Inspiration
Many current reimbursement systems are bogged down by inefficiencies, with repeated back-and-forth between employees and finance teams due to missing materials, unclear policy violations, or manager rejections. Employees often struggle to navigate complex and fragmented reimbursement policies, receiving delayed decisions without clear guidance on how to revise their requests. At the same time, finance teams are burdened with repetitive manual checks and audits, draining valuable time and resources. These challenges are often made worse by fragmented systems that lack a centralized workflow, further slowing down the process and creating confusion across departments.
What it does
Our system is a conversational AI assistant that automates expense reimbursement. It helps employees submit requests by extracting and validating info from receipts and messages, checks policy compliance, routes for approvals, and notifies users of updates. Reviewers receive full context to make informed decisions, and users can track progress through a dashboard. This streamlines the process, reduces manual checks, and ensures clarity and efficiency.
How we built it
We built the system using a modular agent-based architecture, where each agent handles a distinct part of the reimbursement workflow: request submission, policy validation, review processing, and reporting. These agents are organized under clearly structured submodules (request, policy, review, reporting, and dashboard) to keep responsibilities isolated and maintainable. Supporting utility functions such as database access, input validation, and notification delivery are implemented in reusable tools like database.py, validation.py, and notification.py. This setup allows for clean orchestration and easy expansion. The system is designed to run as a conversational interface, enabling seamless interaction through a chatbox while automatically managing validation logic, approval routing, and user notifications in the background.
Challenges we ran into
One major challenge was ensuring consistent identification and tracking of each reimbursement case within a dynamic conversation flow. We had to design a reliable mechanism to associate natural language references with specific case IDs to avoid confusion and ensure smooth follow-ups. Another challenge was verifying attachments: our AI agent occasionally misinterpreted whether a file was actually uploaded. Beyond checking file formats (e.g., PDF, PNG, JPG), we also needed to validate the content to ensure it met justification requirements, such as including dates, amounts, or vendor names, all without adding friction to the user experience.
Accomplishments that we're proud of
We successfully built a fully functional, end-to-end conversational reimbursement system that handles complex workflows with minimal user effort. Employees can submit requests using natural language, and the system dynamically guides them through validation, policy checks, and submission without requiring them to read or understand internal policies. Reviewers receive well-structured summaries with all necessary context, enabling faster and more informed decision-making. We also implemented automated email notifications, approval routing, and a real-time progress tracking dashboard—all integrated seamlessly into a modular architecture. Our system not only reduces processing time and manual errors but also lays a foundation for future AI-driven features such as auto-fill and predictive justification suggestions.
What we learned
Building this system taught us the importance of designing AI workflows that balance automation with user control. We learned that even simple tasks like submitting receipts involve many edge cases from missing attachments to vague justifications, which require thoughtful prompting and flexible error handling. We also gained experience in building modular agent architectures, which helped us isolate responsibilities and scale the system more easily.
What's next for Enterprise Smart Travel Concierge
One key direction is enabling automatic retrieval of reimbursement-related information and materials from sources like emails, calendars, and messaging platforms. For instance, if a user booked a flight or hotel, the system could extract relevant receipts, itineraries, and vendor details directly from their Gmail or calendar events, reducing the manual burden on users. In addition, we are exploring enhancements in content verification using multimodal AI and OCR, allowing the system to better interpret images and PDFs for relevant data points like total amounts or dates.
Log in or sign up for Devpost to join the conversation.