“Search, Forward” with AI: Court Approves AI-Assisted Review Workflow for eDiscovery
Editor’s Note: Courts have been asked, in one form or another, whether AI-assisted review deserves the same judicial confidence that technology-assisted review (TAR) earned over a decade ago. In this article, Phil Favro examines Schulte v. LinkedIn Corporation, where a federal court approved a defendant’s AI-assisted review workflow and rejected the plaintiffs’ demands for additional disclosure metrics. Favro draws a direct line between this decision and the TAR case law that shaped discovery a decade earlier, showing how courts are applying familiar reasonableness and proportionality standards to a new generation of technology. The ruling offers producing parties real reassurance that pre-culling search terms and AI-assisted workflows can withstand scrutiny, while also reminding both sides that ESI protocols remain the operative rulebook for what must be disclosed. Favro’s analysis goes further, unpacking what requesting parties can still argue from TAR-era precedent and where “discovery on discovery” challenges may still gain traction. As courts continue to weigh in on AI’s role in ESI review, Schulte stands as an early but significant signal of where that jurisprudence is heading.
“Search, Forward” with AI: Court Approves AI-Assisted Review Workflow for eDiscovery
By Phil Favro, Contributing Author for HaystackID
A significant development recently occurred regarding the use of artificial intelligence (AI) for identifying, reviewing, and producing responsive electronically stored information (ESI). In Schulte v. LinkedIn Corporation, a San Francisco-based federal court issued an order approving a defendant’s AI-assisted review workflow.[1] In doing so, the court rejected the plaintiffs’ challenges to aspects of the AI-assisted workflow, along with demands that the defendant disclose various metrics concerning the workflow.
Approximately a year ago, I openly queried whether courts would embrace the technological opportunities that AI offered for eDiscovery.[2] Referring to the article that U.S. Magistrate Judge Andrew Peck (ret.) authored for Law Technology News—“Search, Forward”—I highlighted the parallels between the dawn of technology-assisted review (TAR) for eDiscovery in 2011 and the advent of AI-assisted review in 2025. While it was apparent last year that lawyers and litigants were using AI for search and review, I observed that they were also waiting for courts to issue an order declaring that AI—like TAR in 2011—is “better than the available alternatives, and thus should be used in appropriate cases.”[3]
Schulte is tantamount to that order, i.e., a court decision approving the use of AI in connection with the ESI search and review process. To be sure, Schulte does not declare that AI-assisted review is “better than the available alternatives.” Nor does it encourage the use of AI “in appropriate cases.” Instead, Schulte narrowly handles the instant challenges the plaintiffs raised regarding the defendant’s use of AI. Nevertheless, Schulte appears to offer judicial imprimatur regarding AI-assisted review in the same way that Moore v. Public Group provided for TAR. That is a noteworthy development. Moreover, Schulte provides any number of takeaways for clients and counsel who are using generative AI to facilitate their document review efforts.
Schulte and AI-Assisted Review
Schulte is a putative class action that involves antitrust claims against the defendant, LinkedIn Corporation (LinkedIn). The claims arise from paid services LinkedIn offers to its subscribers and alleged overcharges for those services. In discovery, the plaintiffs served their document requests and then requested that LinkedIn disclose its proposed search methodology for identifying responsive documents.
Disclosures regarding the AI-Assisted Review Workflow
LinkedIn disclosed to the plaintiffs that it developed search terms (“twenty-five search strings”) and that it applied those search terms to the universe of potentially responsive information. The result was that LinkedIn narrowed its review to a subset of 204,444 documents from a universe that apparently included terabytes of data.
LinkedIn also informed the plaintiffs that it intended to use AI-assisted review technology to facilitate its document production efforts. In particular, LinkedIn represented that it would use AI to “assist in filtering out non-responsive documents” from the 204,444 subset. LinkedIn also indicated that it would generally rely on AI to handle its review of the remaining subset of documents. As part of that process, LinkedIn would implement quality control measures such as “validating sample populations” through attorney reviewers to ensure the production was reasonable and proportional under the circumstances.
The Plaintiffs’ Concerns
In response, the plaintiffs expressed various concerns about LinkedIn’s proposed search methodology and AI-assisted review workflow. In particular, the plaintiffs objected to LinkedIn’s use of search strings to pre-cull the document universe. The plaintiffs asserted that LinkedIn’s selected AI technology does not contemplate the use of search terms. Instead, they maintained that the AI technology had been designed to both ingest and then make responsiveness calls on an entire data set. Nor would it be appropriate for LinkedIn to use search terms in connection with AI-assisted review. According to the plaintiffs, this is because the search strings would “artificially reduce the ‘target population’ ingested into the [AI] product, needlessly removing potentially responsive documents.”
