Relativity aiR Review: Generative-AI eDiscovery Inside RelativityOne
Relativity aiR is the generative-AI suite that Relativity has folded into its cloud platform, RelativityOne. It is aimed squarely at document review, privilege review, and early case assessment — the workflows where large-language-model reasoning can genuinely compress hours of associate time. This review looks at what aiR actually does today, where client data flows, how it lines up against recent ethics guidance, and who should (and should not) be buying it.
Basil, which publishes this review library, is an on-device meeting notetaker for lawyers and does not compete in eDiscovery. If you use aiR for review, Basil for Law pairs naturally alongside it for confidential meeting capture during the same matter.
What Relativity aiR Actually Does
aiR is not one product; it is a family of GenAI capabilities layered on top of RelativityOne. The three pieces most lawyers will encounter are:
- aiR for Review — reviewers write a case-specific prompt (issue criteria, exclusions, rationale requirements), and the model classifies documents against those criteria and produces natural-language rationales. Relativity positions it as an alternative or supplement to traditional TAR/CAL.
- aiR for Privilege — targets the historically expensive privilege pass: identifying likely privileged material, suggesting privilege categories, and drafting log entries.
- aiR for Case Strategy — a case-assessment layer that lets litigators query a matter's corpus in natural language and surface themes, actors, and timelines.
Underneath, aiR uses large language models hosted through Microsoft Azure OpenAI Service inside Relativity's own Azure tenancy. See Relativity's product overview at relativity.com/ediscovery-software/air-for-review and the broader aiR portfolio page.
The workflow will feel familiar to anyone who has run TAR 2.0: you scope, you calibrate against a validation set, you iterate on the prompt, and you measure. What is new is that reviewers get free-text rationales they can sample and challenge, which changes how QC and defensibility documentation look in practice.
The Confidentiality Lens
This is the section that matters most. Nearly everything a litigator loads into aiR is client data — often subject to protective orders, cross-border restrictions, or third-party confidentiality obligations. Four questions decide whether a GenAI tool is usable in that setting: where does the data go, is it retained, is it used to train models, and what could a subpoena served on the vendor reach.
Where the data goes. aiR runs inside RelativityOne, which is hosted on Microsoft Azure. The inference calls go to Azure OpenAI Service instances that Relativity operates within its own tenancy, not to the public OpenAI API. Relativity describes this architecture and its regional hosting options on its trust and security page. That matters because Azure OpenAI is contractually distinct from consumer ChatGPT: prompts and outputs are not shared with OpenAI and are not used to train foundation models. Microsoft documents this at learn.microsoft.com — Azure OpenAI data, privacy, and security.
Training use. Relativity states that customer data processed through aiR is not used to train the underlying foundation models. This is a meaningful architectural fact and one of the main reasons aiR is defensible in a way that pasting documents into a public chatbot is not — a distinction the Southern District of New York underscored in US v. Heppner (S.D.N.Y. Feb 2026), where Judge Rakoff held that a litigant's chats with a public AI platform were not privileged.
Retention. Prompts and outputs live inside the workspace like any other work product and are governed by the matter's retention configuration. Azure OpenAI's abuse-monitoring pipeline, which by default stores prompts for up to 30 days for content-filter review, can be turned off for eligible enterprise customers — Relativity has confirmed publicly that it operates aiR without that human review pipeline. Confirm the current posture in your order form.
What a subpoena on the vendor could reach. Realistically: the workspace contents, audit logs, prompt/response pairs stored in workspace history, and administrative metadata. This is the same exposure profile as any hosted review platform — it is not unique to aiR. What aiR does not do, based on the documented architecture, is create a second copy of your documents at OpenAI. That reduces, but does not eliminate, third-party risk. Standard mitigations still apply: matter-specific workspaces, aggressive retention settings, and DPAs that pass through your protective-order obligations. Relativity offers a customer compliance and DPA package including SOC 2 Type II, ISO 27001, ISO 27701, and HIPAA support.
None of this protects privilege as a legal matter — privilege is a doctrine, not a product feature. What it does is meaningfully reduce the architectural risk that using GenAI creates a new, discoverable copy of your client's information outside your control.
Ethics-Opinion Fit (ABA 512)
ABA Formal Opinion 512 (July 2024) frames a lawyer's duties around GenAI use through Model Rules 1.1 (competence), 1.6 (confidentiality), 1.4 (communication), 5.1/5.3 (supervision), and 1.5 (fees). aiR maps to those duties reasonably well, provided the lawyer does the work Opinion 512 expects:
- 1.1 competence. aiR requires prompt design and validation set construction. Neither is trivial. Opinion 512's competence requirement means someone on the team needs to understand precision, recall, and how sampling drives defensibility — not just the vendor's marketing.
- 1.6 confidentiality. The Azure-tenant architecture, no-training commitment, and enterprise DPA give you the material terms Opinion 512 tells lawyers to look for before feeding client data to a GenAI tool.
- 1.4 communication. Opinion 512 signals that, in many matters, lawyers should tell clients GenAI is being used on their documents. That is a client-letter update, not a product feature.
- 5.1/5.3 supervision. aiR produces rationales, which help supervising lawyers spot-check reviewer prompts. The reasoning trail is a real advantage over black-box TAR when a court asks how a call was made.
- 1.5 fees. If aiR compresses a 2,000-hour privilege review, you need to have thought about how that affects billing arrangements before the invoice goes out.
