For years, litigators have watched discovery practice strain against a growing gap: the governing ESI protocols in most active cases were drafted before hyperlinked cloud files replaced traditional email attachments, before Slack and Teams became primary business-communication channels, and well before generative AI tools touched document-review workflows. Courts have grown less patient with that gap. In James v. Cerebras Systems Inc., No. 4:25-cv-09361-AMO (N.D. Cal. July 7, 2026), the Northern District of California entered what is likely the first single court order to address cloud-hosted hyperlinked documents, mobile and short-message data, and generative AI training-data workflows together — under one stipulated ESI protocol. This article examines the framework the court adopted, identifies the obligations it places on counsel, and provides practical guidance for in-house counsel and litigators who need to assess whether their current protocols can survive comparable scrutiny.
The Problem Legacy Protocols Were Not Built to Solve
Electronically stored information ("ESI") — the full universe of digital data subject to preservation and production obligations — no longer lives primarily in email inboxes. A Google Drive link embedded in a message is not an attachment in any traditional sense; it is a pointer to a document that may be revised, moved, or deleted independently of the message that references it. The same is true of SharePoint-hosted files shared via URL in Microsoft Teams, or documents linked inside a Slack channel thread. Courts and practitioners have begun calling these "modern attachments," and they present a foundational challenge: are they part of the document that contains the link, or are they separate ESI that must be identified and collected on their own terms?
At the same time, organizations are increasingly using large language models ("LLMs") and other generative AI tools inside their document-review and workflow systems. When AI-assisted review touches potentially responsive data, questions arise about validation, auditability, and whether an automated process can inadvertently alter or discard material that would otherwise be subject to a litigation hold.
James v. Cerebras arose in this environment and required the court to address both problems simultaneously.
What the Protocol Requires
The order's most significant structural choice for hyperlinked files is what might be called a targeted, manual production model. Rather than requiring a producing party to bulk-collect every document linked anywhere in a produced corpus — an obligation that could generate overbroad data dumps and proportionality objections under FRCP 26(b)(1) — the protocol requires the requesting party to identify specific produced documents that contain relevant hyperlinks and then separately request those linked files. The burden of identifying what to ask for rests with the requester; the burden of producing the identified file rests with the producer. That allocation keeps discovery proportionate and avoids the default-to-everything problem that has plagued cloud-native productions.
For generative AI workflows, the order requires that validation parameters be defined and disclosed before AI-assisted review begins, not after a dispute arises. The practical effect is prophylactic: by anchoring AI governance to pre-review disclosure, the protocol substantially reduces the surface area for spoliation claims and procedural deadlocks that emerge when parties learn mid-case that an undisclosed algorithm made consequential decisions about responsiveness or privilege.
Practical Guidance for In-House Counsel and Litigators
Audit your current ESI protocol against the Cerebras framework. Review any standing protocol or proposed protocol in active matters for three explicit provisions: (i) a definition and handling procedure for hyperlinked or cloud-hosted documents; (ii) a short-message and mobile-data section covering Teams, Slack, Signal, and comparable platforms; and (iii) an AI-governance clause that requires pre-review disclosure of any automated or AI-assisted process touching responsive ESI. Absence of any one of these is now a documented gap against a standing N.D. Cal. precedent.
Negotiate hyperlink-handling obligations before the Rule 26(f) conference closes. The targeted, manual production model adopted in James v. Cerebras is a workable framework, but only if both parties understand their respective obligations from the outset. Counsel should propose explicit language identifying which platforms generate hyperlinked content, the mechanism for requesting linked files, and any applicable time limits for those secondary requests.
Define AI validation parameters in writing before review begins. If your organization or outside counsel intends to use any AI-assisted tool — whether for culling, privilege review, or responsiveness determinations — document the tool, the validation methodology, and the error-rate benchmarks before the process starts. That record is your defense against a later spoliation or process-integrity challenge.
Treat short-message data as a first-class custodian population. Mobile device data and ephemeral-messaging platform content should appear in your litigation-hold notice and your collection protocol as named data sources, not as afterthoughts appended to email. The court's explicit inclusion of short-message data in the Cerebras protocol signals that omitting it is no longer a defensible oversight.
Update template protocols now, not at the next dispute. An outdated template circulated in a new matter is itself a risk. In-house legal operations teams should treat James v. Cerebras as a trigger to revise standard ESI protocol templates so that the next case begins from a compliant baseline.
Conclusion
In conclusion, James v. Cerebras Systems marks a meaningful inflection point: a federal court has now encoded into a single order the practical reality that modern litigation data lives in cloud-linked documents, short-message platforms, and AI-assisted workflows — and that discovery obligations must follow the data wherever it resides. There is no one-size-fits-all approach to ESI protocol drafting, and the specific proportionality balance struck in James v. Cerebras will not translate identically to every case or jurisdiction. However, the order's core architecture — targeted hyperlink production, mandatory AI-governance disclosure, and explicit short-message handling — provides a model that courts in other districts are likely to find persuasive. Counsel who build these provisions into their protocols before a dispute arises will be better positioned to manage discovery cost, resist sanctions exposure, and demonstrate the good-faith compliance that judges increasingly expect when parties are working with modern data environments.