Reggie Bot: AI assistant to answer questions about academic policies, procedures, and services

The problem

The United States Military Academy at West Point develops U.S. Army officers through an academic, physical, and military training programme. The Academy's Office of Academic Affairs and Registrar Services has a broad remit covering academic records, end-of-term grade processing, transcript generation, graduation facilitation, and academic success counselling.

Much of this work is administrative and policy-heavy. Staff regularly received a high volume of inquiries from cadets, faculty, other staff, graduates, and external stakeholders about academic policies, procedures, and entitlements. Many of these questions were repetitive and could be answered from existing published guidance, but they still required individual staff attention. This placed a consistent burden on administrative capacity, particularly given that the policies involved are detailed and contain exceptions and edge cases that require careful handling.

The solution

The office developed Reggie Bot, a policy-constrained AI assistant built as a custom GPT within the ChatGPT platform. Rather than functioning as an open-ended chatbot, Reggie Bot was designed from the start as a governed institutional assistant with a tightly defined scope.

How it works
Reggie Bot answers questions about academic policies, procedures, and services by drawing on a curated knowledge base built from the office's existing policies, handbooks, and standard operating procedures. When users ask a question, the bot provides guidance based on this knowledge base and points them to the relevant online resource or the appropriate person for an authoritative answer.

The system was explicitly designed so that it would not provide authoritative answers or make decisions. The team's guiding principle throughout was that human authority remains central. Reggie Bot explains processes and clarifies policies, but always routes users to the office or individual who holds the formal authority to make a determination. Responses are also shaped by the inferred role of the user, whether they are a cadet, faculty member, member of the public, or another stakeholder.

Users access Reggie Bot through a forwarding link that takes them to the custom GPT. The link is also published on the registrar's public webpage, meaning it is available to external users such as parents and prospective cadets as well as the internal academy community.

How it was built
The knowledge base was assembled from materials the office already had: published policies, handbooks, and operating procedures. Through iterative testing, the team identified gaps where the bot produced incorrect answers or hallucinations, and wrote new content to address those gaps, capturing tacit knowledge that had not previously been documented.

A pilot group of department academic counsellors tested the bot before it was published to the wider institution. The computer science and systems engineering departments provided technical feedback, while counsellors from other departments contributed user experience insights. Feedback was used to refine the knowledge base, adjust guardrails, and improve prompting.

The impact

1. Reduced repetitive workload on administrative staff
Reggie Bot now handles routine policy and procedural inquiries that previously required individual staff attention, freeing time for more complex work. Departments that may not have had strong existing relationships with the registrar's office reported finding the bot a comfortable way to access information they might not otherwise have sought out.

2. Staff confidence in AI increased through constrained design
Staff were more willing to experiment with AI because the assistant was intentionally constrained rather than positioned as an autonomous decision-maker. Framing the tool as administrative support rather than automation was important for organisational acceptance.

3. The process exposed gaps in existing institutional knowledge
Building and testing the bot revealed outdated content on the office's internal SharePoint portal and inconsistent terminology across published materials. The process of curating the knowledge base served as an unintended audit of the institution's information quality.

Lessons learned

• Start with restrictions, not openness. The team initially trained the bot with broad access to information and found this produced unreliable results. The most valuable lesson was that narrowly scoping the AI's mission from the outset significantly improves reliability. If doing it again, they would begin with a restriction-first approach rather than constraining after the fact, to ensure that the information the tool could draw from was relevant and clear.
• Defining boundaries is harder than building the tool. The hardest part of the project was not the technical build but defining the authority and boundaries for the bot's responses. Academic policies contain exceptions, overlapping authorities, and edge cases that required careful thought about what the bot should and should not attempt to address. The institutional process of mapping what the bot should know, how it should respond, and where it should escalate required more time and effort than the technical work of building and training the custom GPT.
• Maintaining the knowledge base requires ongoing ownership. The source material the bot draws on needs to be kept current as policies change. This is an operational commitment, not a one-off task.
• Users may interpret guidance as approval. Some users treated the bot's informational responses as if they were official decisions. This reinforced the importance of clear language in the bot's responses and the design principle of always routing users to a human for authoritative answers.
• Policy changes can force platform changes at short notice. A government-wide policy decision required the team to abandon their original platform and rebuild on ChatGPT with little warning. The experience demonstrated that the knowledge and design principles were transferable, but also underlined the importance of not becoming dependent on a single platform.