Legal document review is the stage of discovery where attorneys and support staff work through electronic files, sometimes millions of them, to identify what’s responsive to the case, pull out what’s privileged, and prepare the rest for production to the other side. The work sits inside the framework of the Federal Rules of Civil Procedure, and the consequences for getting it wrong run from paying the opposing party’s fees to losing the case outright. What follows walks through the full arc: what reviewers are looking for, how the process is set up, the technology that drives it, and the protections that exist when something slips through.
What Counts as Discoverable
Federal Rule of Civil Procedure 26(b)(1) defines discoverable information as anything relevant to a party’s claims or defenses that isn’t shielded by privilege. Relevance alone, though, doesn’t entitle a party to every file in the opponent’s possession. The rule also requires that discovery be proportional to the needs of the case, weighing factors like the amount in controversy, the parties’ relative access to information, their resources, the importance of the issues, and whether the burden of producing material outweighs its likely benefit.1Legal Information Institute. Federal Rules of Civil Procedure Rule 26 – Duty to Disclose; General Provisions Governing Discovery
These proportionality factors come up constantly during meet-and-confer negotiations, where opposing counsel work out the scope of what each side will search and produce. The outcome of those negotiations shapes everything downstream: which custodians get searched, which date ranges apply, which keywords run, and ultimately how many documents hit a reviewer’s queue.
Preserving Evidence Before Review Begins
Review doesn’t happen in a vacuum. Before anyone opens a review platform, the parties have a legal duty to preserve relevant evidence, and failing at this stage poisons everything that comes after.
The duty to preserve arises the moment litigation is reasonably anticipated. That trigger can be a demand letter, a government investigation notice, a terminated employee’s threat to sue, or any event that would put a reasonable organization on notice that a lawsuit is coming. Once triggered, the party must issue a litigation hold directing employees to stop deleting or modifying potentially relevant files, including halting any automatic deletion processes that would destroy data on a schedule.
A proper hold identifies the key custodians, the people most likely to have relevant information, and communicates clearly that they must preserve both physical and electronic records. Counsel should follow up to confirm compliance, because a hold letter that goes unread provides little protection when the other side moves for sanctions.
If electronically stored information that should have been preserved is lost because a party didn’t take reasonable steps to keep it, and the data can’t be recovered, Rule 37(e) gives courts two tiers of response. Where the loss merely prejudices the other side, the court can order measures to cure that prejudice but nothing more severe. Where the party intentionally destroyed evidence to deprive the other side of its use, the court can presume the lost information was unfavorable, instruct the jury to draw that inference, or dismiss the case entirely.2Legal Information Institute. Federal Rules of Civil Procedure Rule 37 – Failure to Make Disclosures or to Cooperate in Discovery
What Reviewers Pull Out of the Production
Not every relevant document ends up in the opponent’s hands. Several categories get flagged for withholding or redaction, and the calls are judgment-heavy.
Attorney-Client Privilege
Attorney-client privilege shields confidential communications between a lawyer and client when the purpose of the communication is to seek or provide legal advice. The protection only holds when the client intended the communication to stay confidential. An email from in-house counsel giving legal guidance to a company executive is typically privileged. That same advice forwarded to a third-party vendor who isn’t covered by a common-interest agreement probably isn’t.
Work Product
Rule 26(b)(3) protects documents and other materials prepared in anticipation of litigation by a party or their representative. This covers litigation strategy memos, interview notes, and legal research compiled by attorneys once a lawsuit is reasonably foreseeable. A court can order production of work product if the requesting party shows substantial need and an inability to get equivalent information any other way, but even then, the court must protect the attorney’s mental impressions, conclusions, and legal theories from disclosure.3Legal Information Institute. Federal Rules of Civil Procedure Rule 26 – Duty to Disclose; General Provisions Governing Discovery – Section: Trial Preparation Materials
Mixed-content documents are the hard part. A long email thread might start as a business discussion, shift into legal advice midway through, and return to operational matters. The protocol needs to tell reviewers whether to redact only the privileged portion or withhold the entire thread.
Personally Identifiable and Protected Health Information
Personally Identifiable Information includes data that can distinguish or trace a specific individual’s identity, either on its own or when combined with other linked information. Social Security numbers, financial account details, and government-issued identification numbers are the most common examples reviewers encounter.4U.S. Department of Labor. Guidance on the Protection of Personally Identifiable Information
Protected Health Information is individually identifiable health information that is transmitted or maintained in any form, whether electronic or paper. Federal regulations under HIPAA define PHI broadly and carve out limited exceptions for education records and certain employment records.5eCFR. 45 CFR 160.103 – Definitions
Reviewers tag these documents with specific codes so that the production team can apply redactions before turning files over. Court protective orders frequently dictate exactly how this information must be handled, and ignoring those orders creates liability for the producing party.
