AI-Assisted eDiscovery: What Law-Firm Leaders Should Preserve Before the Process Is Challenged
- Pouya Shafabakhsh
- 7 days ago
- 4 min read
1. Why the Model Is Not the Whole Story
Scope Comes Before Model Quality
Generative AI can accelerate responsiveness review, but a model only sees the population it receives. A high-performing system cannot recover documents that were never collected, omitted through repository choices, or removed through an untested pre-culling boundary. Scope therefore becomes a litigation-management issue before it becomes a model-performance issue. Senior counsel should know which custodians, repositories, date ranges, and source types were inside the review population—and which were not.
Material Changes Need a Decision Trail
Complex discovery does not remain static. Search terms change, custodians appear, repositories are added or removed, instructions evolve, and sampling can reveal unexpected patterns. Those changes are often legitimate. The risk is that a consequential adjustment later cannot be reconstructed. A firm does not need a memorandum for every operational choice, but material changes affecting what can be found, reviewed, withheld, or produced should remain explainable.
2. What the Schulte Order Actually Decided
The June 30 Order Was Case-Specific
In Schulte v. LinkedIn Corp., No. 22-cv-00237-HSG (LB), the Northern District of California addressed LinkedIn’s planned use of Relativity aiR alongside 25 disclosed search strings. Plaintiffs sought to prohibit pre-culling, require aiR across all custodial files, and compel additional metrics including elusion estimates, error rates, and reviewer counts. Magistrate Judge Laurel Beeler denied those requests on the record before the court because plaintiffs had not shown a specific deficiency in the disclosed strings or resulting production.
The Court Still Required Continued Conferral
The order was not a declaration that the search strings were permanently beyond challenge. The court required the parties to meet and confer within 21 days regarding the strings and whether adjustments were warranted. That detail matters. The practical lesson is not that GenAI review received blanket judicial approval; it is that the dispute remained tied to concrete deficiencies, proportionality, burden, and continued conferral.
Text Messages Show Why Preservation Is Contextual
The same order denied broader text-message discovery from designated custodians. The interim ESI order excluded specified messaging sources absent good cause, and the record supported LinkedIn’s selected repositories as reasonable efforts in the circumstances. The result should not be generalized into a rule that text messages are unimportant. Preservation remains fact-specific and should be revisited when claims, custodians, communication habits, or evidentiary needs materially change.
3. Four Records That Matter More Than a Vendor Label
Source Map, Decision Log, Validation File, Privilege Boundary
A modest four-record structure can make AI-assisted discovery more explainable without building a bureaucracy. A source map identifies principal custodians, repositories, date ranges, collection boundaries, and material exclusions. A decision log captures consequential workflow changes. A validation file records what was tested, what exceptions appeared, and what corrective action followed. A privilege boundary separates operational facts from counsel’s legal impressions, mental theories, and litigation strategy.
Validation Should Follow Litigation Risk
A request for another party’s internal AI-review metrics is not automatically proportional, but that does not make internal validation irrelevant. Useful testing follows actual failure modes. Known-responsive documents can test population and instruction quality. Targeted sampling can look for missed responsiveness or clustered errors. Privilege-sensitive categories may justify extra checks. The objective is not a decorative accuracy percentage; it is evidence that gives counsel a reasoned basis for professional judgment.
4. Leadership Questions Before the Challenge
Before a Rule 26(f) conference, a major production, or a dispute about AI-assisted review, senior litigation leadership can ask five questions: What custodians and repositories are inside the population, and what is outside it? What material search, pre-culling, or workflow decisions shaped that population? What validation evidence would alert the team to a consequential miss? Which preservation decisions should be revisited as the matter changes? Could the process be explained proportionately without unnecessarily exposing privileged strategy? Those questions are intentionally technology-neutral. They can be used whether the review relies on search terms, TAR, GenAI, human reviewers, or a combination. They also encourage an earlier conversation between trial counsel, eDiscovery specialists, vendors, and information-governance teams about what evidence should exist before a dispute hardens. That matters because the best reconstruction record is usually created during the work, when reasons, exceptions, and tradeoffs are still known. Models, vendors, and review teams change. The durable question is whether consequential choices can still be reconstructed from evidence rather than memory.
5. FAQ: AI-Assisted eDiscovery After Schulte
Did Schulte broadly approve generative-AI review?
No. The June 30, 2026 order resolved specific discovery motions on the record before the court. It should not be treated as universal approval of GenAI review, Relativity aiR, or any other workflow.
Did the court require LinkedIn to disclose AI metrics?
No. Requests for items including elusion estimates, error rates, and reviewer counts were denied on the record before the court, including because plaintiffs had not shown a specific production deficiency warranting the requested discovery on discovery.
Did the court approve pre-culling in every matter?
No. The court declined to prohibit LinkedIn’s search-string pre-culling on the record before it, while still requiring further meet-and-confer regarding the search strings and possible adjustments.
Does more preservation automatically improve defensibility?
No. Overcollection can increase cost, privilege review, privacy exposure, and noise; under-preservation creates different risks. Preservation remains matter-specific and proportional.
What should a Managing Partner ask for internally?
A proportionate source map, decision trail, meaningful validation evidence, and a clear privilege/work-product boundary are more useful than a generic statement that a tool was approved or human review occurred.
Should every AI interaction be logged?
No. Indiscriminate logging can create unnecessary sensitive records. The better approach is to preserve operational evidence proportionately, focusing on consequential decisions and keeping legal strategy within counsel’s judgment.
What is the core Litigation AI GRC lesson?
The durable governance unit is the decision chain around the technology: what entered review, what changed, how consequential risks were tested, how preservation choices were made, and whether those decisions can be explained under pressure.
Conclusion: Explainability Under Pressure
The next significant AI-discovery dispute may not turn on whether GenAI was used. It may turn on whether the record reveals a reasonable, proportionate, and explainable process. The practical management shift is to build the source map, decision trail, validation evidence, and preservation rationale while the choices are being made—not after opposing counsel, a client, an insurer, or the court asks how the process worked.
We appreciate the completion of the Assessment Form at:
Author: Pouya Shafabakhsh Co-Founder, CAIO & Principal Forensic AI Auditor, Radsam Academy of AI Sovereign Governance. The Architect of North America's: Judicial Forensic AI Audit Standards, AI Governance, Risks & Compliance Standards, Air-Gapped Sovereign Sanctuary AI Audit System.
Comments