The Legal Discovery Process: From Review to Analysis

A litigation team opens a new matter and faces a familiar problem. The other side has produced several hundred documents that need to be reviewed and the deadline is weeks away.. The traditional calculation begins: How many contract attorneys will the review require, for how many weeks, and at what cost?

That calculation changes when an AI model can review the entire production overnight and return the documents sorted by issue, flagged for privilege, and linked to their sources. The team must then answer two different questions: What should it do with the time it has recovered, and can it trust the model's findings?

Getting discovery wrong can be costly in several ways. A missed document surfaces at a deposition. A privileged file goes out by mistake. A deadline slips, and the court takes notice. Sanctions can follow, and the client bears the consequences. Teams that treat discovery as paperwork to survive tend to approach settlement from a position of weakness because they never learned what their own records said. This article examines where discovery costs are concentrated, how the balance between review and analysis is changing, what makes AI-assisted review defensible, and how legal teams can implement it.

The Six Stages of the Legal Discovery Process

The legal discovery process is the pretrial phase in which opposing parties exchange evidence and information relevant to a lawsuit. It includes the preservation, collection, processing, review, analysis, and production of electronically stored information. Legal document review has traditionally consumed most of the discovery budget, which is why AI for legal discovery is reshaping how legal teams conduct document review and analysis.

Discovery follows a familiar sequence captured in the Electronic Discovery Reference Model. Each stage moves the data closer to a form that a party can use in the case.

1. Preservation

During preservation, a party protects data that could matter once litigation is reasonably anticipated. Allowing that data to disappear, even through routine deletion, can expose the party to sanctions.

2. Collection

Collection gathers preserved data from sources such as email accounts, shared drives, and devices. Bringing that data together completely and intact helps keep the later stages defensible.

3. Processing

Processing prepares the collected data for review. It removes duplicates, converts files into a consistent format, and organizes the material so reviewers can work through it efficiently.

4. Review

During review, lawyers examine the collected material and decide what is responsive, what is privileged, and what matters to the case. This stage has traditionally consumed most of the discovery budget.

5. Analysis

Analysis helps a legal team understand the facts, the narrative the documents support, and the legal theory that may emerge from the record. During this stage, the team builds a chronology and identifies the documents that could decide the matter.

6. Production

Production delivers responsive, non-privileged material to the other side. The parties often negotiate the format and timing, subject to the court's requirements.

The bulk of discovery effort and expense is concentrated in review and analysis. Lawyers spend most of their time in these stages, and AI is changing both. The rest of this article focuses on review and analysis.

The Cost of the Legal Discovery Process

The cost of discovery has long been concentrated in document review. Where the Money Goes, a study from the RAND Institute for Civil Justice, places document review at roughly 73 percent of the cost of producing electronic discovery. Collection and processing trail far behind.

Advances in legal software have reduced that burden for more than a decade. Predictive coding, often called technology-assisted review, made first-pass review meaningfully faster. Much of its credibility traces to a landmark 2011 study by Maura Grossman and Gordon Cormack, a legal scholar and a computer scientist. Their research gave the field rigorous evidence that technology-assisted review can be at least as accurate as exhaustive manual review in far less time. Courts and practitioners have leaned on that finding ever since. Yet review stayed the largest line on the bill, because the volume of data kept climbing faster than the tools could cut it.

Review remained expensive because records now live in more places, including email, chat, shared drives, collaboration tools, and messaging apps. A single custodian can generate more data in a month than an entire matter produced a generation ago. Although the tools improved, data volumes grew even faster.

The current shift builds on that foundation. Legal AI adds an analysis layer to the eDiscovery stack. Existing tools continue to collect, process, and host data, while legal AI reviews and reasons over the documents. The eDiscovery tools prepare and organize the data; legal AI adds first-pass review and preliminary analysis. Lawyers remain responsible for verifying the output, applying legal judgment, and deciding how the findings affect case strategy.

