Does AI-assisted review satisfy Rule 26 by itself?

Rule 26 states the scope and duties of discovery rather than approving a specific review technology. Counsel needs a method suited to the claims and defenses, as well as the privilege issues and proportionality factors in the case. AI may support that method when the record includes the population reviewed and task definition. It also needs validation and reviewer supervision, plus exceptions and change history. Technology procurement documents alone leave those matter-specific facts unproven.

What should be logged for an AI-generated review note?

Record the source document identifier with the matter and workflow version. Add the authenticated reviewer and the prompt template with its model identifier. The request ID and output classification come next. Record the reviewer action and time as well. Preserve the generated note as a derivative artifact linked to the source. Avoid placing full privileged text in a general observability store. The matter repository should hold protected content under its access and retention rules.

Can an AI gateway preserve evidence under Rule 37?

A gateway can preserve evidence about model requests routed through its HTTP boundary. Rule 37 preservation covers relevant ESI across the matter, including sources that never enter an LLM request. Legal holds and collection integrity remain separate duties. Source preservation and restoration do too. The gateway record can support the file by showing what crossed the model route and which policy applied at that moment.

Should a model make final privilege or production decisions?

The California guidance keeps professional judgment with the lawyer. It also calls for review and verification, followed by correction of AI outputs. A model can prioritize a privilege queue or propose a production action. A qualified reviewer should approve consequential privilege and redaction decisions, along with disclosure and filing decisions. Preserve the proposal and approval as separate events so later review can identify the human decision.

AI E-Discovery Needs a Reproducible Review and Production Record

An AI e-discovery workflow may rank collected email and summarize a custodian's documents. It may also draft review notes or assemble material for production. Each action changes the path between preserved evidence and the set delivered to another party. The defensible record therefore needs more than a model response. It needs the source population and preservation state. It also needs the prompt and model version plus the reviewer action, followed by the exception queue and production mapping that existed when the team made the decision.

TL;DR

AI-assisted review should preserve the source population and workflow version, plus the reviewer decision and exceptions for every consequential stage.
Federal Rules 26 and 34, plus Rule 37, make relevance, proportionality, usable production, and reasonable preservation steps concrete review points.
Lawyer supervision and confidentiality controls apply to prompts, uploaded documents, model outputs, and autonomous actions.
An AI gateway can govern routed HTTP model traffic while the legal team owns preservation and review methodology, privilege decisions, and production quality.

The workflow record starts with the preserved population

The Federal Rules of Civil Procedure published by the U.S. Courts place electronically stored information inside the discovery process. Rule 26 limits discovery to nonprivileged matter relevant to a claim or defense and proportional to the needs of the case. Rule 37 addresses loss of ESI that should have been preserved when a party failed to take reasonable steps.

An AI review system enters after counsel has identified custodians and sources, set date ranges and legal holds, and chosen collection methods. Its output remains tied to that population, which can change. If a mailbox enters the collection a week later, an earlier ranking run describes a different universe. The record should therefore bind each run to collection identifiers and hashes. It should also capture the hold and custodian scope in effect.

I would reject an AI review report that opens with an accuracy score and omits the population manifest. The shiny number hides the first question opposing counsel will ask: which material was available to the system? A printed manifest with a missing custodian row is more revealing than a polished model dashboard.

Legal AI governance gives the program-level controls. AI e-discovery needs a matter-level execution record beneath them.

Rule 26 requires a review method the team can explain

Rule 26 centers relevance and privilege, along with proportionality and disclosure. AI can help prioritize material, but counsel still needs a review method tied to the matter. Define the task in operational terms: rank for issue A or detect likely privilege for a separate queue. The system may instead summarize only reviewed documents or draft a proposed response for lawyer approval.

The control file should preserve the system instructions and prompt template. It should name the model and version together with the retrieval configuration, then record the sampling method and decision threshold where applicable. The file also needs the validation set and reviewer guidance. Adjudication notes and change approvals belong beside them. When the method changes, start a new version rather than silently overwriting the earlier run.

A useful acceptance test includes obvious responsive documents and hard negatives, privileged examples and attachments, plus scanned pages and near duplicates. Reviewers should record disagreement and resolution. That evidence shows how the team tested the workflow on the matter's own material. Generic vendor evaluation remains procurement evidence.

