aiR for Review is Relativity’s generative AI document review tool that classifies documents for relevance, flags key documents, and applies issue tags using natural-language prompts instead of a trained predictive coding model. Where traditional technology-assisted review (TAR) requires seed sets and iterative machine learning training rounds, aiR for Review works from written review criteria closer to instructing a reviewer than training a classifier. That distinction matters if you’re evaluating AI document review software or weighing a predictive coding alternative for an upcoming matter.
This guide walks through how to configure a project correctly the first time, and provides a recommended workflow for using aiR for review. It is based on our hands-on setup experience because to truly create an aiR for Review project that saves real document review hours often comes down to a handful of decisions made before you ever click “analyze.”
(Note: This assumes aiR is already installed within your Relativity; that setup is covered separately.)
TAR vs. Generative AI Review: Where aiR for Review Fits
When comparing TAR vs. generative AI review approaches, the practical difference is workflow, not just technology. Traditional TAR (predictive coding) needs a human reviewer to code a seed set, train a statistical model, and validate it through recall/precision testing before it’s defensible at scale. aiR for Review instead asks you to describe your relevance criteria in plain language. Use a case summary, examples of what’s relevant, examples of what isn’t, and the model reasons against that description document by document, providing a rationale and citation for every score.
That doesn’t mean aiR skips validation. You still need a representative, human-coded sample to test your prompts against before trusting results across a full population; more on that below. But the setup investment shifts from “train a model” to “write a precise prompt,” which is a meaningfully different skill set for review teams to build.
Not sure whether TAR or generative AI review fits your matter?
Start With Your Prompts, Not Your Documents
The single most important piece of advice is to develop your aiR for review prompts before you start analyzing documents. When you set up a new aiR for Review project, you’ll choose from five analysis types:
- Relevancy review: classifying documents as responsive or not
- Key document identification: flagging documents of heightened importance
- Issue tagging: categorizing already-identified relevant documents by issue
- Confidential Business Information (CBI): classifying documents containing sensitive material like trade secrets or non-public financial data.
- Custom analysis: prompt-driven insights for matter-specific questions, extractions, or classifications, including image review
You can combine relevancy and key document review in a single pass, then layer in issue tagging once you’ve established your responsive set.
Build a Smart Evaluation Set First
Don’t point aiR at your full document population. Instead, start with a small, already-coded sample document set. That sample should be balanced: include known key documents, known responsive documents, and known non-responsive documents. Mixing all three gives the model something real to be evaluated against.
Run your prompts against that small set first. If the results skew too broad or too narrow, you’ll be able to determine where the model is drifting and adjust your prompt language accordingly. Because the set is small, reruns are fast, usually a minute or two, so you can iterate on prompt wording quickly before committing to a full-population run.
Clean Your Data Source Before You Run It
aiR for Review needs text to work with. Documents need either extracted text or OCR text to be analyzable, but there are a few practical limits and cleanup steps worth doing up front:
- Text size cap: Documents with extracted text over 300KB will error out. Pull those from your data source before running.
- Consider scoping to emails only and excluding attachments if that fits your review protocol.
- Screen out file types the model can’t meaningfully analyze: audio, video, or proprietary formats that Relativity itself can’t process natively. These will need to be reviewed manually anyway, so don’t waste a run on errors you’ll just have to investigate afterward.
- Prioritize Custom Analysis parameters early: Designed to leverage aiR in ways that are unique to your matter versus the pre-built analyses for standard review workflows like responsiveness or issues. You define the criteria that is most important for your matter. Custom analysis helps to bridge the gap for expanded file types like images or scanned documents. For example, use it to determine which images contain people within a set of thousands.
A little cleanup on the front end saves time figuring out errors on the backend.
If your team doesn’t have bandwidth for setup and validation, Percipient runs aiR for Review projects end to end.
Setting Up the Project
Once your sample set and prompts are ready:
- Navigate to the aiR for Review tab and select aiR for Review Projects
- Click New aiR for Review Project to launch the setup wizard
- Give the project a clear label; name it after the matter and the type of analysis you’re running (for example, a case name plus “Relevancy + Key Review”)
- Choose your analysis type (relevancy, key documents, confidential information, custom analysis, or issues)
- Set a project use case this is a required field but mainly useful for your own tracking and organization
- Select your data source starting with your smaller, pre-coded validation sample
- Click Create Project
This lands you on the main aiR for Review project page, which has three core areas: your prompt criteria on the left, your data source and versioning controls, and your results.
Iterating on Prompts and Data Sources
You can edit and save prompts at any time except while a document set is actively being analyzed. Every time you save and analyze a new version, aiR tracks a version number, and you can click back into prior versions to see exactly what changed. If an earlier version performed better, you can copy it back in and modify from there.
