How LangSmith and LangChain OSS Help You Meet EU AI Act Requirements

LangChain published a guide describing how its LangSmith and LangChain OSS products can help developers and deployers of high-risk AI systems meet requirements under the EU AI Act, whose compliance deadline is August 2, 2026.

According to the company, the Act targets high-risk AI systems used in areas including credit scoring, medical devices, recruitment, biometric identification, critical infrastructure, and law enforcement. Non-compliance with the high-risk provisions carries penalties of up to €15 million or 3% of total worldwide annual turnover, whichever is higher. The requirements include risk management systems, automatic event logging, transparency to deployers, human oversight mechanisms, post-market monitoring, and incident reporting.

LangChain maps its products to specific articles. For observability and tracing, it cites Article 9's requirement for a living risk management system across the development lifecycle, Article 12's requirement for automatic event logging over the system's lifetime, and Article 13's requirement for traceable, interpretable decisions. LangSmith provides end-to-end tracing of LLM calls, tool invocations, and reasoning steps with structured metadata, an execution graph view, automatic pattern clustering, and custom dashboards that trigger alerts through PagerDuty or webhooks.

On retention, LangChain states that managed cloud retains base traces for 14 days and extended traces for 400 days, with bulk export for long-term archival. It also offers EU data residency through LangSmith EU, plus self-hosted and BYOC deployments where the stack runs in the customer's Kubernetes cluster or cloud region.

For evaluations, LangChain cites Article 10 on data governance and bias examination, Article 13 on transparency for deployers, and Article 15 on declared accuracy levels, adversarial resilience, and protection against common attack surfaces. LangSmith's online evaluators score a configurable sample of production traces, with prebuilt evaluators covering bias and fairness, toxicity, sensitive imagery, hallucination, answer relevance, PII leakage, prompt injection and jailbreaking, API leakage and code injection, correctness, plan adherence, and tool selection.

On human oversight, LangChain cites Article 14's requirement that humans can understand, intervene on, override, and interrupt the system. It points to LangGraph's interrupt primitive for human-in-the-loop control within the agent graph, LangSmith Deployment for checkpointing and resume-from-exact-point recovery, annotation queues for routing traces to reviewers, and webhooks for alerting.

The post includes a crosswalk table and links to get started with LangSmith, speak with an expert, and explore a trust center.