Coding the G.O.A.T. Debate
The websites for popular Artificial Intelligence applications like ChatGPT, Claude.AI, & Gemini receive millions of visitors everyday where users can intuitively create prompts and take the output for their various needs.
These simple interfaces are the most straightforward way to interact with these “closed-sourced frontier models” but may beg the question, “How do you build a ‘smart’ application that has integrated with these Large Language Models for business purposes?”
First we need a brief explanation on what an Application Programming Interface (API) is. An API allows for two software programs to speak to each other.
When I asked both ChatGPT and Claude for assistance in explaining what an API is to a non-technical person, they both referenced the analogy of a restaurant.
You are the customer. You tell the waiter (API) what you want from the kitchen (application) and when your request (information) is ready it is served to you.
Let’s start with an example and then dive into some of the more technical aspects of why this is applied in a business context and how this is made possible.
There are 3 things to pass to the LLM API of your choice:
- Which model (version) are you wanting to leverage (e.g. GPT 4.1 Mini vs GPT 4.5)
- The system message that provides the “role” the LLM adopting (“You are a fanatic”)
- The user message that includes the prompt the model is going to ingest. (“Explain/Create/Draft/etc…)
gpt_model = "gpt-4o-mini"
claude_model = "claude-3-haiku-20240307"
gpt_bot = "You are a quarrelsome enthusiast of the National Basketball League. \
You firmly believe that Michael Jordan is the best basketball player ever."
claude_bot = "You are a contentious fanatic of the National Basketball League. \
You are completely convinced that LeBron James is the best basketball player ever."
gpt_chat = ["Who do you think is the best basketball player ever?"]
claude_chat = ["Who do you think is the best basketball player ever?"]
Define a function for each chatbot, pass the necessary parameters, and get the conversation started.
Business Use Case
Why would a Chief Operating Officer be interested in virtual assistants that talk to each other? A use case that is less subjective and involves dispersing definitive truth in a parent-child hierarchical relationship is examined through the eyes of a company local to Utah.
Given their footprint in sports, entertainment, and community engagement, the Smith Entertainment Group has a variety of customer and employee points of contact that would benefit from implementing a “smart chatbot hub” strategy.
What this means is a centralized conduit for business logic and information, collecting insights and analytics, and coordinating a more seamless experience across applications. This one-chatbot-to-rule-them-all would disseminate information to and from many sources including:
- The Utah Jazz app chatbot “Notey”. (Which should be immediately accessible via a Floating Action Button on the home screen and not buried in the ‘More’ submenu, but that is a UX post for another day).
- A social media bot that responds to direct messages via Instagram, TikTok, etc.
- Updating the Amazon Alexa Utah Jazz skill so it stays current.
- A DeltaCenter.com bot that allows customers to purchase tickets or request refunds to events.
- Employees that might ask for a HR or other policy via a bot on Slack.
- Collecting internal feedback via Qualtrics.
These are all handled by separate systems initially setup in isolation from one another. It could be hard to keep messaging consistent, update policies across platforms, and understand the full customer journey. This hub knows who it’s talking to—fan, donor, partner, or internal staff—and gives the right information consistently across channels.
Measurable Impact
But how can I know if this strategy is making a positive difference? Metrics to track may include:
- Chatbot Engagement Rate = How many visitors are purposefully engaging with the chatbot?
- Session Completion Rate = Percent of users that get what they need.
- Click-Through Rate = How often suggested links to a landing page, form, or product is clicked.
- First Contact Resolution = Percent of inquiries resolved within one interaction.
- Time Saved by Internal Teams
- Satisfaction Scores = Able to be primed by offering a promotion code for customers that rate their interaction. Usage of promotion code can also be reported on.
Final Thoughts
In the end I guess the G.O.A.T. debate wasn’t decided after all. Another optional parameter that can be utilized by an LLM API is referred to as the temperature of the request. Maybe this would have led to a less amenable outcome.
“A lower temperature makes the model more likely to choose the most probable next word, leading to consistent and predictable outputs. Conversely, a higher temperature allows the model to explore less likely options, resulting in more varied and potentially creative responses.“ – Vellum AI
Experimenting with temperature can impact the accuracy of the output, but may give your model a little more personality.
It costs money to send requests to an LLM API. Otherwise I would have made this post a little more dynamic by including two input fields from a package like Gradio that would have allowed a visitor to select from a variety of LLMs, the temperature of that model, and any two athletes to compare their bona fides.