Organizations increasingly recognize the potential of integrating AI and agents into complex business processes yet lack practical guidance on how these processes can be translated into AI-augmented systems. The proposal development process for Request for Proposals (RFPs) is an example of such a process as it requires multidisciplinary expertise, adherence to strict requirements for both content and formatting, and consistent quality across writing from individuals and teams. Creating proposals consumes a lot of time and resources, while at the same time quality is essential as they can be decisive in winning new business.

This research was conducted in collaboration with Schuberg Philis, a mission-critical IT company, to explore how multi-agent systems can be applied to their RFP response procedure. The study adopted a multi-method approach: a literature review builds the theoretical foundation of MAS, contextual research in the form of interviews and training observations mapped the current challenges in proposal development, followed by iterative prototyping with a domain expert to design and refine agent roles and evaluate their outcomes, and concluding with a workshop on human-AI interaction guidelines. The prototyping phase utilized the workflow automation platform n8n to develop and test various agent configurations, orchestration patterns, and knowledge retrieval and sharing mechanisms.

The findings show that MAS offer the most value in knowledge retrieval from fragmented sources, requirement alignment with RFP specifications, and drafting compliant content following Schuberg Philis’ best practices for proposal writing. The proof of concept demonstrated that MAS design requires a careful balance between agent specialization and over-fragmentation, considerations on parallel executions and sequential executions, and defining explicit instructions for agents’ tasks and output hand-offs. Human oversight remains essential for adoption, with traceability and controllable levels of agent autonomy as primary factors for trust.

In addition to the proof of concept, the research delivers a set of design guidelines that translate the findings into actionable steps for developing MAS in business contexts. These guidelines cover the selection of a business process, role and task decomposition, orchestration patterns, iterative agent development, combining agents to a MAS, and human-agent collaboration.

Together, the proof of concept and guidelines provide both a concrete demonstration of a MAS tailored to proposal development at Schuberg Philis, and a practical foundation for organizations that aim to implement MAS in real-world business contexts. While the guidelines offer structured and actionable steps, further research is needed to validate scalability and applicability to other situations.