TechPrufer
Get started
Harvey
Harvey logoHarvey · Feature trail

AML AI Detection Systems

25 posts · Sep 2024 – Sep 2026

What this trail covers

Harvey's platform continues to evolve its AI agent capabilities, shifting the paradigm from prompt-based assistance to task-oriented execution. The system now empowers agents to autonomously plan, research, execute, and deliver finished work products, with lawyers focusing on goal setting and final review. This evolution spans transactional work, litigation, compliance, and in-house operations, aiming to increase leverage by automating repeatable, high-volume tasks. The platform now integrates custom models and LLM systems, such as the Tax AI Assistant developed in alliance with PwC and OpenAI, which combines curated tax datasets with Harvey's LLM expertise and is fine-tuned with feedback from PwC's tax experts. This system answers complex tax queries with accurate citations to source material, powered by in-house models and semantic search technology, and further specialized on tax datasets and expert-annotated examples. Human-guided model training is used to expand specialized models to more tax jurisdictions.

All Harvey trails
  1. Post-Training RLM Agents for End-to-End M&A DiligenceThis post details the technical process of applying post-training techniques, specifically reinforcement learning and self-distillation SFT, within a recursive language model (RLM) harness for M&A diligence. It introduces the LAB Diligence environment with large document contexts and evaluates the performance gains achieved by post-training a Qwen3.5-122B-A10B orchestrator, increasing rubric criteria pass rate from 29.9% to 63.0%. The post also analyzes the division of labor between root and sub-agents in the RLM harness and discusses coverage and cost implications.
  2. How to Run a Compliance-Ready Policy Gap AnalysisThis post introduces the capability to perform compliance-ready policy gap analyses. It details the six steps involved: defining the benchmark and obligations, assembling the policy set, comparing policies with obligations (leveraging AI for this comparison and citing authoritative sources), checking for evidence of obligations in practice (identifying operational gaps), prioritizing gaps by risk, and recording findings with an audit trail. It highlights the use of AI agents for repeatable checks and continuous analysis, contrasting this with traditional sampled reviews. The post also mentions the integration of curated content from data partners and the use of Harvey Vault for bulk analysis.
  3. Harvey Tenet Research PreviewThis post introduces Harvey Tenet, a new post-trained open-weight model based on Kimi K3, developed in collaboration with Fireworks. It details the model's training methodology using asynchronous reinforcement learning (RL) and group-sequence policy optimization (GSPO), focusing on improving performance and cost-efficiency for long-horizon, agentic legal tasks. The post highlights initial results on legal benchmarks (LAB, APEX Agents, Redline Bench) and explores specialized capabilities in M&A Diligence (using Recursive Language Models - RLM), Review Tables (structured data extraction), and Firm Knowledge.
  4. Introducing Harvey IIThis post introduces Harvey II, which significantly enhances the AML AI Detection Systems by integrating matter/project context inheritance and persistent user memory into agent workflows. Agents now operate within a 'Space' that encapsulates all relevant context (documents, permissions, history) for a given matter, eliminating the need for repeated setup. The 'Memory' feature allows agents to learn and apply user-specific preferences and team templates, personalizing the AI's output. This is underpinned by the new Harvey Tenet model, a frontier-level, legal-specific model designed for cost-efficient, long-horizon legal reasoning. The post details how these components combine to enable agents to handle more substantive legal work by starting with pre-existing context and learned preferences.
  5. AI for Regulatory Compliance: Monitor, Interpret, ActThis post details the application of AI to regulatory change management, focusing on the 'interpret' and 'act' stages after initial monitoring. It describes how AI can assess applicability of regulations based on entity, activity, and jurisdiction, and how it can translate interpretations into updated obligations, policies, and controls. The post also emphasizes the importance of human oversight and grounding AI outputs in source materials for auditability and trust. It positions Harvey as a tool that complements existing monitoring platforms by assisting with interpretation and action, rather than replacing them.