During the meet and confer process, the plaintiffs also demanded that LinkedIn disclose additional information about its AI-assisted review technology and process. They requested several details including:
- Whether LinkedIn would use a seed or training set of documents;
- Validation metrics including error, recall, and precision rates;
- How LinkedIn would address coding conflicts between the AI technology and attorney reviewers; and
- Whether the AI technology would merely prioritize documents for review or make final calls on the responsiveness of documents.
LinkedIn’s Positions
LinkedIn responded to the plaintiffs’ concerns with various disclosures. First, LinkedIn clarified that its AI-assisted review workflow did not require a seed or training set of documents. In addition, LinkedIn explained that its selected AI technology would make final responsiveness calls and that its quality control measures would include attorney review of “samples taken from each responsiveness type.” Relying on the stipulated ESI protocol the court entered as an order, LinkedIn declined to make additional disclosures regarding its review workflow. LinkedIn also maintained that its use of pre-culling search strings was proper and complied with ESI protocol requirements. Finally, LinkedIn argued that it would be unduly burdensome to run the entire universe of potentially responsive information through AI.
The Court Approves the Use of Search Terms
The court (Magistrate Judge Laurel Beeler) agreed with all of LinkedIn’s positions and denied the plaintiffs’ discovery motion.
As an initial matter, Judge Beeler rejected the plaintiffs’ argument that LinkedIn should not be allowed to use search terms in connection with AI-assisted review. Judge Beeler observed that the plaintiffs had not raised any deficiencies with LinkedIn’s search terms. That alone doomed the plaintiffs’ search term challenge.
In addition, the court concluded that TAR case law—which allowed producing parties to use “search terms to pre-cull documents before providing them to technology-review platforms”—was equally applicable to AI-assisted review.[4] In particular, Judge Beeler emphasized that pre-culling search terms satisfied “the reasonableness and proportionality standards of Rules 26(b) and 34(b)(2).” Moreover, it would be unduly burdensome for LinkedIn to run its entire universe of documents through the AI technology: “LinkedIn would be required to feed multiple terabytes of documents into [the AI technology], resulting in significant costs related to processing, hosting, and human review.”
Accordingly, the court approved LinkedIn’s workflow blending search terms and the AI technology. Judge Beeler also indicated that the plaintiffs could raise with LinkedIn any concerns they had with the search terms and ordered the parties to meet and confer on this issue.
The Court Rejects the Plaintiffs’ Disclosure Demands
In addition, Judge Beeler rejected the plaintiffs’ demands that LinkedIn disclose several additional metrics regarding its use of AI. According to the court, the plaintiffs requested details such as “elusion estimates, the document error rate, and the number of human reviewers being used to validate [the AI technology’s] predictions.”[5]
Those requests, however, were not appropriate for two reasons. First, Judge Beeler found that LinkedIn met its disclosure obligations under the ESI protocol. LinkedIn had revealed that it would use AI to remove nonresponsive information from its universe of documents, along with several other items regarding its ESI workflow and its selected technology. As Judge Beeler reasoned, “[t]hese disclosures more than satisfy the demands of the Interim ESI Order.”
Second, Judge Beeler considered the metrics the plaintiffs demanded to be “discovery on discovery,” which is typically disfavored and generally authorized only upon a showing of a “specific deficiency” with the producing party’s response efforts. In this instance, the plaintiffs had only shown that the subset of LinkedIn’s document universe (after running the 25 search strings) was 204,444 documents. That, by itself, was not enough to establish a process or response deficiency.
Takeaways from Schulte for AI-Assisted Review
Judge Beeler’s order in Schulte is the first of its kind and provides the type of “Search, Forward,” general endorsement that lawyers and clients have long sought for AI-assisted review. Beyond the court-sanctioned approval of LinkedIn’s AI-assisted review workflow, Schulte offers other instructive takeaways for those using or considering AI-assisted review.
First, it’s noteworthy that Schulte adopted TAR-era case law to sanction LinkedIn’s use of search terms to pre-cull its document universe before running data through the AI-assisted review process. This has long been the majority rule under TAR case law given the high costs of applying TAR to all potentially responsive custodian data. With Schulte extending that rationale to AI-assisted review, producing parties should feel some assurance that courts may examine their AI workflows through the lenses of proportionality and reasonableness that courts have often used to address TAR use questions.
Nevertheless, requesting parties may feel some assurance from Schulte as well. While Schulte rejected the workflow and disclosure issues the plaintiffs raised, requesting parties could persuade other courts that different aspects of TAR-era case law favorable to them—such as the cooperation and disclosure requirements from Progressive v. Delaney and In re Valsartan—should apply to AI-assisted review workflows.[6]
Next, Schulte highlights that courts are reluctant to authorize “discovery on discovery,” absent fact-specific good cause. This offers producing parties an opportunity to get their process “right” without undue meddling or the need for vast disclosures of information relating to their AI-assisted review workflow. And yet, the opportunity still exists for “discovery on discovery,” particularly if producing parties fumble their AI-assisted review process or otherwise fail to make productions that satisfy reasonableness and proportionality standards.
