by: Rachel Fernandez, Director of Sales Engineering
CAL workflows have become the standard for technology assisted review, but the emergence of generative AI and large language model-based classification is changing what is possible in document review. The question for most practitioners is no longer whether AI has a place in eDiscovery, but when to use it, how to configure it, and how it fits alongside the TAR workflows already in place.
This article covers what to consider before selecting an AI classification workflow, and how those decisions play out in both Relativity and Everlaw.
AI Classification vs. TAR
Traditional TAR workflows leverage machine learning to predict document coding. The model learns from human coding decisions and applies those patterns to rank unreviewed documents. This methodology has been established in case law and has a clear validation methodology.
Generative Artificial Intelligence (GenAI) tools work differently. Instead of learning from coding decisions made by humans, the large language model (LLM) is given a prompt describing relevance and applies that instruction across the document population. The LLM can evaluate context, nuance, and meaning in ways that traditional machine learning cannot.
The two primary tradeoffs to plan around are cost and court precedent. Currently, AI classification tools carry a higher per-document cost in most configurations, and this technology does not yet have the same depth of court precedent behind it that TAR workflows have accumulated over the past decade. These are not reasons to avoid it, but factors to account for when planning your workflow and your defensibility strategy.
Understanding how these differences affect your matter starts with an honest evaluation of your data and your goals.
Evaluating Your Matter for AI Classification
Much like TAR, GenAI tools have several factors that contribute to the success and return on investment for your matter. Each should be considered carefully before committing to a workflow.
- Volume and Cost Structure. GenAI tools can be applied to matters of any size, but cost planning requires particular attention. Most AI classification tools carry a per-document cost for analysis, which compounds quickly on large data sets. Planning is also made more complex by the iterative nature of prompt development, which often requires multiple rounds of analysis before results are fully reliable. Building a buffer into your budget for this iteration is important and allows you to reach defensible results without being constrained mid-project by cost overruns.
- Matter Complexity and Issue Count. One of the strongest advantages of AI classification over traditional TAR is its ability to code for multiple issues simultaneously. Matters involving various legal issues, layers of nuance, or a need for rapid classification across a large document population are particularly well suited for this technology. Where TAR requires separate workflows for separate issues, AI classification can return layered results in a single pass, reducing both time and the manual effort required to manage parallel review tracks.
- Document Quality. Similar to CAL, most current AI classification tools perform best on good quality extracted text. Some tools are specifically designed to process a broader range of document types including audio files and images, which offers more flexibility than traditional TAR workflows. These expanded capabilities may carry additional cost, so identifying your document composition during scoping allows you to select the right tool and anticipate the full cost structure before the review begins.
- Prompt Design and Relevance Definition. Because AI classification is instruction-driven rather than training-driven, the quality of your prompt is directly proportional to the quality of your results. This places a premium on having a precise, well-documented relevance definition before the workflow begins. Vague or overly broad instructions produce inconsistent results. This is an area where working with an experienced consultant early makes a meaningful difference.
- Defensibility Planning. AI classification workflows are newer and carry less established court precedent than TAR. This does not make them indefensible, but it does require more deliberate documentation of your methodology. Be prepared to explain your prompt design, your validation approach, and your quality control process clearly. Courts and opposing counsel are increasingly familiar with AI in document review, and a well-documented AI workflow is becoming more accepted. An undocumented one is not.
AI Classification Alongside TAR
AI classification does not have to replace CAL altogether, or in all situations. In many matters, layering both tools can complement various workflows while driving efficiency and a stronger work product.
One common hybrid approach is to use AI classification for early case assessment and then leverage a full CAL workflow to further process large document sets. Another is to use CAL to surface likely responsive documents and run that subset through AI classification for issue tagging. A third approach is to use AI classification on document types that CAL cannot process effectively, such as images or audio transcripts, and feed those results back into the broader review workflow alongside your CAL output. The decision of which tool to lead with should follow from your matter evaluation rather than platform defaults or team familiarity.
An experienced partner can help consult on custom workflows that best suit the specific needs of your matter.
Platform Considerations
As with CAL, the platform you work in shapes how AI classification is configured, monitored, and documented. Understanding the differences between Relativity and Everlaw in this context helps ensure you are selecting the right environment for your matter.
- Relativity. Relativity’s AI classification tool is aiR for Review, which leverages GPT and is integrated directly within the platform. aiR classifies documents based on a drafted prompt describing relevance and scores them on a graduated numeric scale rather than a binary relevant or not relevant decision. This scaling makes it significantly easier for the tool to capture nuance across complex matters. Relativity’s existing review infrastructure integrates directly with aiR results, meaning you can build review queues, batching logic, and quality control workflows around aiR classifications the same way you would with Active Learning rankings. Prompt design is the most critical setup step. Work with your team to develop clear, specific instructions that reflect the matter’s relevance definition, and test your prompt on a sample population before applying it to the full data set. Consider Relativity when your matter requires deep configurability, multi-issue classification, and tight integration between AI classification and your existing review infrastructure.
- Everlaw. Everlaw’s AI tools include document summarization, predictive coding, and AI assisted analysis features that sit alongside its traditional predictive coding workflow. The platform’s approach to AI classification reflects its broader design philosophy of making powerful tools accessible to teams without deep technical backgrounds while still providing the metrics needed for a defensible workflow. Teams already working in Everlaw can layer AI assisted analysis into their existing environment without significant additional configuration, and the platform’s visual reporting makes it easier to communicate AI classification results to client teams or supervising counsel who may be less familiar with the technology. Consider Everlaw when accessibility and collaboration between review teams are priorities, or when your client’s in-house team is participating in the review process and values transparency in how results are communicated.
Summary
The integration of AI classification into eDiscovery workflows is not a distant development. It is happening now, across matters of all sizes and types. The practitioners who are getting the most out of these tools are not necessarily the ones with the most technical expertise. They are the ones who are asking the right questions at the scoping stage, selecting workflows deliberately, and documenting their process carefully throughout.
Across all three articles in this series, the throughline is the same: the decisions made at the beginning of a matter determine the quality of everything that follows. Whether you are scoping custodians, evaluating your data for CAL, or selecting an AI classification workflow, a methodical and well-documented approach is what separates a defensible process from an expensive one.
The Sandline team works across both platforms and across the full range of TAR and AI classification workflows. We are here to help you navigate these decisions and build a process that works for your matter.