  7. Corporate Compliance AI: Keeping Up With Regulatory ChangeThis post details the application of AI to corporate compliance, specifically focusing on how in-house teams can use AI to interpret regulatory change, map obligations, and update policies. It introduces the concept of 'Corporate Compliance AI' as distinct from AI governance, highlighting its role in accelerating the interpretation and implementation phases of regulatory change management. The post outlines an AI-assisted workflow that includes synthesizing regulatory updates, producing plain-language summaries, mapping discrete obligations, identifying policy gaps by comparing new requirements against existing documents, and drafting policy revisions. It also emphasizes the importance of an audit-ready record and the division of responsibilities between AI and legal professionals, along with the security considerations for using AI in compliance work. The post mentions the integration of agentic workflows for this purpose.
  8. AML AI Software: Detection Systems Plus the Legal LayerThis post elaborates on the two-layer architecture of AML AI software, distinguishing between detection systems and the legal judgment layer. It details how legal AI supports AML compliance by interpreting rules across jurisdictions, updating policies, reviewing KYC/CDD/EDD files, and documenting investigations. The post emphasizes that legal AI complements, rather than replaces, AML detection systems and highlights the continued responsibility of human compliance professionals for judgment and accountability. It also touches upon enterprise-grade security requirements for legal AI platforms, including data handling and auditable records.
  9. AI Tax Research: Get Reliable Answers to Complex QuestionsThis post details the application of AI agents to tax research, specifically focusing on answering complex, multi-jurisdictional questions, moving from research to drafted memos, and keeping up with changing tax rules. It highlights the use of purpose-built AI tax research software that grounds answers in primary authority, reflects current law, includes verifiable citations, and has been reviewed by a tax professional. The post emphasizes the security aspects of handling client tax data and outlines how Harvey approaches tax research by applying its legal AI platform to tax professionals' needs, including co-building custom tax models with PwC. It also provides best practices for using AI in tax research.
  10. AI in Legal Research vs. Traditional Tools: What ChangedThis post details the evolution of Harvey's legal research capabilities, introducing AI-native research that moves from keyword-based searches to natural language questions. It explains how the system synthesizes answers with supporting authority, grounds claims in authoritative legal databases, and emphasizes source transparency to combat hallucination. The post highlights the integration with existing legal research databases and the importance of lawyer verification. It also discusses how to evaluate AI legal research platforms by asking about sources, traceability, handling of negative treatment, performance measurement, and data handling, referencing Harvey's published benchmarks on BigLaw Bench and Legal Agent Bench.
  11. AI Trademark Search: From Clearance to Brand ProtectionThis post introduces the application of AI to trademark search and brand protection. It details how AI models are used to compare marks based on sound, meaning, and visual similarity across various registries, moving beyond simple keyword matching. The post explains how AI assists legal professionals in analyzing search results, drafting clearance opinions by evaluating likelihood of confusion factors, responding to office actions, and supporting trademark watch and enforcement efforts. While the core AML AI Detection Systems remain, this post describes a new capability within that broader system focused on the trademark lifecycle.
  12. Extending Legal Agent Bench to M&A Due Diligence | HarveyThis post extends the Legal Agent Bench (LAB) to evaluate AI agents on M&A due diligence workflows. It introduces synthetic VDRs with tens of millions of tokens of context and thousands of documents to test agents' ability to identify and remediate transaction risks. The post details the challenges of diligence, including context management, exhaustive review, and contextualized judgment, and outlines the development of specialized agents to address these issues.
