Relativity connects AI review to the legal evidence workflow

Relativity applies AI to the evidence-heavy work of litigation and investigations: organising legal data, reviewing documents and building a matter’s factual record. The platform now presents RelativityOne as Relativity aiR. Its appeal is the connection between analysis and the established review process, but buyers need to distinguish included capabilities, metered use and newer products that remain in limited access.

01 / ProductAI belongs inside the evidence process

The platform overview positions Relativity aiR across collection, processing, document review, production and case strategy. It also describes Assist, integrations and APIs. These are parts of a larger evidence workflow, rather than a stand-alone question box. A useful result must remain connected to the documents, reviewer decisions and matter context that justify it.

The aiR Review documentation describes using a subject-matter expert’s criteria and extracted document text to return recommendations, reasons and citations. The documented approach moves from developing a prompt on a small sample to validation on a separate sample and then application to a larger set. That sequence gives reviewers a way to assess the method before expanding it.

claiR is a newer assistant for broader matter work, with skills and source-linked outputs. Its product page labels access as Advanced Access, while the pricing page calls it Coming Soon. Treat it as a distinct availability question. This article’s proposed exercise uses the documented aiR Review process and does not depend on receiving claiR.

The company’s AI principles acknowledge limitations, including errors and bias, and emphasise control, clarity and validation. That is the right framing for procurement. A cited recommendation is useful evidence about how an analysis was reached; it is not an automatic determination that a document should be produced, withheld or treated as legally significant.

02 / AudienceA review protocol gives the AI something testable to do

The strongest starting point is a team with a defined issue, a representative document collection and reviewers who can explain the inclusion rules. An internal investigation might need to identify communications about a specific decision. A litigation team might need to apply a negotiated responsiveness definition. In both cases, examples and counterexamples are essential to assessing the output.

The fit is weaker when the main need is ordinary drafting or when no one can agree what counts as relevant. Before software can help consistently, the team must resolve the question it is asking. A large review platform can organise complex work, but it cannot supply the missing agreement between matter leads about scope and treatment of ambiguous evidence.

OpenText offers a comparison for broader enterprise information and discovery requirements. Harvey is relevant where the centre of work is a legal AI workspace rather than a production-oriented evidence review. Compare the complete matter process, including data preparation and review responsibility, so that an attractive interface does not obscure a missing operational step.

03 / WorkflowA proposed issue review separates development from validation

Consider an authorised historical investigation with a known set of documents about a disputed commercial decision. Select examples that support the issue, superficially similar material that does not, and difficult borderline cases. This is a proposed pilot, not a report of testing. The expected decisions should be recorded independently before the team begins adjusting instructions.

Have the subject-matter expert write the issue definition in ordinary language. State what belongs, what does not and which facts change the decision. Avoid relying on a private shorthand that only the original reviewer understands. A protocol that needs constant verbal explanation is not yet a stable basis for comparing human and AI recommendations.

Use a development sample to identify ambiguous wording and improve the criteria. Keep a record of the prompt versions and the reason for each change. If an example is used to teach the distinction, do not then count success on that same example as independent evidence of generalisation. Separate learning from evaluation in the team’s own records.

Apply the settled criteria to a different validation sample with a known outcome. Review false positives and missed relevant documents separately. An irrelevant document that consumes a reviewer’s time has a different consequence from important evidence that disappears from the prioritised set. Examine the source passages behind consequential recommendations rather than accepting an apparently plausible explanation.

Only then expand to a larger collection, with sampling and escalation responsibilities agreed in advance. Preserve disagreements for the matter lead to resolve. Where a document’s significance depends on its surrounding conversation or an attachment, ensure that reviewers inspect that context. A per-document recommendation should not be allowed to stand in for a broader factual conclusion.

Finish with a short evaluation record showing the criteria, sample design, error patterns, corrections and usage consumed. Ask someone outside the prompt-development work to reproduce the main conclusions. The practical outcome is a review process the team can explain and maintain, not simply a large number of classified documents or an impressive first demonstration.

04 / PricingIncluded AI capabilities do not establish unlimited analysis

The distinction between a licence and consumption is central to the business case. Ask the vendor to work through the proposed sample, expected production volume and likely reruns using the actual agreement. A team that repeatedly refines criteria may consume resources differently from one applying a settled protocol once. The public material does not settle a particular customer’s allowance.

Also include data preparation, hosting duration, specialist support and reviewer effort in the comparison. A cheaper analysis pass may not reduce the total matter cost if the team must repeat ingestion or spend substantial time resolving poor extracted text. The useful economic unit is a review stage that meets the agreed quality requirements, with its associated data and staffing costs visible.

05 / DistinctionThe validation path is as important as the model

The most valuable design idea is the explicit move from criteria development through a separate validation exercise to broader use. That gives a legal team an operational structure for learning where a method works. It also makes changes reviewable: if the issue definition moves, the team can decide which prior decisions need to be revisited instead of silently treating them as current.

Calder7 describes Relativity’s security organisation and customer-facing security capabilities. For an evidence platform, this is a distinct part of the evaluation from answer quality. Administrators need to understand permissions, monitoring and incident responsibilities for the actual matter environment, while reviewers need a clear route from a recommendation back to the source.

The wider platform can also keep review work connected to later case preparation and production. That continuity is useful when a document changes status or a conclusion needs correction. The adoption question is whether the organisation can follow those changes without creating parallel spreadsheets whose decisions diverge from the authoritative review record.

06 / QuestionsAsk what the model sees and what the contract enables

The review documentation explains that basic document analysis treats documents independently and uses extracted text. That sets an important boundary for the pilot. Test a relevant message whose meaning depends on an attachment or adjacent communication, and decide how the review protocol handles that dependency. Do not assume the model automatically reconstructs every relationship in the matter.

Next, confirm language and data characteristics. A method that works on clean English correspondence may behave differently on scans, multilingual records or terse technical material. Include those conditions in a representative sample if they will appear in the actual collection. Describe the observed limitation precisely rather than translating a successful demonstration into a general accuracy claim.

Finally, obtain the account’s current feature status and commercial usage schedule. Some detailed consumption information leads into customer-community material, and no customer contract was inspected for this article. This Blueprint is based on public sources, without a live tenant test. Availability, processing configuration, review quality and matter-specific obligations remain questions for the evaluating team.

07 / DecisionChoose a defensible review method before expanding the collection

Relativity deserves evaluation when AI must sit inside a substantial legal evidence process with accountable review and downstream handling. Begin with a known issue and independent validation material. Scale when the team can explain the errors, usage and correction path, rather than when the first answer happens to read well.