%20(1).png)
.png)
Simplifying mailing cost estimates by a natural chat experience
How I leveraged the strengths of a chatbot to enhance intuitiveness
Finding the right USPS service among their diverse options is challenging
USPS offers various mailing services. It has a mailing price calculator feature on its website, but finding the most suitable service for a user’s specific package can be challenging.
%20(1).png)
%20(1).png)
%20(1).png)
How might we help users calculate postal prices and choose the best USPS service without overwhelming or confusing them?
Chatbots can provide personalized, easy and natural experience like talking to a real postal officer
A chatbot can effectively address the problems mentioned above due to the following advantages.
The chatbot calculates mailing costs for each user’s unique package and suggests the best service
The bot asks users about their service preferences, such as the cheapest or fastest option, and then collects package details, including dimensions, weight, destination and origin ZIP codes, and drop-off date and time. After confirming that all inputs are correct, it provides a price estimate and recommends the best USPS service with a brief explanation of its features.
Efficient, calm and helpful
In the initial phase of the project, I defined the bot's personality as follows. Overall, I envisioned the interaction as similar to USPS’s in person service.
Planned and documented how the bot works
I created a sample script, intent training data, and a flowchart to support prototyping. Drawing on my engineering background, I incorporated error handling early and defined how the chatbot would manage edge cases.
The bot's personality was well conveyed, but some users overlooked service-based price differences and recommendation features
I built a prototype in Voiceflow and tested it with 3 participants, 1 somewhat familiar with USPS and 2 unfamiliar. The scenario focused on returning an online purchase through USPS. After using the chatbot, participants were asked to complete the same task with the existing GUI to compare the 2 interfaces.
Some users misunderstood the bot’s capabilities, assuming it could only perform calculations, while it can also provide recommendations and service comparisons. I revised the introduction to communicate these capabilities more clearly.
.png)
.png)
Users wanted more information, such as nearby post office locations and operating hours, so I added those features.
The bot, by default, asked for width, height, and length one by one, but all testers preferred entering all dimensions at once.
.png)
.png)
To keep the conversation concise, I kept it brief and instead included a link for users to explore more details if they wish.
If they provide only one dimension, the bot adapts by prompting for the remaining dimensions—length and height—separately.
If users explicitly indicate that they are only interested in the flat rate service (fixed price), the bot skips the dimension request and asks whether the package weighs less than 70 pounds (maximum allowed weight).
If the ZIP code is incorrect or users indicate they forgot it, the bot prompts them to provide their address instead.
Then, the bot asks users if they are sending from their current location. If yes, it fetches their current location. Otherwise, it prompts them to enter the origin ZIP code.
Ideally, I wanted to implement a calendar input field. However, due to Voiceflow's feature limitations, I used a standard chat input.
(Insurance)
Understand the strengths of conversational UI and intentionally choose the most appropriate solution
Through testing, I learned that some users prefer the chatbot to explore services and find the best option, while others prefer the GUI for quicker answers without fully understanding the service differences. It is important to give users the freedom to choose their preferred interface to address diverse needs.
In this project, AI was used only in the backend for tasks such as identifying city names from ZIP codes. It was not used to directly generate responses in order to avoid inconsistent results where precision is critical. However, I am interested in exploring AI generated conversations in the future.
%20(2).jpeg)
.jpeg)
.png)
%20(1).png)
%20(1).png)
