everlaw will notify the customer without undue delay, provide timely information as it becomes known or is reasonably requested, and promptly take steps it deems necessary and reasonable to contain, investigate and remediate. The clock therefore starts at a verified incident rather than a suspicion, and no hour figure is stated. 'Security Incident' is defined as a breach of Everlaw's security leading to accidental or unlawful destruction, loss, alteration, unauthorised disclosure of or access to Customer Data processed by Everlaw or its sub-processors. Schedule 3 says the incident process aligns with NIST SP 800-61, that incidents are logged in a tracking system audited annually, and that an incident response test is run annually. Customer Terms section 3.4 makes the customer report suspected incidents to security@everlaw.com. Neither document mentions the Notifiable Data Breaches scheme or the OAIC.
Ask: Will you commit to notice within 72 hours of discovering a suspected incident affecting our data, including at a sub-processor, with the facts we need for a Notifiable Data Breaches assessment?
Notice of changes to models, sub-processors and controls
Public claim only
DPA section 11.1 gives at least 10 days' notice before a sub-processor is added or replaced. DPA Schedule 3 states that Everlaw 'may update the Security Measures from time to time and without notice, provided that such updates and modifications do not materially decrease the overall security of the Service'. The AI FAQ says Everlaw continually evaluates advances in foundation models, may engage new sub-processors from time to time, and will notify customers accordingly. The framework says its development approach allows Everlaw to shift between third-party service providers as needed, and that new features and improvements are released monthly. Customer Terms allow fee changes on 30 days' notice before a Renewal Term. No model-version pin, no per-output model identifier and no pre-change testing window are published.
Ask: What written notice will we get before the model or vector-database provider serving our instance changes, and can we test the new configuration on a sample before it becomes the default?
Export and exit
Public claim only
DPA section 6.2 gives a 30-day grace period after termination or expiry for the customer to export all Customer Data, alone or with Everlaw's support, after which the deletion instruction takes effect. Customer Terms section 12.5 says that when an Order Form ends all rights and access to the Service terminate, including access to Case Materials, and that Everlaw will send a final invoice. Section 12.2 requires 30 days' notice of non-renewal. The documents read do not name export formats, do not say whether AI outputs, prompts and audit logs are included in an export, and contain no escrow or certified-deletion undertaking.
Ask: Which formats and which record classes are covered by the 30-day export, including AI outputs and the AI usage audit log, and what does an export cost?
Connectors and onward exposure
Not confirmed
For the AI suite itself, the framework describes a closed loop: tasks are one-time generation requests to a model, with no link-following or web-browsing, and Deep Dive's only additional component is the vector database sub-processor. Collection connectors, application programming interface access and connections that let external assistants reach case materials are platform-level matters and were not examined for this profile; they are recorded against the platform profile. The sub-processor list is the only published enumeration of third parties that receive customer data read here.
Ask: List every connector, application programming interface and external assistant connection that would be enabled for us, what each one receives, and how we disable each without a support ticket.
Open-web access from the AI features
Public claim only
The AI governance framework states that 'Everlaw AI tools do not offer features or functionality that allow link-following or web-browsing based on interactions with the LLM', and that all features use grounding, in which task-relevant information from the customer's own documents is placed in the model's context window rather than relying on the model's embedded knowledge. It adds that many tasks take no user input at all, which it presents as reducing prompt-injection exposure. These statements are in a vendor governance document, not in the Customer Terms or the DPA.
Ask: Will you warrant in the agreement that no AI feature can retrieve from the open web on our projects, and how is that enforced technically rather than by design intent?
What the AI features can do without a person acting
Public claim only
The framework says users can opt in and out of the AI features, that a recognisable icon always indicates when generative AI is in use, and that 'every existing task in Everlaw AI is run as a one-time generation request', or for bulk work a series of single requests. Coding Suggestions proposes codings with a rationale and in-text citations rather than applying them as final. Every request is logged with input, output, prompt, user, time and documents affected, and surfaced in a user-facing dashboard. The documents read describe no agent that plans and executes multi-step actions, and no tenant-level kill switch is documented.
Ask: Can any AI feature write to a coding field, produce, share or export without a person confirming the action, and how do we freeze the whole suite immediately if needed?
Human review before output is used
Public claim only
Customer Terms section 8.5 states that the customer is responsible for its use of the AI suite, acknowledges that Output is probabilistic and may not be accurate, reliable or otherwise appropriate, and 'should independently evaluate Output for accuracy and appropriateness for Customer's use case(s), including through human review, and modify Output as necessary before relying on, publishing, or otherwise using Output'. Section 8.4 disclaims any warranty as to the accuracy, completeness or reliability of Output. The framework says Everlaw encourages human verification of all generative outputs and builds verification pathways into the interface. No product control forces a named reviewer to sign off before an output is exported or produced.
Ask: Can we require named reviewer sign-off before an AI output is exported, produced or placed in a draft, and can that requirement be enforced in the product and evidenced in an audit export?