Earlier tools helped lawyers prioritize documents for human review. A model can review all of them and explain each classification in plain language. As a result, first-pass review becomes less of a bottleneck, and analysis can begin earlier.

AI Document Review and the Move From Review to Analysis

As the cost of first-pass review falls, legal teams can devote more time to analysis. That work includes building a chronology, conducting targeted legal research, testing a legal theory against the record, identifying the handful of documents that could decide a matter, and examining the other side’s production for gaps that could change the case.

A chronology built from thousands of emails can expose a timeline the other side would rather not explain. A single document found early can reframe a settlement conversation. This work can influence outcomes, but it often had to wait until first-pass review was complete.

Teams using legal AI report a similar pattern. At Carvana, the legal team estimates that each lawyer now recovers 7 to 10 hours a week and uses that time to advise clients, develop strategy, and exercise judgment. Its structured litigation workflows have saved more than 800 hours so far this year, and the lawyers spend the recovered time on the counseling and case strategy needed to keep pace with the business.

Similar results appear elsewhere. Independent research from RSGI shows that the most active users save around 11 hours a week. Teams can redirect time from initial document reading to legal judgment and case strategy.

Harvey operates in this layer. It helps litigators review and categorize large sets of documents, analyze deposition transcripts and expert materials, build chronologies, and identify gaps in factual and legal analysis. The AI performs first-pass review, categorization, extraction, and preliminary analysis. Lawyers verify the output, apply legal judgment, and determine how the findings affect case strategy. The model's conclusions are starting points rather than final legal determinations.

The Verification Standard for AI-Assisted Review

Discovery has always depended on defensibility. A party must be able to explain to a court and its opponent how it searched, reviewed, and produced documents. When technology-assisted review emerged, judges asked whether a process that allowed a machine to cull documents could meet that standard. Decisions such as Da Silva Moore v. Publicis Groupe indicated that it could, provided the party supported its method with sound statistical sampling and validation.

That standard assumed that a human could not read everything, so it accepted a defensible sample as evidence of a defensible whole. AI changes that premise. When a model can review the entire set, the question becomes whether a lawyer can verify each decision. Defensibility depends on verification, which requires the ability to compare the model's reasoning with the underlying document.

Privilege is where this matters most. Producing a single privileged document by mistake can waive protection over an entire subject, and the party may need to show the court that its review was reasonable. A model that flags every potentially privileged document and explains why it flagged each one gives reviewers a defensible basis for the decisions they approve. A record of those decisions can support the party if the production is challenged.

The tool's design determines whether reviewers can perform that check. Legal AI can produce fluent, confident text without a reliable way to trace the answer to its source. A tool built for legal work links each answer to the underlying material so reviewers can open the document and confirm a privilege or responsiveness decision. That traceability helps turn fast review into defensible review.

Harvey is designed around that requirement for traceability. It grounds its answers in sources a legal team trusts and keeps its reasoning open to inspection, allowing a reviewer to verify each decision against the underlying document. Harvey's report The Legal AI Governance Imperative in Practice describes legal teams moving toward evidence-based oversight and treating auditability as a standing requirement. Both practices align with what a court expects when a party defends its process.

The duty remains with the lawyer. The duty of technology competence requires lawyers to understand the tools they use well enough to supervise them. A qualified lawyer reviews the model's output before the team relies on it, and that supervision is part of a defensible process.

How to Run Document Review With AI

Understanding how to use AI as a lawyer begins with defining the standards the model will apply. In document review, the most useful setup applies the matter’s issue codes, privilege criteria, and legal theory so the review reflects the matter and draws on the organization’s precedent. Generic criteria produce generic results, and legal matters are rarely generic.

Teams should define a review protocol for the matter and run first-pass review against that protocol. Reviewers should then verify a representative sample of the model's decisions and examine every high-stakes or borderline call, including close privilege questions. The team should also maintain a record of how each decision was made so it can explain the process if a court or opposing party asks. This procedure allows the team to measure performance and correct errors before producing documents.

