Sage Intacct AI Agents and Model Context Protocol (MCP) Study Guide

Introduction to Sage Intacct AI Agents and Gateway Louis Stereo, Product Manager for the Platform Business Unit at Sage Intacct, introduces the session focused on launching AI agents using the Model Context Protocol (MCP). The session includes contributions from Eric Cunningham, Principal Product Manager for Data Platform, and Chad from Whitley (Wipley), a strategic partner. The presentation is designed for a diverse audience, ranging from highly technical developers to non-technical users and 'citizen developers.' Stereo emphasizes three core pillars of Sage's AI strategy: Confidence, where every answer provided by the AI is explainable and interrogated; Control, where business data is handled with user-specific context and permissions; and Logic, where every AI action is logged via API calls, ensuring full auditability. # The AI Gateway and MCP Server Architecture The AI Gateway is technically an implementation of an MCP server. It provides entitlement to use the Sage Intacct REST API. To utilize this service, customers must meet a specific prerequisite: possession of a Web Services Developer license. Upon subscribing to the AI Gateway add-on, Sage provisions a 'sender ID' with a flag marked 'AI enabled' as true. The cost for this add-on is approximately $360\,USD$ per year. Currently, the MCP server provides read-only access with coverage for specific modules: Accounts Payable (AP), Accounts Receivable (AR), Cash Management, General Ledger (GL), Purchasing, and Order Entry. Coverage is expected to expand over time. The system utilizes two primary services: the Model Service for interrogating schemas and the Query Service for retrieving data. Connection is established via OAuth using a Client ID and Secret generated from the App Registry developer portal. Users plug the MCP server URL into a compatible client, such as Claude. The service is initially available to customers and partners in the United States and Canada. # Use Case Demos: Automated Integrations and Utilities Stereo demonstrates the 'art of the possible' with two primary use cases. The first involves a HubSpot to Journal Entry integration. In this scenario, an MCP client reads 'deals' from HubSpot, reasons over the Sage Intacct Chart of Accounts, and generates schema-accurate journal entries in bulk. This is achieved through 'artifacts' in Claude that create copy-pasteable REST API calls. Stereo notes that write access is intentionally omitted to maintain a 'human-in-the-loop' principle for security and trust. The second demo involves an Artist Consignment Settlement utility. Designed for a music retailer, the utility calculates a $70\%$ payout for independent artists based on a custom 'payout' field on the vendor object and AR invoices tagged with vendor dimensions. The AI generates specific API calls for AP bills, a process that previously required 30 to 40 hours of manual coding, now completed in 15 minutes via the MCP server. # Questions and Discussion During the Q&A, an audience member asks about the role of the Sage Protocol AI Gateway in routing questions. Stereo explains it acts as a gateway to multiple Large Language Models (LLMs), routing domain-specific questions to the most appropriate model (e.g., ChatGPT or Claude). Another participant asks about 'governors' or limits on large queries to prevent API overages. Stereo admits that specific governance policies are still being finalized but strongly recommends testing in a sandbox environment initially. Sandbox API usage is not subject to performance tier overages. A final question addresses multi-entity environments. A user noted that queries sometimes default to the top level rather than specific business units. Stereo clarifies that the Query Service includes properties for entity-specific information (using Location IDs or Entity IDs) and suggests specifying the entity ID during the connection or query process. # The Sage Agent Builder Program The Sage Agent Builder is a partner-exclusive program and development framework for creating custom AI agents within Sage Copilot. It involves a strict certification and governance process managed by a committee of product managers. Partners register their use cases on the developer portal ($developer.sage.com/intac$). If approved, they undergo technical, security, and legal reviews. Agents are distributed through an in-product Sage Marketplace and the AWS Marketplace. These agents can connect to external systems like Salesforce or Procore to bring insights into Intacct. While the program is currently for Independent Software Vendors (ISVs), there are long-term discussions about opening it to customers. Partners host the infrastructure and provide their own LLMs. The agent lifecycle includes registration, validation in lower environments (sales demos), certification, and finally, publication to the AWS and Sage Marketplaces. # Wipley Job Insights Case Study Chad from Wipley presents the 'Job Insights Agent,' a vertical solution for construction and project-based companies. The agent allows project managers and CFOs to ask questions in natural language, such as 'What is my cost to completion by cost code?' or 'Show me the Earned Value Management (EVM) metrics.' The architecture involves an 'agent port gateway' hosted on AWS. Because construction formulas (like EVM) are complex, the agent uses a controlled process to ensure consistent results. There are two tiers: a 'Client Essentials' tier using only Sage data, and a 'Sage plus Procore' tier that blends data from both systems. This allows for reconciliation, such as identifying direct costs in Procore that have not yet been recorded in Sage. The runtime process involves Sage Copilot making a POST request to a web endpoint, providing a prompt, an ID token (exchanged for a REST token), and a session ID to maintain conversation context. # Sage Data Cloud and Snowpoint Eric Cunningham introduces the Sage Data Cloud (released in R1 February 2024), which utilizes Snowflake for 'Data Cloud' functionality. This allows users to access Sage Intacct data via a 'private share' and run direct SQL queries instead of using APIs. This method is highly scalable, avoiding the limitations of API call rates. The data is structured into curated views based on Sage's SQL APIs. Cunningham demonstrates creating a dashboard and a 'cash summary' AI directly within Snowflake using Python, showcasing that users can build powerful analytical tools and AI summaries within their existing data warehouse infrastructure without needing the MCP server. # Roadmap and Future Developments The roadmap for AI Gateway includes hardening existing coverage and eventually introducing 'human-in-the-loop' write privileges, asynchronous/offline support, and reporting services. The Agent Builder program is currently in beta with three ISVs. Early summer (June) marks the Early Adopter (EA) phase with the launch of the in-product marketplace. General Availability (GA) is scheduled for late summer. Future updates will include a 'Task Agent' button in the marketplace to immediately load prompts into Copilot and a 'context shift' feature in the UI to maintain agent selection across queries. All agents will be required by policy to explain how they arrived at a specific response, ensuring transparency.