Key Takeaways
- Harvey II and Litera’s Lito agents move legal AI from single-prompt tools to multi-step process automation, including retaining attorney preferences.
- Agentic AI outputs require full traceability for legal defensibility, with systems like Visus LLC’s capturing detailed audit trails.
- Agentic systems accelerate data-heavy tasks such as eDiscovery triage, a design pattern also applied in security operations by Cisco. Harvey and Litera both shipped agentic legal platforms in August 2026, and the gap between them and traditional legal software is wider than a feature update. Harvey II can retain attorney preferences across sessions; Litera’s Lito now handles contract drafting and negotiation inside Microsoft Word without leaving the document. Both moves point to the same bet: that lawyers will hand off multi-step workflows to AI agents before they hand off judgment.
Legal AI Moves to Goal-Directed Work
Harvey‘s Harvey II platform includes a “Memory” feature that retains individual attorney preferences across agents and applications, so the system builds context over time rather than starting cold on every task. Litera updated its Lito AI agent to handle contract drafting and negotiation as a single end-to-end workflow inside Microsoft Word and Litera for Windows. These are not incremental improvements to existing tools; they are attempts to own the full process, not just the prompt.
The shift matters most in eDiscovery, where the old model was AI-assisted document review and the new model is AI-managed review pipelines. LegalOn Technologies added five new AI agents in February 2026 covering playbook creation, intake responses and contract translation, framing the output as “review-ready legal work under attorney control.” That framing is deliberate. The agent does the legwork; the attorney owns the conclusion. For in-house teams running lean, that distinction is the whole value proposition, and it connects directly to how other firms are cutting contract review hours through legal AI.
Agents Accelerate Evidence Triage
In eDiscovery, early case assessment maps almost perfectly onto what agentic systems do well: sequential, data-heavy tasks with a defined endpoint. Alvarez & Marsal describe agents that categorize matter severity, pull data across multiple systems and produce summary reports for senior investigators before a human reviewer opens a single document. The triage phase shrinks; the volume reaching costly human review shrinks with it.
Cisco showcased its Instant Attack Verification in August 2026, built for Security Operations Centers running tier-1 and tier-2 investigations. The agent triages incoming detections, enriches alerts with context, filters false positives and prioritizes real threats. It then correlates evidence across endpoint, network, cloud and identity data, reconstructing attack timelines and recommending containment actions. Cisco frames the ambition as “100x scalability, quality, and speed” in security operations, that is the company’s own claim and has not been independently tested. The structural difference from older alert-scoring tools is that this agent queries systems actively and adapts its investigation path as new evidence comes in, rather than running a static scoring pass.
The Audit Trail Problem
Visus LLC put it plainly in an August 2026 post: “The AI said it” is not a defensible position. Their rebuilt tracing layer records the exact user question, retrieved document chunks with IDs and ranking scores, model version, system prompt, final generated answer and a cryptographic hash to detect tampering. In high-stakes configurations, the system refuses to answer if it cannot produce citations above a minimum confidence threshold, routing the question to a human with the full decision record attached.
Thomson Reuters introduced what it calls a “Fiduciary-Grade AI™” framework, requiring that every material AI output be traceable to a curated, verified legal source that a qualified professional can independently locate and cite.
Preserving Human Control Through Design
The American Bar Association stated in January 2026 that best practices for agentic AI tools require human oversight to guard against hallucinations and potential sanctions for fabricated output. Cisco’s Instant Attack Verification gates high-impact actions behind human approval and feeds analyst corrections back into the system as a learning loop. The design logic is incremental autonomy: let the agent handle data collection and correlation, keep humans on the decisions that carry consequences.
Vectra AI which focuses on network detection and response, takes a similar approach in its agentic SOC design, automating data collection, correlation and prioritization while keeping analysts in control of the actual security decisions. The pattern across both legal and security deployments is consistent: agents absorb the volume work, humans absorb the accountability. That division only holds if the handoff points are clearly defined and the audit record makes clear which decisions were made by whom, something worth weighing alongside broader questions about what happens when agents operate without that oversight.
Governing AI Autonomy
General counsel are no longer just evaluating AI tools, they are governing systems that can make decisions without continuous human sign-off. A PYMNTS.com article from August 2026 describes legal departments managing that exposure through AI addenda in vendor agreements, covering acts and omissions, legal violations and third-party liability. The addendum is becoming standard contract hygiene for any vendor relationship that involves autonomous AI.
An FTI Consulting and Relativity report put generative AI use in corporate legal departments at 87% in 2026, nearly double the prior year. Summarisation and contract review are driving most of that adoption. The governance infrastructure, policies, design artifacts, runtime enforcement logs, human approvals, access reviews, test results and incident records, has to scale with the adoption rate, not lag behind it. Auditable records are what convert an agent from a productivity tool into something a regulator or opposing counsel can examine.
Originally published at https://autonainews.com/harvey-ii-and-lito-agents-advance-legal-ai-workflows/
Top comments (0)