In litigation, automated document review means using machine learning and AI agents for the first pass of discovery review: ranking documents by how likely they are to be responsive (within the scope of the discovery requests), applying preliminary issue codes, clustering similar documents and email threads, and flagging potential privilege for attorney review. Automated review builds on technology-assisted review (TAR), also called predictive coding, which learns from attorneys' coding decisions to prioritize the rest of a collection.
Linear review by associates and contract attorneys remains a sound way to work through a collection, and many teams use it for smaller matters. Document review automation becomes worth the setup when the volume of electronically stored information (ESI) or the deadline outgrows what linear review can cover. Automation moves likely responsive records to the front of the queue and keeps attorney time for privilege calls and for case work that builds on the review, such as evidence cataloging and deposition summaries.
What follows explains how automated review relates to TAR, walks through the workflow from ingestion to production, and shows where attorneys stay in the loop and what keeps the process defensible.
How Automated Document Review Relates to TAR and Predictive Coding
TAR ranks documents based on attorney-coded examples, and automated review adds concept search, clustering, and agent-handled tasks around that ranking. Two TAR approaches are common, and the choice depends on how stable the coding criteria are and how much attorney feedback the team can sustain:
TAR 1.0 (simple passive or simple active learning): Attorneys code a training set until the model stabilizes, then classifies the remaining documents. Changes to the coding criteria mean revisiting the training set.
TAR 2.0 (continuous active learning): The model keeps learning as attorneys code, re-ranking the queue so the most likely responsive documents surface next. TAR 2.0 needs a steady flow of attorney decisions.
Natural language processing (NLP) adds concept search, which finds documents about the same event even when witnesses use different terms, and clustering, which groups near-duplicates and related documents.
Courts have accepted TAR for years. In Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125 (S.D.N.Y. 2015), Magistrate Judge Peck wrote that "it is now black letter law that where the producing party wants to utilize TAR for document review, courts will permit it." Defensibility still depends on the protocol and the validation record behind each use.
The Automated Document Review Workflow From Ingestion to Production
Each stage below has a defined output and a point where an attorney or litigation support specialist signs off before the next stage starts.
Ingest and Process ESI
Processing extracts text, preserves metadata, and keeps family relationships intact, so an attachment stays linked to its parent email. Collections can include email, chat exports, office files, and scanned paper records. Before releasing anything for review, litigation support confirms the collection scope and logs processing exceptions, such as encrypted or corrupt files.
Cluster and Thread Similar Documents
Near-duplicate detection, email threading, and concept clustering group related material, so one reviewer can code an entire thread or cluster consistently. Clusters also show the shape of a collection early, such as a large group of automated system notices that can be reviewed in bulk. The review manager decides how clusters are assigned and which ones need closer reading.
Rank Documents by Likely Responsiveness
The model ranks the collection by how likely each document is to be responsive, learning from a seed set (a starter group of attorney-coded documents) or from the queue as attorneys code it. Counsel first defines the issues and coding criteria, then validates a sample of the model's calls against them and adjusts the instructions where the model and the attorneys disagree.
Apply Preliminary Codes
The system applies preliminary responsiveness, issue, and confidentiality codes based on document content and attorney feedback. Preliminary coding works only when those definitions are stable. Ambiguous documents go to a senior reviewer, and when the claims or requests change, counsel revises the coding rules so the model stops enforcing outdated criteria.
Reviewers unsure how a coding definition applies can ask Datagrid's Mentor Agent, which answers from the connected review protocol and names the section it relied on.
Flag Potential Privilege
Privilege screening flags communications involving counsel, legal domains, or language that resembles legal advice, and sends those documents to attorneys first. Attorneys decide privilege based on participants, purpose, attachments, and possible waiver. Many teams also seek an order under Federal Rule of Evidence 502(d) so a disclosure connected with the case does not waive privilege there or in other federal or state proceedings.
Validate and Produce
Before production, the team samples coded documents for quality control and runs an elusion test, which samples documents predicted to be nonresponsive to estimate how much responsive material was missed. The validation record captures how each sample was drawn and why the team stopped reviewing, so counsel can explain the process if the opposing party or the court asks.
Agentic AI Document Review in Litigation
Agentic AI document review extends TAR from ranking into the repeatable work between attorney decisions. Teams automating document review for faster litigation typically hand AI agents tasks like these:
Route exceptions: Send ambiguous coding calls and potential privilege flags to the assigned senior reviewer.
Extract entities and dates: Pull names, organizations, and dates across custodians (the people whose files were collected) to build a working case timeline.
Compare coding patterns: Show where reviewers apply the same instruction differently, so the review manager can clarify the protocol.
Assemble validation records: Collect samples, overturned decisions, and coding changes as the review proceeds.
Track discovery obligations: Compare requests, review assignments, and production dates, and flag commitments that are close to due.
For quick factual questions about a collection, such as which custodians discussed a particular contract, attorneys can use Datagrid's Fast Search Agent, which answers from connected documents and lists its sources for follow-up.
Where Attorneys Stay in the Loop
Attorneys keep the coding criteria, every privilege determination along with the privilege log (the list of withheld documents and the privilege claimed for each), the decision to stop reviewing, and approval to produce.
Keeping those decisions with attorneys matches the duties lawyers already carry. The American Bar Association's ABA Formal Opinion 512 (July 2024), which addresses generative AI tools, says lawyers must reasonably understand the capabilities and limitations of the AI tools they use. The opinion warns that uncritical reliance on AI output can produce inaccurate advice, and says supervising lawyers must make reasonable efforts to ensure others in the firm comply with their professional obligations.
What Makes Automated Review Defensible
Supervision also means being able to defend the process later, and a defensible review is one counsel can explain and repeat. Before relying on automated output for production, confirm the review record includes:
A written protocol: Issues, coding definitions, escalation rules, and who has sign-off authority.
Validation by sampling: Documented samples of coding calls and the elusion test, with the method used to estimate recall (the share of responsive documents the review found).
Traceable coding decisions: For any coded document, reviewers can see the text, metadata, code history, and the reviewer who made each change.
Repeatable searches: Saved searches tied to the protocol that return the same results when rerun.
Access controls: Role-based access to privileged and confidential material, with an audit trail.
Automation adds work of its own. Validation and protocol updates take attorney time, and continuous active learning can drift when the case theory changes mid-review.
Handle the Work Around Discovery Review With Datagrid's AI Agent
Around the review itself, Datagrid's AI agents can take on the document-heavy work and leave coding judgment with the litigation team:
Document classification: Sort collections by document type, custodian, and date range before review begins.
Witness document sets: Gather the documents a witness sent, received, or appears in, with dates and participants, for deposition preparation.
Search across matter documents: Answer questions about the collection using semantic search, with links to the source documents.
Review summaries: Draft summaries of clusters or document sets for the review manager to check.
Counsel keeps every privilege call and the decision to produce.
Get started with Datagrid on documents from a closed matter, where the team already knows the right answers.
Frequently Asked Questions About Automated Document Review
Do Courts Accept Technology-Assisted Review?
Yes, US courts have accepted TAR since Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012), and the Rio Tinto decision in 2015 called it black letter law when the producing party chooses to use it. Courts still look at whether the process was reasonable and validated.
Can AI Agents Make Privilege Decisions?
No, AI agents can flag potential privilege and prioritize those documents, but attorneys decide privilege based on context, participants, purpose, and waiver risk.
How Should Documents Be Prepared for Automated Review?
Process the collection into searchable text, preserve metadata and family relationships, confirm collection scope, and log processing exceptions before release. Define the issues, coding criteria, escalation rules, and the attorney with sign-off authority before any model is trained.
