MaCo Onboarding Agent: Exploring its Multi-Agent AI Architecture

[SonamSingh]

By Sonam Singh, Francisco Freitas

Overview

As AI-driven transformation reshapes enterprise systems, intelligent onboarding becomes essential for improving customer experience and accelerating product adoption. In MaCo (SAP Market Communication for Utilities), we are driving an innovation initiative by developing the MaCo Onboarding Agent—a multi-agent AI solution.

The MaCo Onboarding Agent is designed to engage with customers before they subscribe to MaCo saas. On request, the agent is intended to collect essential information such as Global Account, subaccount, and subdomain, and generate an onboarding incident to initiate and manage the onboarding process. This removes manual overhead, especially for customers with multiple subscriptions.

Using a vectorized knowledge base from MaCo documentation, downstream service documentation, and FAQs, the assistant offers real-time, context-specific support. For example, based on the customer's monthly message volume, the assistant can recommend the appropriate subscription plan, for testing or production use, by referencing the relevant information maintained in the FAQs. This ensures customers receive timely and contextually accurate guidance, improving their overall experience.

Architecture: Engineered for Scalability, Modularity, and Intelligence

The MaCo Onboarding Agent is built on a supervisor-level multi-agent architecture, designed with a modular and extensible technology stack:

• Python with LangChain and LangGraph: Serves as the orchestration layer for agent-based interactions. LangChain provides structured tooling for chaining LLM prompts and external calls, while LangGraph introduces persistent, event-driven workflows with stateful memory management and multi-agent routing logic, exposed as a RESTful API.
• SAP CAP (Cloud Application Programming) using Node.js
• SAP BTP (Business Technology Platform)
• SAP HANA for Semantic Search and Context Retrieval 
  MaCo documentation, downstream service documentations and references, and FAQs are converted into vector representations stored in SAP HANA's vector database. This allows the assistant to perform semantic searches and retrieve information based on meaning rather than exact keywords. At runtime, relevant content is dynamically retrieved from the vector store to build context for responses.
• Multi-Model Support (e.g., GPT-4.0, Gemini): Enables runtime selection of multiple large language models based on use case requirements such as accuracy vs. latency trade-offs, tokens constraints, and response time requirements. The classification logic is also handled by LLMs, based on the prompt and input context.

Supervisor-Agent Routing Flow

Supervisor Agent: This agent acts as the orchestrator. It receives user inputs—ranging from questions to actionable requests, to determine an appropriate response strategy. To do this effectively, it uses structured prompts, including simulated human messages, to guide the language model's classification. Depending on the nature of the query, it may engage one or more of the specialized sub-agents.

Knowledge Retrieval Agent: This RAG agent utilizes vectorized knowledge base to search for relevant information chunks, followed by summarization using an LLM to provide a context-aware response.

Incident Management Agent: This agent handles actionable user requests related to incidents, such as creating, updating, or retrieving incidents based on processing logic and integrated system rules.

Fallback Agent: This agent is engaged when the user query is unclear, out of scope, or lacks relevant information. It helps maintain a meaningful conversation by steering the dialogue in a helpful direction.

User Interface

What’s in Progress

The MaCo Onboarding Agent extends its capabilities beyond a standard chatbot assistant. Integration with JIRA and ServiceNow APIs is also underway, enabling a dedicated agent to create and manage incidents.

Business Benefits

Efficient Onboarding Process

• Accelerates the onboarding process by providing context-specific support and resolving queries in real-time
• Particularly valuable for enterprise customers with high-volume onboarding requirements by automating manual tasks.

24/7 Real-Time Support

• Leverages a vectorized knowledge base for real-time, context-aware assistance.
• Once subscribed, customers go through multiple setup steps. The assistant provides on-demand access to knowledge from MaCo documentation, downstream service references, and FAQs — enabling users to resolve known issues or questions independently. This reduces reliance on support teams and minimizes delays caused by time zone differences or limited team availability.

Continuous Customer Engagement

• Keeps customers informed on ongoing issues with downstream services or MaCo issues by gathering information from FAQs.
• Ensures ongoing transparency and trust during the onboarding lifecycle.

Conclusion

The MaCo Onboarding Agent demonstrates how intelligent automation and a multi-agent solution can enhance the efficiency and consistency of early customer interactions. By providing access to relevant information and supporting decision-making with contextual guidance, the solution enables a more scalable and responsive customer experience.

Ongoing enhancements would explore extended capabilities, including deeper integrations, improved coordination between agents, and alignment with evolving customer and product needs.