When you’re ready to expand beyond your initial sample, there are three ways to adjust your data source:
- Run against a new sample set of similar size to check consistency
- Run a validation step, sampling from your full targeted review population and routing results into a review queue useful for pairing aiR output with human quality control and evaluation metrics in Review Center
- Apply the prompt set to your entire document population once you’re confident in your prompts and ready for full-scale analysis
Reading the Results
After a run, you’ll see summary metrics at the top of the page: counts of borderline, relevant, key, and non-relevant documents, along with a version history showing what changed between runs. It is important to note that aiR sends one document at a time instead of all at once or in batches so that documents are not disregarded by the model.
Each document gets a score from -1 to 4:
| Score | Description |
|---|---|
| -1 | Error: The document either encountered an error or could not be analyzed. For more information, see How document errors are handled. |
| 0 |
Junk: (For aiR versions below 2026-04-14) The document contains no useful information or is considered “junk” data, such as system files, an empty document, or sets of random characters. Refer to the aiR for Review Jobs tab for the aiR version used for the job. Highly Not Relevant: (For aiR versions 2026-04-14 and above) The document either contains no useful information or is predicted highly not relevant to the case or issue. Refer to the aiR for Review Jobs tab for the aiR version used for the job. |
| 1 | Not Relevant: The document is predicted not relevant. aiR did not find any evidence that it relates to the case or issue. |
| 2 | Borderline Relevant: The document is predicted to be borderline relevant. aiR found some content that might relate to the case or issue. It usually has citations. |
| 3 | Relevant: The document is predicted to be relevant to the issue. Citations show the relevant text. |
| 4 | Highly Relevant: The document is predicted to be very relevant to the issue. aiR found direct, strong evidence that the content relates to the case or issue. Citations show the relevant text. |
Alongside each score, aiR provides the following:
- Rationale: the reasoning behind the classification, including relevant considerations (for example, noting a document relates to one employee’s termination but not another’s)
- Citations: specific language pulled from the document that supports the score, up to five citations per document for Relevance and Key Documents analyses, and one citation per issue for Issues analysis
- Summaries: a newer feature that lets you filter on topics independent of your prompt criteria, giving you a broader content overview of what’s actually in your document set
Writing Prompts That Actually Work
aiR for Review prompt criteria section is where most of your effort should go. A few helpful specifics:
Case summary: This should read like your factual background pulled from the complaint and your review protocol. Relativity’s own documentation recommends capping the matter overview at around 20 sentences. Avoid legal jargon, stay concise, and write in an active voice.
Relevance criteria: Be concrete rather than procedural. Instead of describing what you want the model to look for in the abstract, give specific examples: “a document is relevant if it involves Jane Smith’s termination,” not “review documents for termination issues.” Concrete positive and negative examples instruct the model far more effectively than general instructions.
Field selection: For relevancy review, choose a field with a single responsive/non-responsive answer rather than one allowing multiple outcomes; this keeps the model’s task well-defined instead of open-ended.
Issue tags (optional): Adding issue tags gives the model more concrete texture to work with. For example, a “delays” issue tag might include a description like evidence of scheduling setbacks or shipment delays. Note that issue tags used this way inform relevancy scoring; they don’t get scored themselves unless you run a separate issue analysis.
Key documents: Select the field tied to your key document coding, and provide clear examples of what meets that bar the same way you’d define it in a review protocol.
aiR for Review works best as an iterative tool, not a one-shot black box. Build a small, balanced training set. Write concrete, example-driven prompts. Test on a sample before scaling to your full population. And use the rationale, citations, and summary features to actually understand why the model is scoring documents the way it is — not just what it decided.
Get the setup right, and aiR for Review becomes a genuine force multiplier for your review team rather than another tool generating results you don’t fully trust.
Frequently Asked Questions
What is Relativity aiR for Review? aiR for Review is a generative AI feature within Relativity that classifies documents for relevance, key document status, and issue tags based on natural-language prompts, rather than a trained statistical model.
How is aiR for Review different from traditional TAR or predictive coding? Traditional TAR trains a model on a human-coded seed set through iterative statistical rounds. aiR for Review instead evaluates documents against written relevance criteria and returns a score, rationale, and citation for each one, making it faster to configure but still reliant on a validated sample before full-scale use.
What does the aiR for Review scoring scale mean? Scores range from -1 to 4: -1 flags an error document; 0 is a junk or highly not relevant response; 1 suggests likely non-responsive; 2 indicates borderline or needs-further-review; 3 indicates likely responsive; and 4 indicates highly relevant.
What file types or document limitations does aiR for Review have? aiR for Review requires extracted or OCR text to analyze a document, and documents with extracted text over 300KB will error out. Audio, video, and proprietary file types that Relativity can’t natively process should be excluded before running a project.
Can aiR for Review replace human document reviewers? No aiR for Review is designed to accelerate and prioritize review, not eliminate human quality control. Best practice pairs aiR output with a human validation sample and review queues for borderline or high-stakes documents.
Percipient pairs aiR and other technology with expert human review on every matter.