Organizations using this approach report concrete results. Carvana turned its templates into reusable, logic-driven review workflows that operate inside Microsoft Word and perform issue spotting at the level of a mid-level associate. At Bridgewater, a supplier contract review that once took two days now takes two hours. The legal team has also reduced the time required to review large batches of bespoke trading agreements by more than 95 percent.

These workflows are easier to incorporate into legal practice when they operate within the tools teams already use. Harvey works within Word, Outlook, and related tools and can apply an organization’s precedent to the review, helping keep the work consistent across a matter while leaving the reviewer in control.

The Shifting Balance Between In-House Teams and Outside Counsel

The changes in discovery are also affecting the relationship between in-house teams and the firms they hire. According to research from RSGI, more than half of the organizations studied now describe the tool as foundational to their legal work, and 89 percent of firms say it allows them to take on more work. In-house teams are increasingly performing work they once sent to outside counsel.

That additional capacity changes what an in-house team sends out. At the Adecco Group, a legal department spanning 45 countries has reduced its reliance on outside counsel by 5 to 10 percent and retains more work internally, using Harvey for tasks ranging from contract review to litigation preparation and strategy modeling. The team's lawyers report saving up to 8 hours a week and directing that time to work they choose to keep. Outside counsel can then focus on validating analyses and advising on the most difficult questions.

The effect differs on each side. An in-house team can perform more early analysis internally and purchase outside judgment with greater precision. A law firm can demonstrate its AI capabilities and resulting capacity as clients direct work to firms that can use these tools effectively. Both groups can use discovery analysis as a source of competitive advantage.

None of this removes the need for outside counsel. It shifts outside counsel's work toward the judgment calls that carry the most risk. Firms that work with in-house teams as analytical partners are more likely to retain that work.

The Legal Discovery Process and Where Cases are Won

AI does not eliminate legal judgment. It reduces the time required for first-pass review and allows legal teams to begin chronology development, issue analysis, and case strategy earlier. A defensible process still requires traceable sources, documented methods, and lawyer oversight.

Within that lawyer-led process, Harvey can review and sort large document sets, build chronologies, identify gaps in a case, and ground its answers in sources that lawyers can inspect. By applying an organization’s standards, it helps teams move from faster review to earlier, better-supported analysis.

To see how Harvey could support a specific discovery matter, book a demo.

Frequently Asked Questions About Legal Discovery

How long does the discovery process take?

Discovery often runs for months, and in large cases it can extend beyond a year. The timeline depends on the volume of data, the number of parties, and the number of disputes over what must be produced. Review and analysis usually require the largest share of that time.

What is the difference between discovery and eDiscovery?

Discovery is the pretrial exchange of evidence in a lawsuit. The term eDiscovery refers to discovery involving electronic information, such as email, chat messages, and files. Almost all modern discovery is eDiscovery because most records originate in digital form and never exist on paper.

Is AI-assisted document review defensible in discovery?

AI-assisted document review can be defensible when a qualified lawyer remains in control. Courts have long accepted technology-assisted review, and AI can fit within a defensible process when the tool grounds its answers in sources that reviewers can check and a lawyer verifies the decisions. Maintaining a clear record of the process further supports its defensibility.

Why is document review the most expensive part of discovery?

Review requires a person to make a judgment about each document by deciding what is relevant, what is privileged, and what matters to the case. That work has traditionally required many hours of work by lawyers and contract attorneys, which is why review has accounted for most of the cost of producing discovery.

What is a technology-assisted review?

Technology-assisted review, sometimes called predictive coding, uses document review software to prioritize and classify documents in discovery. A lawyer trains and validates the tool, which then applies those judgments across a large document set. Technology-assisted review made review faster and helped set the stage for AI in the discovery process.