AivexaNewsSearch
AI news for builders and product teamsChecked every hour

How Orange Business Built Live Intelligence with LangChain and LangSmith

Collected Oct 13, 2026

Orange Business, the B2B arm of Orange Group, built its flagship AI platform Live Intelligence on LangChain and LangSmith. The platform is deployed across 120,000 Orange employees and has been adopted by more than 150 customers in France and Europe, according to LangChain's account of the work.

The starting point was the Live Intelligence Portal, which let employees create simple AI agents and trigger basic API calls. That covered entry-level personal tasks, but Orange Business wanted agents that could orchestrate complete business processes across fragmented enterprise systems, govern who can access what, and run reliably at production scale. Reproducing the orchestration, deployment, monitoring and observability infrastructure that enterprise-grade agents require would have taken more than a year of development, so the team looked for a foundation instead of building from scratch. That decision produced Live Intelligence Studio.

Orange Innovation, the group's innovation scouting arm, evaluated roughly 20 major cloud providers and specialized AI vendors. LangChain was selected for model neutrality and portability, which let Orange Business keep full ownership of the agent architecture and manage workloads in line with European data sovereignty requirements. LangSmith was chosen for evaluations and observability. Mathieu Ducrot, Head of AI Products at Orange Business, said the observability capabilities in LangSmith were significantly ahead of what the team could realistically build internally, and that building it themselves would likely have delayed the roadmap by twelve to eighteen months.

Live Intelligence has three components aimed at different users. Studio is for AI engineers and agent developers designing complex cross-system workflows. It includes an LLM-as-a-Service catalog with models hosted on Orange Business infrastructure and public cloud, an agent builder powered by LangSmith Fleet, MCP support, agent observability through LangSmith, and orchestration tooling. Portal is the interface for business end users; the examples given are HR teams drafting job descriptions, legal teams reviewing NDA clauses, software teams auto-generating test scripts, and helpdesks building IT support agents. Workflows built in Studio can be deployed directly into Portal. Console gives enterprise administrators usage monitoring, performance tracking, and enforcement of security and compliance policies.

Two registries add organizational governance on top of individual agents. The Agent Registry lets teams discover, reuse, version and govern agents and prompts across the organization, so agents become shared assets with lifecycle management instead of each team rebuilding the same capability. For prompt management, Orange uses LangSmith Context Hub to maintain different prompt versions per environment, which allows prompt changes to be tested and validated before reaching production without changing code. The MCP Registry is a centralized catalog of trusted tools, connectors and external capabilities. Agents reach enterprise systems such as Microsoft 365, CRMs, internal APIs and custom services through a controlled, auditable layer. The effect, as described, is to pull business logic into the platform layer and out of prompts so teams do not re-solve authentication, security and governance individually.

Evaluation runs as a three-layer stack built on LangSmith. The RAG layer uses golden datasets with question-answer pairs and expected source documents, checking retrieval recall, grounding quality and answer faithfulness. The agent-behavior layer measures tool selection accuracy, task completion rates, refusal behavior, robustness against prompt injection, and adherence to business constraints. The production layer monitors traces for latency, cost, model usage, failed runs, human feedback and escalation cases. Ducrot said that debugging an agent used to be difficult because failures were hidden inside model behavior, and that LangSmith lets engineers inspect each execution step and determine whether a problem comes from retrieval, the prompt, the model, a tool call or the business logic. That shared visibility means AI engineers, backend engineers, product managers and QA can work from the same traces and evaluation results rather than separate views.

Regression evaluations are enforced before changes reach production. Engineers can compare agent versions, replay historical traces and validate behavior before deployment. For workflows with multiple agents and tool calls, the full execution tree shows where a workflow failed, which agent or tool was executing, the input to that step, the response received, how long individual steps took, how many agents and tools ran in the workflow, and how each step contributed to the overall execution. Orange also builds datasets from trace examples representing good and bad agent behavior for offline evaluations and continuous prompt refinement.

The reported result is improved operational efficiency across teams, with engineers and subject-matter experts sharing one workflow for agent development and improvement. Ducrot described the split as the technical team owning application and agentic workflow implementation while the business team participates in prompt improvement and quality evaluation, with LangSmith turning prompt iteration, tracing and evaluation into a continuous improvement loop. Orange plans to extend that loop to online evaluations for richer feedback. The team is also working on automating more complex processes, including agents supporting the entire insurance claim process with CRM and legal workflow integrations, and an agent that monitors security sensors in a port for field-engineer emergency intervention.

This account is a case study published by a vendor, so the deployment figures, timelines and efficiency claims come from Orange Business and LangChain rather than independent measurement. In practice, the architecture is worth reading as a reference shape for enterprise agent platforms: model-neutral orchestration, a shared agent and tool registry, and evaluation split across retrieval, agent behavior and live production traces. The registries are the part most teams underestimate. Centralizing connectors and prompts is what makes governance auditable, but it also creates a platform team that owns the shared layer while product teams depend on its release cadence. A likely trade-off is slower initial onboarding for teams accustomed to self-serve tooling, offset by not rebuilding authentication, security and evaluation for every new agent.

Why it matters: Developers and product teams building regulated or multi-team agent platforms get a concrete pattern for separating orchestration, governance and evaluation, plus evidence that observability tooling can substitute for a year-plus of internal platform work. Teams should note that the reported gains depend on committing to shared registries and pre-deployment regression evaluations, which is an organizational change as much as a technical one.

Read at LangChain

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Learn how Orange Business used LangChain and LangSmith to build Live Intelligence, a governed AI platform for agents, evals, observability, and production scale.