Stepping Up to the Plate: Human-in-the-Loop and What Courts and Regulators Expect

In our previous post in this series, we discussed applying protective orders to limit the use of public LLMs on opponents’ sensitive and confidential documents, including three recent case law rulings that addressed disputes related to this issue, and two examples of protective order language on which to model your own request.

AI is rapidly becoming embedded throughout the eDiscovery lifecycle. Generative AI can help legal teams assess collections during early case assessment (ECA), identify responsive documents, detect potentially privileged communications, summarize evidence, construct timelines, analyze issues, and perform quality control. Those capabilities can dramatically reduce the time and effort required to manage large volumes of electronically stored information (ESI).

But greater automation does not eliminate the need for attorney oversight. For courts and regulators, the fundamental principle remains largely unchanged: legal teams are responsible for the work performed on behalf of their clients, regardless of the technology used to perform it. In this post, we’ll discuss why having humans in the loop remains central to defensible AI usage and best practices for ensuring a defensible approach to leveraging AI in discovery.

AI Changes the Workflow, Not the Responsibility

Court acceptance of technology-assisted review (TAR) by Judge Andrew Peck in the case Da Silva Moore v. Publicis Groupe & MSL Group established years ago that attorneys do not necessarily have to manually review every document to conduct defensible discovery. GenAI extends that concept further.

Instead of predicting whether documents resemble attorney-coded examples, GenAI systems can apply natural-language instructions to documents and explain why a document may be responsive, privileged, or relevant to a particular issue. Attorneys can ask questions of large document populations, generate summaries, identify important evidence, and potentially automate significant portions of first-pass review.

But delegating work to AI is not the same as delegating responsibility.

Humans still need to be in the loop to prompt or instruct AI, determine whether its output is sufficiently reliable for the intended purpose, and establish appropriate validation and quality-control procedures. The degree of human oversight should depend heavily on the consequences of getting the answer wrong.

AI-Driven Review Is An Iterative Process

One of the biggest mistakes eDiscovery professionals can make is treating attorney review as one of the last steps in an AI-driven process. Meaningful oversight should occur throughout the workflow.

Consider a GenAI document review involving hundreds of thousands of documents. Attorneys might provide the system with natural-language prompts describing responsiveness issues and allow the technology to classify the collection. Waiting until the classifications are complete before assessing their quality risks an incorrect interpretation of the review criteria affecting the entire population.

Instead, it’s important to establish checkpoints at critical stages. Attorneys can draft and periodically review the initial prompts provided to the AI, evaluate early samples of classified documents, examine borderline results, refine prompts where necessary, and conduct additional validation until the prompts are refined.

This iterative process on a small subset of documents is human driven – it’s literally humans in-the-loop and is vital to ensuring the entire collection is accurately classified. Reviewing and correcting AI classifications on a few sample documents can prevent wholesale misclassifications and ensure greater overall accuracy.

Privilege Requires Additional Caution

Privilege identification illustrates why risk-based oversight matters.

GenAI may be particularly useful for identifying privileged documents because it can evaluate context rather than relying exclusively on attorney names, domains, keywords, or other traditional indicators. GenAI may recognize that a communication involves a request for legal advice even when no obvious privilege terms appear or, vice-versa, understand that communications involving attorneys do not rise to the level of legal advice.

Still, errors in privilege classification can have significant consequences.

A defensible workflow might therefore use AI to identify and categorize potentially privileged material while requiring attorney review of higher-risk categories before production. Legal teams should also sample AI identified non-privileged documents to test whether potentially protected communications were missed.

That second step is particularly important. Reviewing only the documents the system identified as privileged tells the team little about false negatives lurking elsewhere in the collection.

Bottom line: The stakes are higher when using AI for privilege determination, so human oversight is critical to ensure defensible classifications.

ECA Allows More Flexible Oversight

Conversely, the appropriate level of validation can look considerably different during ECA.

At that stage, attorneys may use GenAI to identify important custodians, summarize communications, locate documents discussing particular issues, construct preliminary timelines, or evaluate the strengths and weaknesses of claims and defenses.

