Generative AI is creating new possibilities in eDiscovery, but every matter brings its own questions, data, and objectives. Sometimes, the analysis a legal team needs does not fit neatly into a predefined workflow.
That is where Custom Analyses in Relativity aiR for Review comes in.
During a recent Page One Quick Bytes session presented in collaboration with Relativity, Page One’s Chief Operating Officer Andrew Milauskas and Relativity Senior User Experience Researcher Melissa Maloney explored how Custom Analyses gives teams more flexibility to classify information, extract specific data, and analyze both text and images based on the needs of a particular matter.
The goal is not simply to add another AI capability. It is to give teams greater control over what they ask AI to analyze, how they ask it, and what they want returned.
Why Custom Analyses?
aiR for Review already provides purpose-built analysis types for common eDiscovery needs, including relevance and issues. These workflows provide structured outputs designed specifically for those use cases.
But as teams began working with aiR for Review, Relativity saw users pushing beyond those predefined analyses.
Teams were incorporating additional instructions into prompts to extract information or answer questions outside of traditional relevance and issue analysis. At the same time, feedback around capabilities such as document summaries and topics pointed toward the same need: users wanted the ability to customize AI analysis for their own use cases.
Custom Analyses was developed to provide that flexibility.
Rather than beginning with a predetermined analysis type, teams can define the concept, criteria, and logic they want aiR for Review to apply to a document set.
That creates opportunities to use AI for questions that are highly specific to the matter at hand.
What Can Custom Analyses Do?
Custom Analyses currently supports two primary types of analysis: text analysis and vision analysis.
Text Analysis
For text-based documents, teams can configure an analysis to evaluate extracted text on a document-by-document basis.
Depending on the use case, that could mean asking aiR for Review to:
- Summarize a document
- Provide a yes/no answer based on defined criteria
- Categorize or label information
- Extract names, dates, amounts, or other details
- Identify information matching a particular condition
This can be applied across communications, contracts, policies, reports, transcripts, forms, and other text-based material.
For example, Andrew demonstrated an analysis of a fictional baseball agreement. Instead of simply summarizing the agreement, the analysis extracted specific financial information, including base salary, performance bonuses, and other compensation details.
The same concept can be applied to many different types of matters: identify the information you need, define how it should be returned, and run that analysis across the appropriate document population.
Vision Analysis
Custom Analyses also extends beyond extracted text.
Vision analysis can evaluate supported image files such as JPEGs, PNGs, and GIFs (and more coming soon!). This opens additional possibilities for evidence that may otherwise require significant manual review.
A team could analyze:
- Photographs
- Screenshots
- Handwritten notes
- Image-only exhibits
- Construction site images
- Forms
- Other visual evidence
During the demonstration, one example included a handwritten note that traditional optical character recognition did not successfully capture. Custom Analyses was able to interpret the handwritten content, provide a summary, and categorize the document as a handwritten note.
Another example used a medical form, where the analysis could extract information such as pain levels and details about how an injury occurred.
For matters containing thousands of images or handwritten records, the ability to ask targeted questions of that content can create an entirely different starting point for review.
From Prompt to Insight
The flexibility of Custom Analyses also means that the quality of the analysis depends heavily on how the task is defined.
Within a Custom Analyses project, teams can create individual insights representing the information they want returned for each document.
For example, imagine a collection of expense receipts. Instead of simply asking AI to analyze each receipt, the team could create an insight asking:
Does this receipt show a total meal expense greater than $100?
The prompt can then establish boundaries around that question. It might specify that the analysis should use the final total rather than the subtotal, should not use the gratuity amount independently, and should not calculate a new total when one is not provided.
That level of specificity matters.
A useful Custom Analysis starts with a clearly defined question and a clearly defined expected output.
Building a Better Custom Analysis
During the session, Andrew compared building a Custom Analysis to preparing for game day: you need the right game plan before stepping onto the field.
There are several elements that can help create a stronger analysis.
Start with one clear topic.
Determine exactly what you are trying to learn from the document set.
Give the analysis a single task.
Rather than asking one insight to accomplish several different objectives, define one document-level question at a time.
Establish boundaries.