How the Review Itself Runs
The review protocol is the instruction manual that keeps hundreds of reviewers making consistent decisions across millions of documents. It originates from lead counsel and litigation support, and every reviewer should treat it as the single source of truth for the project.
A strong protocol starts with a case summary that gives reviewers enough factual background to understand what they’re looking at. It then defines the exact criteria for coding a document as responsive or non-responsive, spells out the privilege standards, and explains how to handle edge cases like partially relevant spreadsheets or email chains where only one message touches on a case issue. The protocol also specifies the tagging palette: the set of codes available in the platform for categorizing documents. Typical tags go beyond responsive/non-responsive to include categories like “hot document,” “confidential — attorneys’ eyes only,” or issue-specific codes tied to particular claims. Sample documents showing reviewers how to tag a tricky memo or multi-attachment email do more to prevent inconsistent coding than pages of abstract instructions.
The actual review happens inside specialized platforms such as Relativity, Everlaw, and DISCO. These tools display the document text alongside a coding panel where reviewers apply the tags defined in the protocol. Most reviewers work through batches, discrete sets of documents assigned to one person, typically grouped by custodian, date range, or keyword hit.
Each document gets a decision. The reviewer reads the content, examines the metadata (sender, recipients, date, file type), and applies the appropriate responsiveness and privilege codes. When an email and all its attachments share the same relevance and privilege status, most platforms allow bulk coding, applying one set of tags to the entire document family at once, which saves significant time on large populations. Once every document in a batch has been coded, the reviewer submits it as complete. The system logs every action on every document, creating an audit trail that matters both for quality control and for defending the review if it’s challenged later.
Catching Mistakes Before They Ship
Reviewers make mistakes. On a project with fifty contract attorneys coding documents for twelve hours a day, inconsistency is inevitable. Quality control protocols exist to catch those errors before they contaminate the production.
The most common approach is statistical sampling. A supervisor pulls a random sample from completed batches and re-reviews them against the protocol. Acceptance sampling takes this further by testing whether a reviewer’s error rate falls below a predetermined threshold, often 10%, and flagging batches that fail for full re-review.
Two metrics matter most. Recall measures the percentage of truly responsive documents that reviewers correctly identified. Elusion measures the percentage of documents coded as non-responsive that were actually responsive, essentially the rate at which relevant material slips through. A well-run review drives the elusion rate as close to zero as possible. In low-yield populations where responsive documents are rare, elusion testing is often more practical than recall calculations because it focuses the sampling effort where mistakes matter most.
Technology-Assisted Review and Generative AI
Manual review of every document in a large collection is slow, expensive, and produces inconsistent results. Technology-assisted review uses machine learning to prioritize the documents most likely to be relevant, cutting both cost and time.
In its most common form, continuous active learning, the system learns from every coding decision a reviewer makes. As reviewers tag documents, the algorithm continuously re-ranks the remaining population so that the most likely responsive documents surface next. Review continues until the algorithm has exhausted the high-probability documents and the elusion rate on what remains is acceptably low. This improves on earlier approaches that required a fixed training phase before the machine could rank anything.
Federal courts have approved technology-assisted review since 2012, when a judge in the Southern District of New York recognized computer-assisted review as an acceptable method for searching electronically stored information in appropriate cases.6Justia Law. Da Silva Moore v Publicis Groupe et al, No 1:2011cv01279 A later decision in the same court noted that courts generally don’t dictate how parties respond to discovery requests; the choice to use computer-assisted tools belongs to the producing party.7Justia Law. Rio Tinto PLC v Vale, SA et al, No 1:2014cv03042
Large language models are now entering the workflow, handling tasks from initial responsiveness coding to privilege identification. The legal standard hasn’t changed: the question remains whether the review process was reasonable, and the same statistical measures that validate traditional technology-assisted review (recall, precision, and elusion) apply to generative AI tools. Legal teams using these tools still carry the professional obligations of competence and diligence, which in practice means understanding how the model works, testing its outputs against human reviewers, and documenting the validation process for court if needed.
Producing Documents and Logging What’s Withheld
Once review is complete and quality control has signed off, the litigation support team assembles the production set: the collection of responsive, non-privileged documents that gets turned over.