AI audit trails explains the event record. The matter file supplies the legal context that makes those events understandable.

Rule 34 keeps production format inside the control file

Rule 34 permits requests for documents and ESI. It also addresses the form of production. Material should be produced as kept in the usual course of business or organized to correspond with request categories. If a request leaves the form unspecified, production must use the ordinarily maintained form or another reasonably usable form.

AI summaries can help reviewers find material. The summary is usually a derivative artifact, however, and may omit an attachment relationship or spreadsheet formula. Hidden fields and message threading details needed for production may also be missing. The production process should point back to the source item and preserve family relationships with metadata, plus redactions with load-file mapping.

The workflow record should identify every AI-derived field. A review note generated by a model needs a label that separates it from source evidence and human annotation. If a model proposes redactions, a qualified reviewer should approve each consequential redaction before the production set closes. The final export then receives its own manifest and validation report.

My preference is blunt: treat generated summaries as review aids until a human signs the production decision. E-discovery punishes elegant shortcuts that sever a conclusion from its source record.

Confidentiality and supervision apply to the model route

The State Bar of California's 2026 Practical Guidance for generative and agentic AI addresses competence and confidentiality, along with communication and candor. Lawyers must understand material capabilities and risks. They must also review and correct AI outputs while preserving professional judgment. The guidance warns that prompts and uploaded material may be used or shared, then stored or transmitted in ways that create confidentiality concerns.

For e-discovery, provider eligibility belongs in the matter's control file. The team should record data-use terms and retention settings. It should also record processing location where relevant and approved matter classifications. Access policy should bind the user and matter to the model route. A reviewer assigned to Matter A should not gain prompt access to Matter B because both matters share a review platform.

Agentic features need an action boundary. A system may draft a search or tag a document. It may also propose a response. Filing and production release, along with privilege calls and substantive legal conclusions, need meaningful lawyer supervision. The review platform should preserve the proposed action and the approving person. It should also preserve the time and resulting artifact.

Gateway evidence covers one bounded segment

An LLM gateway sees an authenticated HTTP request after the review application has assembled prompt content. It can evaluate supplied user and matter context, classify the payload, and enforce an approved model destination. It can then inspect the response and write a decision record. That path can prove whether a routed model request was allowed or denied under a versioned policy, including any modification made before the decision.

The gateway has no view of a mailbox that the collection process missed. It cannot prove that a legal hold reached every source, evaluate the statistical validity of a review protocol, or guarantee that the final load file preserves family relationships. Those controls belong to legal and collection owners, together with the review and production owners.

Join the records with stable identifiers. The matter run should reference the gateway request ID. The gateway record should carry the matter and workflow version supplied by the application, without copying full privileged content into general telemetry. LLM audit log retention covers storage after the request decision exists.

A matter acceptance file should survive a later challenge

Build one acceptance file for each material workflow version. Include the source-population manifest and preservation reference. Follow them with the task definition and reviewer protocol. Add the model and prompt identifiers with the retrieval settings. Test fixtures and exception results come next, followed by approvals. Finish with production mapping and a signed release decision.

Run negative tests at the HTTP boundary. Use a reviewer outside the matter, then use an unapproved model route. In a separate test, use a prompt containing a restricted classification. Preserve each denial with policy version and request ID. Test approved traffic and confirm that the response returns to the correct matter workspace.

Run separate legal-process tests. Remove an attachment from a family, then verify production validation catches it. Change a custodian population and confirm that the workflow version becomes stale. Introduce a model-generated citation to a nonexistent document and confirm reviewer verification blocks it. The gateway contributes evidence to only the routed model tests.

The file shows how the team preserved and reviewed evidence, while also showing how model traffic was controlled.

DeepInspect

DeepInspect governs authenticated user or agent traffic deliberately routed through an HTTP LLM endpoint. A legal review application can supply the user and matter, along with the workflow version and requested destination. DeepInspect can then apply identity-aware data and route policy before inspecting the response. It produces a signed, tamper-evident decision record tied to the request ID.

Preservation and review methodology remain with the legal and e-discovery teams, as do privilege judgment and production validation. DeepInspect supplies evidence for the model-request segment inside that larger matter file. Book a technical deep dive at deepinspect.ai.