SOW Generation

Background

Zazmic’s project management team frequently engages in solution workshops with clients, which are a crucial step in defining project scope and deliverables. However, the subsequent process of creating a formal Statement of Work (SOW) was a manual and time-consuming task. It required PMs to sift through various documents—including meeting transcripts, estimate spreadsheets, and presentation slides—to extract and synthesize information into a coherent document. This process was a significant administrative burden and a bottleneck in the project lifecycle. To address this, an internal project was initiated to leverage Zazmic’s expertise in generative AI to create a tool that could automate this critical workflow. The goal was to build a sophisticated solution that could intelligently process unstructured data and generate accurate, consistent SOWs with minimal human intervention.

Business Challenges

The project management team identified several key challenges with the traditional SOW creation process:

• Inefficient Data Extraction: Manually parsing and consolidating information from long, unstructured documents was time-intensive.
• Information Discrepancies: Maintaining consistency across different data sources was difficult, especially when changes were made to the project scope during a workshop.
• Hallucination Issues: A straightforward approach of inputting all source data into a single LLM call resulted in poor output quality and factual inconsistencies.
• Scalability Limitations: The manual process did not scale efficiently for workshops that included multiple use cases or complex project phases.

Solutions Delivered

Zazmic’s development team created an AI-powered service designed to overcome these challenges. The solution automates the transformation of workshop deliverables into a high-quality SOW document. The system is built on a modular architecture that integrates seamlessly with Google Drive to handle both input file retrieval and the publishing of the final SOW document.
The core solutions delivered are:

• Modular Architecture: The system employs a multi-step process, with separate LLM calls for each section of the SOW. This approach minimizes hallucination, reduces latency and cost, and allows for precise control over the content of each section.
• Intelligent Data Processing: The service first pre-processes all input files, using Python to extract only the necessary information from each document. For instance, it isolates the “backlog” tab from the estimate spreadsheet and the executive summary from the workshop SOW.
• Context Aggregation & Grounding A crucial intermediary step uses LLMs to summarize each data source into a unified knowledge base. This prepared context is then used to ground all subsequent content generation, ensuring accuracy and relevance. All information is grounded on the estimate to ensure that there is no scope creep coming from themes discussed in customer calls that are not part of the SOW.
• Template-Based Generation: The final SOW is generated from a pre-defined Google Docs template. The service creates a copy of the template and replaces designated placeholders (e.g., ##client_name##) with the content generated by the LLMs. This guarantees consistent formatting and a professional final document, including handling multi-level bullet points and bolding.
• Flexible Filtering The tool allows users to specify which use cases from the workshop to include in the SOW, providing the flexibility to generate documents for specific parts of a multi-use case project.

Outcomes

The SOW generation tool has had a significant impact on Zazmic’s internal operations:

• Massive Productivity Gains: The average time for a product manager to create an SOW has been reduced from 2-3 hours to just 30 minutes.
• Improved Consistency and Accuracy: By grounding the generated content in the estimate spreadsheet, the tool ensures that every SOW accurately reflects the final scope of work and is free of inconsistencies and scope creep.
• Reduced Manual Effort: The administrative burden of SOW creation has been largely automated, reducing the team’s effort to only reviewing the document for potential gaps and inconsistencies.

This case is an example of how generative AI can be integrated into existing business workflows to automate complex, administrative tasks, thereby freeing up valuable time and resources for high-impact activities.