Produced documents are typically converted to TIFF images or PDF files. TIFFs are static images that can’t be searched on their own, which is why productions include load files alongside them. Load files are structured data files, usually in .DAT or .OPT format, that contain the metadata and extracted text for each produced document. Standard metadata fields include the document’s author, date created, date sent for emails, recipients, custodian, file name, and a hash value used for authentication and deduplication. The receiving party loads these files into their own platform, which reassembles the documents with their metadata for searching and analysis. Format specifications are typically agreed during meet-and-confer or set by standing orders. Produce in a format the other side can’t use, and you’ll be doing it again on your own dime.
Every document withheld from production on privilege or work product grounds must be logged. Rule 26(b)(5) requires the withholding party to expressly state the privilege claim and describe the withheld document in enough detail that the other side can evaluate whether the claim holds up, without revealing the privileged content itself.8Legal Information Institute. Federal Rules of Civil Procedure Rule 26 – Duty to Disclose; General Provisions Governing Discovery – Section: Claiming Privilege or Protecting Trial-Preparation Materials
In practice, most courts require privilege logs to include the document’s date, author and position, recipients and positions, a general description of the document’s nature, and the specific legal basis for withholding it.9United States District Court for the District of Nebraska. The Dreaded Privilege Log – Rules and Practical Tips Where a document is partially privileged, teams often redact the protected portions and produce the rest rather than withholding the whole thing. That approach preserves parent-child relationships in email threads and reduces log volume. Redactions are tracked with notations explaining what was removed and why, which gives the receiving party enough transparency to challenge specific decisions if they believe the privilege claim is unfounded.
When Privileged Material Slips Through
Even with careful review, privileged documents sometimes end up in a production. When you’re pushing through hundreds of thousands of files on a deadline, inadvertent disclosures happen. Federal Rule of Evidence 502 provides two layers of protection.
Under Rule 502(b), an inadvertent disclosure in a federal proceeding doesn’t waive the privilege if the holder took reasonable steps to prevent the disclosure and acted promptly to fix the error once discovered.10Legal Information Institute. Federal Rules of Evidence Rule 502 – Attorney-Client Privilege and Work Product; Limitations on Waiver The “reasonable steps” analysis looks at the review process as a whole: the size of the production, the procedures used, and the resources available.
Rule 502(d) offers stronger protection. A federal court can enter an order declaring that any disclosure connected to the pending litigation doesn’t waive privilege, regardless of whether the disclosure was inadvertent or intentional, and that protection extends to other federal and state proceedings as well.10Legal Information Institute. Federal Rules of Evidence Rule 502 – Attorney-Client Privilege and Work Product; Limitations on Waiver In practice, these court-entered clawback orders require the receiving party to return all copies of an inadvertently produced document within a set number of business days after notification and prohibit them from using the information until the court rules on the privilege claim.11United States District Court Southern District of Florida. 502(d) Clawback Order Long Form Getting a 502(d) order in place early is one of the smartest moves a litigation team can make. It doesn’t eliminate the need for a thorough privilege review, but it provides a safety net if something slips through.
Sanctions for Getting It Wrong
The consequences for review failures go beyond losing a privilege claim. Under Rule 26(g), every disclosure and discovery response must be signed by at least one attorney of record. That signature certifies that the attorney conducted a reasonable inquiry and that the disclosure is complete and correct. If a court finds the certification was made without substantial justification, it must impose an appropriate sanction on the signer, the party, or both, which can include an order to pay the opposing party’s reasonable expenses and attorney’s fees caused by the violation.12Legal Information Institute. Federal Rules of Civil Procedure Rule 26 – Duty to Disclose; General Provisions Governing Discovery – Section: Signing Disclosures and Discovery Requests, Responses, and Objections
Spoliation sanctions under Rule 37(e) are more severe. When electronically stored information that should have been preserved is lost through a party’s failure to take reasonable steps, and the data can’t be recovered, the court can impose curative measures proportional to the prejudice. If the destruction was intentional, the court can presume the lost evidence was unfavorable, give an adverse inference instruction, or dismiss the case outright.2Legal Information Institute. Federal Rules of Civil Procedure Rule 37 – Failure to Make Disclosures or to Cooperate in Discovery
The gap between “we ran a reasonable process and missed some things” and “we didn’t bother trying” is where most sanctions fights play out. Documenting your methodology, quality control results, and validation statistics isn’t just good practice. It’s the evidence you’ll need if the other side claims your production was deficient.