A final key point arising from Schulte is the importance of ESI protocols. Judge Beeler repeatedly emphasized that LinkedIn’s disclosures satisfied the court-ordered ESI protocol’s requirements. Had the ESI protocol mandated additional disclosures that LinkedIn failed to make, perhaps the result from Schulte would have been different.[7] Whether entered as an order or a Rule 29 stipulation, ESI protocols have long been considered determinative of parties’ rights and responsibilities in discovery. Accordingly, it behooves both requesting and producing parties to examine the provisions in a proposed protocol before it becomes an order governing how the parties will handle discovery.
Schulte is undoubtedly the first of its kind, reflecting a “Search, Forward” moment for AI and eDiscovery. It will not be the last. Expect other court decisions to impact the AI-assisted review landscape in the months and years to come.
[1] Schulte v. LinkedIn Corp., No. 22-CV-00237-HSG (LB), 2026 WL 1905851 (N.D. Cal. July 1, 2026). [2] Phil Favro, Will Courts “Search, Forward” with eDiscovery Case Law on AI?, HaystackID Blog (Aug. 7, 2026). [3] Moore v. Publicis Groupe SA, 287 F.R.D. 182, 191 (S.D.N.Y. 2012). [4] The court cited Livingston v. City of Chicago, No. 16-cv-10156, 2020 WL 5253848 (N.D. Ill. Sep. 3, 2020) and In re Biomet M2a Magnum Hip Implant Prods. Liab. Litig., No. 3:12-MD-2391, 2013 WL 1729682 (N.D. Ind. Apr. 18, 2013) as supporting its determination on this issue. [5] In the parties’ joint discovery letter brief, the plaintiffs demanded “(1) the recall, precision, and elusion estimates with 95% confidence intervals; (2) the richness estimate and validation sample size used; (3) whether ‘borderline’ documents were treated as relevant in the validation . . . (4) whether human reviewers were shown [the AI technology’s] predictions while coding the validation sample—which . . . ‘could influence the reviewer too heavily,’ . . . (5) the number of human reviewers used to validate [the AI technology’s] predictions, and their titles; (6) the document error rate and the disposition of documents [the AI technology] could not analyze (e.g., corrupted images that have been included in LinkedIn’s productions to-date); (7) the withholding and reversal rate for all human-based quality control; and (8) given LinkedIn’s rolling, multi-repository productions, the validation results for each population or wave (e.g., as documents from additional custodians, Google Drive, and Slack are added), as [the AI provider’s] own guidance instructs that re-validation may be required when the population changes.” Schulte v. LinkedIn Corp., No. 22-CV-00237-HSG (LB) (N.D. Cal.), ECF No. 186 (Discovery Letter Brief Regarding LinkedIn Search Methodology), dated June 23, 2026. [6] Progressive Cas. Ins. Co. v. Delaney, No. 2:11-CV-00678-LRH, 2014 WL 3563467 (D. Nev. July 18, 2014); In re Valsartan, Losartan, & Irbesartan Prods. Liab. Litigation, 337 F.R.D. 610 (D.N.J. 2020). [7] See James v. Cerebras Systems, Inc., No. 4:25-cv-09361-AMO (N.D. Cal.), ECF No. 74 (Stipulation and Order regarding the Production of Electronically Stored Information (“ESI”) and Hard Copy Documents), dated July 7, 2026 (imposing by stipulation and court order far more extensive disclosure obligations regarding AI-assisted review than Schulte).
About Phil Favro
Phil Favro is the founder of Favro Law PLLC, where he counsels clients on ESI, AI, and discovery issues and serves as a special master, mediator, and expert witness. Phil is nationally recognized for his expertise on ESI, discovery, and information governance, with courts acknowledging his credentials. See, e.g., Oakley v. MSG Networks, Inc., No. 17-CV-6903 (RJS), 2025 WL 2061665 (S.D.N.Y. July 23, 2025). This background makes Phil particularly well-suited to counsel clients and advise courts on information-related issues. As a special master, Phil is acclaimed for his collaborative approach, working with parties to find stipulated solutions to complex issues. For disputes that require adjudication, he is renowned for the clarity and vigor of his written dispositions, which are available on legal search engines.
About HaystackID®
HaystackID® solves complex data challenges related to legal, compliance, regulatory, and cyber requirements. Core offerings include Global Advisory, Cybersecurity, Core Intelligence AI™, and ReviewRight® Global Managed Review, supported by its unified CoreFlex™ service interface and eDiscovery AI® technology. Recognized globally by industry leaders, including Chambers, Gartner, IDC, and Legaltech News, HaystackID helps corporations and legal practices manage data gravity, where information demands action, and workflow gravity, where critical requirements demand coordinated expertise, delivering innovative solutions with a continual focus on security, privacy, and integrity. Learn more at HaystackID.com.
Assisted by GAI and LLM technologies.
SOURCE: HaystackID