Because those results generally inform attorney decision-making rather than directly determine what will be produced or withheld, a greater degree of uncertainty is acceptable. This is the stage for developing insights, where legal professionals can use AI to learn about the document collection rather than dealing with the concerns of document production. Attorney validation of AI results should be proportional to risk.

Turning Oversight into Validation

Sampling is one of the most important tools for meaningful AI oversight because it enables legal teams to evaluate large-scale automated decisions without manually repeating the work.

Depending on the workflow, validation procedures might include:

• Random sampling: Selecting documents randomly from AI-classified populations to estimate overall performance.
• Targeted sampling: Reviewing higher-risk categories, unusual classifications, or documents where the system’s conclusions appear uncertain.
• Null-set sampling: Sampling documents classified as nonresponsive to determine whether responsive material has been incorrectly excluded.
• Privilege QC: Sampling both potentially privileged documents and documents classified as nonprivileged to evaluate false positives and false negatives.
• Issue-specific testing: Evaluating performance separately for different claims, custodians, document types, languages, or other characteristics where performance may vary.

Metrics such as recall, precision, and elusion can provide additional insight into system performance. But metrics alone don’t establish defensibility. Attorneys must determine what those results mean in the context of the matter and whether additional review, sampling, or workflow adjustments are warranted.

The Importance of Escalation Procedures

Effective AI workflows should also define when a decision requires more human oversight.

An escalation procedure might require human intervention when the AI encounters ambiguous responsiveness criteria, potentially privileged communications involving unusual participants, conflicting classifications, highly confidential information, novel document types, or material involving particularly important issues.

Agentic AI makes these procedures even more significant. As systems become capable of performing sequences of discovery tasks with less human intervention, legal teams need clearly defined boundaries identifying what the system can do autonomously, and which decisions require human approval.

The goal is not human involvement in every action. It is humans in the loop where they’re most needed: to make consequential decisions.

Production QC Is the Last Line of Defense

No matter how sophisticated the underlying AI workflow becomes, production QC remains critical.

Before ESI leaves the legal team’s control, attorneys and eDiscovery professionals should apply checks to verify that the production reflects the intended scope, privileged material has been appropriately addressed, agreed-upon metadata has been provided, redactions have been correctly applied, production specifications have been followed, and unexpected documents or data have not slipped through. Just as they do with any other production.

Document Your Work

Defensibility ultimately involves more than achieving a good result. Legal teams should be able to explain how they reached it.

That means documenting important elements of an AI-assisted eDiscovery workflow, including the technology and model used, instructions or criteria applied, attorney review checkpoints, sampling methodology, validation results, identified errors, workflow adjustments, escalation decisions, and final QC procedures.

While many courts try to avoid “discovery on discovery” unless significant production issues are identified, it’s better to have the documentation that you may not need than to not have it when you need it. Also, regulatory agencies are often more likely to request documentation of the process, so it’s best to establish a consistent approach to documenting your AI workflows.

Conclusion

Let’s face it: manually reviewing everything AI touches would eliminate the ROI for using the technology in the first place. AI can dramatically change who – or what – performs the work in eDiscovery. And it can considerably streamline eDiscovery workflows.

But when courts and regulators ask whether the resulting process was reasonable and defensible, the answer will still depend on whether knowledgeable legal professionals exercised meaningful judgment over that process. AI may continue to do more of the work. But the legal team still owns the outcome. The responsible attorney is the one that certifies the production – regardless of the workflows used to get there. This post builds on earlier installments in our Stepping Up to the Plate series covering the fundamentals of AI, lawyer competence, hallucinations and verification, confidentiality risks, and protective orders. Catch up on the series, and stay tuned as we continue exploring the practical and ethical considerations shaping responsible AI use in legal practice.

Next time, we’ll discuss the importance of practical checklists for internal AI policies.

For more regarding Cimplifi specialized expertise regarding AI & machine learning, click here.