Autonomous large language model (LLM) agents are increasingly deployed in electronic discovery (e-discovery), but compounding errors across multi-step reasoning chains can constitute legal malpractice. According to a paper published on arXiv, agentic workflows operating over privileged document corpora exhibit a class of failure termed "trajectory collapse": an early misclassification silently propagates, rendering an entire privilege review invalid.
The Failure: Trajectory Collapse in Autonomous E-Discovery
The paper makes three contributions. First, it proposes a structured taxonomy of agentic failures in legal information retrieval, organized by functional stage. Second, it introduces a four-layer verification architecture spanning planning, reasoning, execution, and uncertainty quantification, designed to intercept these failures before they compound. Third, it presents a preliminary simulation study on a synthetic e-discovery corpus that demonstrates how mandatory Human-on-the-Loop (HOTL) escalation thresholds reduce privilege-waiver risk relative to fully autonomous baselines.
Architecture: Four-Layer Verification with Human Escalation
The proposed architecture integrates human oversight at critical decision points. Calibrated uncertainty thresholds trigger escalation to attorney review only when the model's confidence is low. This approach contrasts with fully autonomous systems that may proceed blindly through a chain of reasoning steps.
Simulation Results: 61% Risk Reduction
In the simulation study, the researchers found that calibrated uncertainty thresholds can reduce privilege-waiver risk by up to 61% versus fully autonomous deployment, while routing fewer than one quarter of documents to attorney review. This suggests that a targeted human-on-the-loop approach can dramatically lower legal risk without overwhelming legal teams with false positives.
Implications for Legal Technology Procurement
For CTOs and technology leaders in legal departments, the findings underscore the importance of incorporating human oversight mechanisms when deploying LLM agents for e-discovery. Rather than relying on end-to-end automation, a human-on-the-loop orchestration layer can catch early misclassifications before they cascade. The paper's taxonomy of failures may also serve as a framework for evaluating vendor solutions. As the authors note, compounding errors across multi-step reasoning chains can constitute legal malpractice, making risk mitigation a procurement priority.
| Metric | Fully Autonomous | Human-on-the-Loop (HOTL) |
|---|---|---|
| Privilege-waiver risk | Baseline | Up to 61% reduction |
| Documents routed to attorney review | 100% (implicit) | Fewer than 25% |
The four-layer verification architecture provides a blueprint for building more robust AI-assisted discovery tools. While the simulation was conducted on a synthetic corpus, the results point to a viable path for safer deployment of LLMs in high-stakes legal contexts. Legal technology buyers should look for solutions that implement explicit uncertainty quantification and mandatory escalation protocols.