New York practitioners should also read NYC Bar Formal Opinion 2025-6 (December 2025), which layers additional supervision and disclosure expectations on top of Opinion 512.
Defensibility and Workflow
The practical case for aiR is that GenAI can review documents with reasoning that a human can inspect. That is different from TAR, where a model spits out a rank and the defensibility story is statistical. With aiR, defensibility rests on: (1) the prompt as a written protocol, (2) a validation sample scored against reviewer ground truth, and (3) targeted QC of the model's rationales.
Two things to watch. First, prompts are protocols — treat them like a review manual, version them, and produce them if challenged. Second, rationales are persuasive precisely because they read like a junior associate's memo, which is exactly why they need sampling. Confident-sounding wrong answers are the failure mode of every LLM, and eDiscovery is no exception. Federal courts have been receptive to well-documented AI-assisted review since Da Silva Moore; the standard is reasonableness and proportionality, and that standard travels with aiR.
On the privilege side, aiR for Privilege is genuinely interesting because privilege review is the last redoubt of manual-only work at most firms. Even a modest lift in first-pass accuracy is economically significant. Just remember that a log entry drafted by a model is still your log entry once you serve it.
Pricing and Who It's For
Relativity does not publish list pricing for aiR. In practice, aiR is sold as an add-on to RelativityOne subscriptions and is typically priced per-document or per-matter with volume-based tiers, negotiated through Relativity or a certified partner. If you want current numbers, request a quote via the aiR product page or ask your existing Relativity partner.
Who it fits:
- AmLaw and litigation boutiques already on RelativityOne running matters where review costs exceed roughly six figures. The math works because aiR displaces reviewer hours, not because it is cheap on its own.
- Corporate legal departments with recurring investigations or second-request workflows and an existing Relativity footprint.
- Service providers offering managed review — aiR is increasingly table stakes in RFPs.
Who it does not fit: small firms without an existing RelativityOne relationship, or matters where the whole corpus is a few thousand documents. Those cases are better served by lighter-weight review tools or a well-designed manual pass.
Verdict
aiR is one of the more credible GenAI-for-review offerings on the market because Relativity did the boring architecture work — dedicated Azure tenancy, no training on customer data, enterprise DPA, disable-able abuse monitoring, workspace-scoped retention. It does not eliminate privilege risk, it does not turn a bad prompt into a good review, and it does not replace the lawyer's supervision duty under Model Rule 1.6 and Opinion 512. It does compress the economics of large review meaningfully, and it produces a reasoning trail that is easier to defend than a rank score.
| Pros | Cons |
|---|---|
| Runs inside Relativity's own Azure tenancy; no data sent to public OpenAI. | Requires existing RelativityOne subscription — significant floor cost. |
| No customer data used to train foundation models. | List pricing not published; procurement takes time. |
| Enterprise DPA, SOC 2 Type II, ISO 27001/27701, HIPAA support available. | Prompt design and validation are real skills; there is a learning curve. |
| Natural-language rationales help supervision, QC, and defensibility narratives. | Rationales can be confidently wrong — sampling is not optional. |
| Privilege module targets the most expensive part of most reviews. | Overkill for small matters or firms not already in the Relativity ecosystem. |
If your firm is already on RelativityOne and running large or recurring review, aiR is worth a serious pilot on a bounded matter with a defined validation protocol. If you are shopping platforms from scratch, evaluate aiR alongside comparable offerings from Reveal, DISCO, Everlaw, and Casepoint on both economics and data-handling terms.
For confidential meeting capture that stays off the cloud entirely while your team runs eDiscovery in aiR, see Basil for Law.
This review is for information only and is not legal advice.
Frequently asked questions
Does Relativity aiR send my client's documents to OpenAI?
No. aiR uses large language models hosted through Microsoft Azure OpenAI Service inside Relativity's own Azure tenancy. Prompts and documents are not sent to the public OpenAI API, and Microsoft's Azure OpenAI terms state that customer data is not shared with OpenAI or used to train foundation models.
Is Relativity aiR compatible with ABA Formal Opinion 512?
aiR's architecture (dedicated tenancy, no training on customer data, enterprise DPA, workspace-scoped retention) gives lawyers the material terms Opinion 512 asks about. But compatibility is not automatic — the lawyer still has to run competence, supervision, communication, and fee analyses on the specific matter.
How is Relativity aiR priced?
Relativity does not publish list pricing. aiR is sold as an add-on to RelativityOne, typically on a per-document or per-matter basis with volume tiers, and is quoted through Relativity or a certified partner.
What could a subpoena served on Relativity reach?
Realistically, workspace contents, audit logs, prompt and response history stored in the workspace, and administrative metadata — the same exposure profile as any hosted review platform. The Azure-tenancy design means there is not a separate copy of your documents sitting at OpenAI.
How does aiR differ from traditional TAR?
Traditional TAR produces a rank or classification with a statistical defensibility story. aiR uses generative models to classify documents against a written prompt and produces natural-language rationales, which change how QC and defensibility documentation look in practice.
Is aiR appropriate for small firms?
Usually not. aiR requires a RelativityOne subscription and its economics assume large review volumes. Small firms or small matters are typically better served by lighter-weight review tools or a well-designed manual pass.
Meeting notes with no server to subpoena
Basil transcribes and summarizes entirely on-device — privilege-safe by architecture. See Basil for Law →
This review is for information only and is not legal advice.