Proof of Concept (PoC) for Enterprise AI Copilot Accessible via Microsoft Teams

About Our Client

Our Client is a US real estate investment company that focuses on healthcare facilities.

Need to Verify the Feasibility of AI-Driven Process Optimization

The Client has an extensive knowledge base that holds information on its internal processes, such as employee onboarding and investment portfolio management. The knowledge base contains diverse file formats, including PDF and XLSX.

The Client wanted to leverage the full potential of the accumulated knowledge and best practices with the help of a custom AI tool. As envisioned by the Client, AI was meant to analyze the data on business processes from the knowledge base and generate intelligent recommendations for their improvement. The Client turned to ScienceSoft for AI engineering services to verify the feasibility of this idea.

Project Discovery to Engineer Requirements for Enterprise AI

ScienceSoft assembled a team of a business analyst, two .NET developers, and a solution architect for the project. To kick off the discovery, we conducted a number of Q&A sessions to get a clear idea of the Client’s expectations for the AI solution.

According to the Client’s vision, the solution was to scan process-specific files (e.g., procedure documents for employee onboarding, investment portfolio reports) and generate recommendations for improving the related processes. Each recommendation had to include exhaustive information, such as step-specific roles and timelines, suggested KPIs, and references to the documents used by AI when generating the recommendation. The Client’s subject matter experts (SMEs) would then discuss the relevance of the AI output and update the knowledge database with the meeting recap. Finally, the AI engine would use the experts’ feedback to learn and adjust future recommendations.

Taking into account the innovative nature of the solution and the impact it could have on the organization’s business performance, the Client and ScienceSoft agreed to start with a proof of concept (PoC) to test the AI’s ability to generate relevant recommendations. The capabilities for updating the output based on human feedback were not included in the PoC, but they would become a part of the full-featured version.

Choosing Optimal AI Services for Document Analysis and Process Recommendations

To choose the optimal AI services for the solution, ScienceSoft ran a test to compare the following options:

• Claude Sonnet (suggested by the Client).
• Azure AI Services (Azure OpenAI + Azure AI Search).
• Microsoft Copilot Studio integrated with Azure Open AI.
• Microsoft Copilot integrated with Microsoft Teams and Azure OpenAI + Azure AI Search.

As the test knowledge base, we used four HR-related PDF files that contained descriptions of HR procedures with step sequences, references to relevant laws and regulations, and screenshots (e.g., of employee surveys). The files were uploaded to Microsoft Blob Storage, which served as a data source for all the testing scenarios.

ScienceSoft used the same test prompts for all four options and compared the results across such parameters as output relevance and accuracy, the ability to adjust the generated content, capabilities for model customization in the future, and the visual representation of the output. Microsoft Copilot (both independently and integrated with Microsoft Teams) demonstrated the best results, while the other two options had significant disadvantages (e.g., the need to convert Excel files to PDF and the poor quality of responses to improvement requests).

ScienceSoft suggested the last option (Microsoft Copilot integrated with Microsoft Teams) as the optimal one. Since the Client uses Microsoft Teams as its primary business messenger, the company’s employees will be able to use it as an interface for sending prompts and receiving AI responses.

Azure-Based Architecture for Streamlined Data Processing and AI Integration

After choosing the main technologies, ScienceSoft created a high-level solution architecture with the following components:

• Microsoft Teams serves as the main user interface.
• Azure Web App, a custom web application for data orchestration, automates repetitive data processing tasks such as data search and analysis to enable smooth communication between solution components.
• Azure OpenAI hosts the AI model and enables its operation.
• Azure AI Search is a cloud search engine that scans the knowledge base, indexes the information, and serves data to OpenAI for analysis.
• Azure Blob Storage keeps the knowledge base files. ScienceSoft recommended creating containers dedicated to specific business processes — this will optimize solution performance, as AI will need to process only relevant data instead of scanning the whole database.
• Azure Key Vault stores access keys to AI services.

Proof of Concept for Enterprise AI Solution Ready in 8 Days

With ScienceSoft’s assistance, the Client was able to prove the feasibility of its idea to optimize business processes with the help of AI. After a series of tests performed by ScienceSoft, the Client received an optimal tech stack (Microsoft Copilot integrated with Microsoft Teams and Azure AI services) that provides the most accurate recommendations for improving business operations, allows users to adjust the AI output, and easily fits into the company’s established workflows. The Client also received a high-level solution architecture that features dedicated components for data orchestration and security enablement. With the resulting PoC, the Client has a solid foundation for further solution development.

Technologies and Tools

Azure OpenAI, Azure AI Search, Microsoft Copilot Studio, Microsoft Teams, Azure Blob Storage, Azure Key Vault, Claude Sonnet, .NET.