Tell the system what should and should not qualify. Positive and negative examples can help clarify the distinction.
Plan for edge cases.
Real-world data is rarely perfectly consistent. Consider what should happen when information is blank, unclear, illegible, formatted differently, or falls directly on a defined threshold.
Define the output.
Determine whether you need a yes/no answer, a category, extracted text, an amount, or another structured result.
These decisions can make the output easier to evaluate and more useful downstream.
Test Before You Scale
One of the most important steps in building a Custom Analysis happens before running it across the entire document population: test it.
A prompt that appears straightforward may encounter unexpected variations once it reaches real-world data.
Andrew shared an example involving handwritten employee timesheets. The objective was to identify records where an employee worked a certain set of hours but did not record a required lunch break.
On paper, the question sounds relatively simple.
The actual documents introduced significantly more complexity. Some employees might leave the lunch field blank. Others might enter a zero, place an X in the field, record a decimal, or complete the form in another unexpected way. Some handwriting may be difficult to interpret altogether.
Those variations need to be considered when designing and validating the analysis.
Instead of immediately running an analysis across thousands of documents, teams should isolate a representative sample, evaluate the results, refine the prompt, and confirm that the output behaves as expected.
AI can accelerate the analysis, but the legal team still determines whether those results are appropriate for the matter.
Where Can Custom Analyses Be Used?
The flexibility of Custom Analyses means the potential applications extend across many different types of matters and investigations.
During the Quick Bytes session, Andrew highlighted several examples:
- Construction matters: analyzing photographs for potential hazards, conditions, or other visual evidence
- Handwritten records: extracting and evaluating information that may not be accessible through traditional text-based workflows
- Corporate investigations: identifying information related to potential bid rigging or other defined conduct
- Employment matters: detecting hostile language, policy concerns, or specific incident details
- Product liability: locating and summarizing information related to alleged defects
- Financial investigations and fraud: extracting and normalizing monetary information or identifying unusual payment structures
These are only starting points.
The broader opportunity is the ability to begin with the question that matters to the case rather than trying to fit every question into the same predefined analysis.
A Simple Framework: Define, Refine, Validate
For teams beginning to experiment with Custom Analyses, the process can be distilled into three steps:
1. Define the question.
Identify the specific document-level question or output you need.
2. Refine the instructions.
Add the appropriate context, boundaries, edge cases, examples, and formatting requirements.
3. Validate the results.
Run the analysis against a representative sample and evaluate the output before expanding it across the broader document population.
The technology may be sophisticated, but successful implementation still depends on a disciplined workflow.
What’s Next for Custom Analyses?
Relativity is also continuing to develop the capability.
During the session, Melissa shared several areas currently being explored, including expanded image support, more structured output options, rationale and consideration citations, additional ways to provide context, and greater reusability.
That reusability is particularly notable. As teams develop prompts and analyses that work well for recurring use cases, Relativity is exploring ways to save and reuse individual insights across workspaces, as well as templatize entire Custom Analyses projects.
Over time, that could allow teams to turn successful AI workflows into repeatable processes rather than rebuilding them matter by matter.
Making AI Fit the Matter
One of the most valuable aspects of Custom Analyses is right there in the name: custom.
Legal data does not arrive in one standard format, and the questions teams need to answer vary significantly from one matter to another. A workflow that works perfectly for a financial investigation may have little relevance to an employment dispute, construction matter, or internal investigation.
Custom Analyses gives teams another way to adapt aiR for Review around those specific needs.
But flexibility does not eliminate the need for process. The strongest workflows begin with a clearly defined objective, account for the realities of the underlying data, and validate results before scaling.
The opportunity is not simply to ask AI more questions.
It is to ask better-defined questions of the right data—and build a repeatable process around the answers.
Want to see Custom Analyses in action?
Watch the full Page One Quick Bytes session, Custom Analyses in Relativity aiR for Review, featuring Andrew Milauskas, Chief Operating Officer at Page One, and Melissa Maloney, Senior User Experience Researcher at Relativity.
Watch the Quick Bytes: https://youtu.be/cqlJl2DHbbo?si=hvkMauvO0idQhYMa