Technology-Assisted Review (TAR) and Predictive Coding

When an organization must review thousands or millions of documents — for litigation, a government investigation, a large public-records request, or an internal compliance matter — reading every item by hand is slow, expensive, and surprisingly error-prone. Technology-assisted review (TAR) is the family of techniques that uses software, and increasingly machine learning, to help human reviewers find the relevant material in a large collection faster and more consistently than manual review alone. Predictive coding is the best-known form of TAR: a workflow in which a machine-learning model learns from human coding decisions and then ranks or classifies the remaining documents by their likely relevance.

TAR sits at the intersection of records management, e-discovery, and broader information governance. The same machine-assisted methods that surface responsive documents in a lawsuit can help an organization identify records that are redundant, obsolete, or trivial; locate material subject to a legal hold; or triage a backlog of unstructured content. Understanding how TAR works — and where its limits lie — has become a core competency for records and information professionals, not just litigators.

How predictive coding works

At its heart, predictive coding is supervised machine learning applied to documents. A reviewer (typically a subject-matter expert who knows the matter) codes a set of documents as relevant or not relevant. The system uses those decisions as training examples, building a model of the linguistic and metadata features that distinguish responsive documents from non-responsive ones. The model then scores the rest of the collection, and the team reviews the highest-ranked documents first.

Practitioners generally distinguish two workflow generations:

• TAR 1.0 uses a one-time training phase. An expert codes a fixed seed set or several rounds of training documents, the model is finalized, and it then classifies the remaining population. Training and review are separate steps.
• TAR 2.0 (continuous active learning, or CAL) blends training and review into a single loop. The model continuously re-ranks the collection as reviewers code documents, repeatedly serving up the items it now believes are most likely relevant. Because learning never stops, CAL adapts to new concepts that emerge mid-review and usually requires no carefully constructed seed set.

Most modern deployments favor continuous active learning because it is simpler to operate, tends to reach high recall efficiently, and is more forgiving of imperfect early decisions.

Measuring effectiveness: recall, precision, and proportionality

TAR is evaluated with the same metrics used across information retrieval. Recall measures the share of truly relevant documents the process actually found; precision measures the share of retrieved documents that are genuinely relevant. There is an inherent tension between the two, and the goal is rarely perfection — it is a reasonable and proportional result.

Teams typically validate a TAR effort using statistical sampling. A random sample of the documents the model set aside as non-relevant is reviewed by humans to estimate how many relevant items were missed (the elusion or null-set rate). If that estimate is acceptably low, the review can defensibly stop. This emphasis on reasonableness rather than absolute completeness mirrors the proportionality principle embedded in the Federal Rules of Civil Procedure, which ask parties to weigh the burden of discovery against the needs of the case rather than demanding exhaustive manual review.

Defensibility and the role of transparency

Courts in the United States have broadly accepted TAR as a legitimate, and sometimes superior, alternative to manual or keyword-only review — but acceptance hinges on a defensible process. The Sedona Conference and similar practitioner bodies have published influential guidance describing what a defensible TAR workflow looks like: clear protocols, qualified reviewers training the model, documented decisions, and validation through sampling.

Several principles recur:

• Document the methodology. Record how the model was trained, what was reviewed, and how the stopping point was justified. This audit trail is what makes the result reproducible and defensible.
• Cooperate and disclose where appropriate. In litigation, parties increasingly negotiate TAR protocols, and good-faith transparency about the approach reduces disputes.
• Keep humans in the loop. TAR prioritizes and accelerates human judgment; it does not replace it. Privilege calls, confidentiality, and edge cases still demand expert review.

TAR within records management and information governance

Although TAR matured in the courtroom, its techniques apply directly to everyday records work. The same classifiers that find responsive documents can support information governance by identifying ROT (redundant, obsolete, and trivial) content for defensible disposition, flagging potential records buried in shared drives and mailboxes, or surfacing sensitive material such as personally identifiable information. These uses connect TAR to records-management standards like ISO 15489, which frame records processes — capture, classification, appraisal, and disposition — as deliberate, documented activities that machine assistance can make more consistent at scale. Explore related material on the information governance hub.

A note on standards is worth making here. Records professionals once looked heavily to functional specifications such as the DoD 5015.2 standard to define system capabilities. NARA, however, retired its endorsement of that standard in 2022 in favor of the Universal Electronic Records Management Requirements developed through the Federal Electronic Records Modernization Initiative (FERMI). The shift reflects a broader move toward outcome-based, technology-neutral requirements — an environment in which adaptive tools like TAR fit naturally, because they are judged by results and defensibility rather than by conformance to a rigid feature checklist.

Limits and good practice

TAR is powerful but not magic. Model quality depends on the consistency of human coding, the clarity of the relevance definition, and the nature of the collection — heavily numeric spreadsheets, images, or audio may need specialized handling. Concept drift, ambiguous criteria, and poorly scoped collections all degrade results. The remedy is disciplined process: qualified reviewers, stable definitions, statistical validation, and honest documentation. Used this way, technology-assisted review turns an impossible volume of unstructured information into something an organization can review, govern, and defend.