Use Case Definition

For this lab, I created Pack Support, a custom GPT for Wolfpack Outfitters, NC State’s online store. I built it around a simple but important business challenge, the customers want quick answers, especially when an order is involved. Many of the questions a store receives are routine, but they still affect trust, satisfaction, and whether the shopping experience feels smooth or frustrating. I designed Pack Support to focus on order-related help first. That includes order status, shipping updates, tracking guidance, pickup questions, order-change requests, and missing or incorrect items. It can also help with returns, exchanges, sizing, store hours, and general store procedures. I thought this was a strong use case because the information already exists on the official store website, but customers still need it in a faster and easier format. Instead of making staff answer the same questions repeatedly, a custom GPT can handle first-line support and point customers in the right direction.

 

GPT Design

I gave the GPT a clear role, a clear tone, and strict boundaries. The voice was designed to feel warm, helpful, collegiate, and dependable. I wanted it to sound like a real support teammate, not a script. I also built empathy into the instructions because customer service usually works better when the response feels human and calm, especially if the customer is confused or frustrated. The most important design choice was limiting Pack Support to official information from shop.ncsu.edu. It was instructed to use the Customer Support page for shipping, tracking, pickup, returns, exchanges, and order questions, the Size Guides page for sizing help, and the Locations & Hours page for store address, hours, and contact details. If the site does not clearly confirm something, the GPT should not guess. It should say that the detail could not be verified and suggest the customer to call or email for more information. I also added guardrails so it would never invent policies, prices, inventory, delivery timelines, discount codes, or live order details. It should not claim access to backend systems, payment records, or tracking tools unless the customer provides that information directly in the chat. That matters because customer-facing AI needs to be accurate and trustworthy, not just fast.

 

Testing and Results

I tested Pack Support with realistic customer scenarios. These included questions about how to track an order, where guest shoppers receive tracking updates, whether an order can be changed after checkout, where in-store pickup happens, whether a final sale item can be returned, how to use the Adidas footwear size chart, and what the official store hours and contact information are. Overall, the GPT performed best when the question matched the official site closely. It was especially strong with order-related questions, which was exactly what I intended. The responses worked best when they started with a short, direct answer and then gave the next step only if needed.

 

The biggest issue I found during testing was overconfidence. Sometimes an answer sounded polished even when the official site did not fully confirm the detail. I refined the instructions so the GPT would separate verified information from anything uncertain and avoid sounding too sure when it should escalate the question. I also adjusted the tone, so the replies felt more natural, conversational, and less templated. After that, the GPT became more balanced, more trustworthy, and better aligned with the role of a first-line support assistant rather than a replacement for store staff.

 

Business Value and ROI

Pack Support creates value in a practical way. First, it gives customers faster answers to routine questions. Second, it reduces support friction because customers do not have to dig through multiple pages to find basic information. Third, it gives the store a more consistent support experience because the GPT follows the same approved guidance every time. The clearest return on investment is time savings. If Pack Support handles even 20 routine questions per week and saves about 5 minutes of staff time per question, that equals about 100 minutes per week. Over a 32-week academic year, that is roughly 53 hours saved. This is only an estimate, but it shows how even a small use case can create measurable operational value. There is also likely a customer-side benefit. If shoppers get quick answers about shipping, pickup, returns, or sizing, they may feel more confident completing a purchase instead of leaving the site.

 

 

Ethical and Risk Analysis

The biggest risk is misinformation. If the GPT gives the wrong answer about a policy or store procedure, it could create confusion and hurt trust. Data privacy is another concern, especially if customers share unnecessary order or payment details. There is also the risk that customers may assume the GPT has access to live systems or decision-making authority when it does not. To manage these risks, I built governance directly into the instructions. Pack Support is limited to official Wolfpack Outfitters information. It is told not to guess, not to claim access to backend systems, and not to promise outcomes. When a situation requires human review, it directs the customer to the official store contact information. This project showed me that a custom GPT can create real business value when it is focused, grounded in trusted information, and clear about its limits.

 

Pack Support Custom GPT Link:

https://chatgpt.com/g/g-69c72ae6c21481918c7bf8c42f5663b9-pack-support