Orange Business, the B2B arm of Orange Group, has deployed its Live Intelligence AI platform across 120,000 employees, and more than 150 customers in France and Europe have adopted it. The company built Live Intelligence Studio with LangChain and LangSmith after estimating that developing the orchestration, deployment, monitoring and observability infrastructure for enterprise agents from scratch would take over a year. Orange Innovation evaluated roughly 20 cloud providers and specialized AI vendors. It chose LangChain for model neutrality and portability, which let Orange retain control of its agent architecture and manage workloads for European data sovereignty requirements; it selected LangSmith for evaluations and observability. Mathieu Ducrot, Orange Business’s head of AI products, said building comparable observability internally would likely have delayed the roadmap by 12 to 18 months.

Live Intelligence has three components. Studio lets AI engineers and agent developers design workflows across enterprise systems, using a catalog of language models hosted on Orange infrastructure and public clouds, an agent builder powered by LangSmith Fleet, MCP support, and LangSmith observability and orchestration tools. Portal gives business users a governed interface for tasks such as HR teams drafting job descriptions, legal teams reviewing NDA clauses, software teams generating test scripts and helpdesks building IT support agents. Workflows created in Studio can be deployed to Portal. Console gives administrators tools to monitor usage and performance and enforce security and compliance policies.

Two registries support organization-wide governance. The Agent Registry lets teams discover, reuse, version and govern agents and prompts. Orange uses LangSmith Context Hub to manage prompt versions for different environments and test changes before production without changing code. The MCP Registry is a centralized catalog of trusted tools, connectors and external capabilities. Agents reach systems such as Microsoft 365, CRMs, internal APIs and custom services through this controlled, auditable layer, moving business logic out of prompts and reducing the need for each team to rebuild authentication, security and governance.

Orange runs three production evaluation layers with LangSmith: retrieval-augmented generation (RAG) quality checks use question-answer datasets, expected source documents, retrieval recall, grounding quality and answer faithfulness; agent behavior checks cover tool selection, task completion, refusals, prompt-injection robustness and business constraints; and production trace monitoring tracks latency, cost, model usage, failed runs, human feedback and escalations. Engineers can inspect execution steps to locate whether a problem came from retrieval, a prompt, model, tool call or business logic. Teams share traces and evaluation results, compare agent versions, replay historical traces and run regression evaluations before production deployments. Orange also builds datasets from examples of good and bad agent behavior for offline evaluations and ongoing prompt refinement.

The platform has improved operational efficiency by giving engineers and subject-matter experts a shared workflow for developing and improving agents. Ducrot said technical teams own applications and agent workflows while business teams participate in prompt improvement and quality evaluation. Orange plans to expand collaboration into online evaluations and is working on more complex processes, including agents supporting the full insurance claims process with CRM and legal integrations, and an agent monitoring port security sensors to support emergency intervention by field engineers.

Orange Business, the B2B arm of Orange Group, has deployed its Live Intelligence AI platform across 120,000 employees, and more than 150 customers in France and Europe have adopted it. The company built Live Intelligence Studio with LangChain and LangSmith after estimating that developing the orchestration, deployment, monitoring and observability infrastructure for enterprise agents from scratch would take over a year. Orange Innovation evaluated roughly 20 cloud providers and specialized AI vendors. It chose LangChain for model neutrality and portability, which let Orange retain control of its agent architecture and manage workloads for European data sovereignty requirements; it selected LangSmith for evaluations and observability. Mathieu Ducrot, Orange Business’s head of AI products, said building comparable observability internally would likely have delayed the roadmap by 12 to 18 months.

Live Intelligence has three components. Studio lets AI engineers and agent developers design workflows across enterprise systems, using a catalog of language models hosted on Orange infrastructure and public clouds, an agent builder powered by LangSmith Fleet, MCP support, and LangSmith observability and orchestration tools. Portal gives business users a governed interface for tasks such as HR teams drafting job descriptions, legal teams reviewing NDA clauses, software teams generating test scripts and helpdesks building IT support agents. Workflows created in Studio can be deployed to Portal. Console gives administrators tools to monitor usage and performance and enforce security and compliance policies.

Two registries support organization-wide governance. The Agent Registry lets teams discover, reuse, version and govern agents and prompts. Orange uses LangSmith Context Hub to manage prompt versions for different environments and test changes before production without changing code. The MCP Registry is a centralized catalog of trusted tools, connectors and external capabilities. Agents reach systems such as Microsoft 365, CRMs, internal APIs and custom services through this controlled, auditable layer, moving business logic out of prompts and reducing the need for each team to rebuild authentication, security and governance.

Orange runs three production evaluation layers with LangSmith: retrieval-augmented generation (RAG) quality checks use question-answer datasets, expected source documents, retrieval recall, grounding quality and answer faithfulness; agent behavior checks cover tool selection, task completion, refusals, prompt-injection robustness and business constraints; and production trace monitoring tracks latency, cost, model usage, failed runs, human feedback and escalations. Engineers can inspect execution steps to locate whether a problem came from retrieval, a prompt, model, tool call or business logic. Teams share traces and evaluation results, compare agent versions, replay historical traces and run regression evaluations before production deployments. Orange also builds datasets from examples of good and bad agent behavior for offline evaluations and ongoing prompt refinement.

The platform has improved operational efficiency by giving engineers and subject-matter experts a shared workflow for developing and improving agents. Ducrot said technical teams own applications and agent workflows while business teams participate in prompt improvement and quality evaluation. Orange plans to expand collaboration into online evaluations and is working on more complex processes, including agents supporting the full insurance claims process with CRM and legal integrations, and an agent monitoring port security sensors to support emergency intervention by field engineers.