  13. Training a Legal Agent With Applied Compute | HarveyThis post details the process of post-training a frontier legal model (GLM-5.1) on Harvey's Legal Agent Benchmark (LAB) in collaboration with Applied Compute. It covers grader alignment using GPT-5 Mini for cost-efficient evaluation, optimization of the legal-agent harness with features like token restriction and compaction, and full-parameter reinforcement learning on Applied Compute Agent Cloud. The analysis highlights domain-level shifts, tool-use optimization (reducing read calls and payload tokens), and behavioral improvements in artifact completeness, specificity, and grounding, resulting in a model that outperforms previous benchmarks.
  14. Claude Opus 4.8, Now Live in HarveyThis post announces the integration of Claude Opus 4.8 into Harvey's AML AI Detection Systems. It details the model's performance improvements on the Legal Agent Benchmark (LAB) and BigLaw Bench, highlighting specific gains in legal accuracy, drafting, and risk assessment. It also includes qualitative feedback from evaluators regarding the model's strengths and areas for future development, such as handling ambiguity.
  15. Post-Training Open Legal Agents With Baseten ResearchThis post details initial research and experiments towards open-weight legal agents, specifically post-training a 27B open-weight model using Baseten Research's pipeline. This pipeline leverages the Legal Agent Benchmark (LAB) signal and a legal-agent harness designed for long-horizon legal work. The research demonstrates that post-training can address challenges in domain expertise, cost, and governance for specialized legal agents. Experiments show that training an agent end-to-end against LAB criteria can shift its behavior towards more comprehensive document reading, similar to frontier models. A natural-language compaction strategy was implemented and benchmarked, showing significant improvements in the all-pass rate for deliverable quality. The post-training of Qwen3.5-27B using Iterative SFT (iSFT) in this compaction harness brought its performance into the closed-source frontier band on LAB, highlighting the need for co-development of post-training and harness optimization.
  16. Legal Agent Benchmark Initial Results | HarveyThis post introduces the Legal Agent Benchmark (LAB) and presents initial results from baselining frontier models. It details the methodology for evaluating agents on complex legal tasks, including the all-pass standard and practice area breakdowns. Key findings include the current limitations of frontier models in completing legal work end-to-end, the 'jagged intelligence' observed across different legal specializations, and the substantial cost and latency associated with high-performing models. The post also provides a qualitative analysis of agent behaviors, categorizing actions like 'Read,' 'Search,' 'Execute,' 'Write,' 'Validate,' and 'Edit,' and correlating specific behavioral patterns (e.g., thorough research before drafting, self-correction) with improved performance. This contributes to understanding the trade-offs in model selection and agent design for legal applications.
  17. Claude Opus 4.7, Now Live in HarveyThis post announces the availability of Claude Opus 4.7 in Harvey, highlighting its improved reasoning calibration, stronger performance on ambiguous tasks, and gains in substantive accuracy. It provides specific metrics from the BigLaw Bench evaluation (90.9% score, 45% perfect scores, 88% >= 0.80) and qualitative feedback on its strengths and minor polish gaps compared to Opus 4.6.
  18. Opus 4.6, Now Live in HarveyThis post announces the availability of Claude Opus 4.6 in Harvey, highlighting its improved substantive accuracy, source grounding, and multi-step legal task handling. It details Opus 4.6's performance on the BigLaw Bench (90.2% score, 40% perfect tasks) and notes observations on its response calibration under uncertainty, specifically the tendency to over-contextualize when less confident. The rollout plan to US and EU clients is also mentioned.
  19. Harvey Co-Builds Custom Model for Tax with PwCThis post details the co-development of the Tax AI Assistant with PwC, a custom model and LLM system for tax professionals. It highlights the process of combining curated tax datasets (legislation, case law, revenue guidance, PwC tax materials) with Harvey's LLM expertise and fine-tuning with feedback from PwC tax experts across 10+ jurisdictions. The system uses in-house models and semantic search, further tuned on tax datasets and expert-annotated examples. A key aspect is the human-guided approach, incorporating expert responses and guidance to optimize the dataset and model, followed by a feedback loop with engineers and tax experts to improve output quality. The post also presents results from a blind side-by-side human preference evaluation where the Tax AI Assistant was preferred 91% of the time by experts.