自社のAzure OpenAIから、ServiceNowのMCPサーバー経由でNow Assistに接続する方法（Entra IDでのユーザー引き当て対応） 自社のAzure OpenAIから、ServiceNowのMCPサーバー経由でNow Assistに接続する方法（Entra IDでのユーザー引き当て対応）

When an employee types "show me the latest five incidents" into an internal AI chatbot, the answer is often a confident but generic explanation rather than an actual list drawn from the ticketing system. The gap is structural: a large language model, on its own, has no channel to query live business data. This article walks through a pattern for closing that gap, connecting an organization's own Azure OpenAI deployment to ServiceNow through a Model Context Protocol (MCP) server, while resolving user identity through Microsoft Entra ID so that each request carries the permissions of the person who made it.

The Model Context Protocol is an open standard, introduced by Anthropic in late 2024, that defines a uniform way for AI applications to reach external tools and data. Instead of building a bespoke integration for every back-end system, a developer exposes capabilities through an MCP server, and any MCP-capable client can discover and invoke them. The approach has spread quickly, with major vendors shipping both servers that expose their platforms and clients that can consume them. ServiceNow's MCP server fits the server side of that picture, offering a standardized entry point into incident, request, and other table data managed in the Now Platform.

Now Assist is ServiceNow's generative AI layer, embedding summarization, search, and assisted authoring across workflows such as IT service management. Connecting through it, rather than calling raw REST APIs directly, means the model can lean on ServiceNow's own context and guardrails. In the configuration described here, Azure OpenAI is the reasoning engine on the client side, using its tool- or function-calling capability to decide when a question requires real data and to issue the corresponding MCP call.

The identity piece is what makes the setup viable for enterprise use. Microsoft Entra ID, formerly Azure Active Directory, authenticates the user and maps them to a corresponding ServiceNow identity, so the assistant does not run with a single shared service account that could over-expose data. With user mapping in place, a request for "my open incidents" returns only the records that user is entitled to see, and audit trails remain meaningful. This on-behalf-of pattern is central to deploying AI assistants that touch regulated or access-controlled systems.

Several prerequisites sit behind the scenes. An Azure OpenAI resource must be provisioned in the organization's own tenant, which gives teams more control over data residency and network boundaries than consumer AI services typically allow. The ServiceNow instance needs the MCP server enabled and the relevant Now Assist capabilities licensed. The Entra ID configuration involves registering the application, defining the scopes it can request, and establishing trust between the Microsoft identity platform and ServiceNow, usually through OAuth.

The broader context is a shift from chatbots that merely generate text to agents that take grounded actions against systems of record. Retrieval-augmented generation addressed part of the problem by injecting documents into prompts, but it does not let a model perform live, permission-aware queries or write changes back. MCP, alongside vendor function-calling features, is one of the mechanisms now competing to become the connective tissue between models and enterprise software. Comparable efforts include direct ServiceNow integrations from other AI platforms and a growing catalog of MCP servers for tools such as GitHub, databases, and collaboration suites.

Organizations weighing this approach should treat it as an architecture pattern rather than a turnkey product. Details such as token handling, rate limits, and how tool results are formatted back into the model's context can materially affect reliability, and the specifics are likely to change as both the MCP specification and ServiceNow's implementation evolve. The combination illustrated here nonetheless points to a practical template: keep the model in your own Azure tenant, expose enterprise systems through standardized MCP servers, and let Entra ID ensure that the assistant never sees more than the user behind it is permitted to. For teams already invested in Microsoft and ServiceNow, it is an incremental path toward AI that answers from real data rather than plausible guesses.