As contact centers migrate to the cloud, the need for transformative AI has never been greater. ChatGPT, a recently released Generative AI model, has taken the world by storm and sparked the interest of business executives. Its ability to generate fluent and useful text has understandably captured the attention of many leaders, across a wide variety of industries, looking to stay ahead of the curve.
At Cresta, we’ve been working with world-class companies, such as Holiday Inn and Blue Nile, to leverage Generative AI for the past four years. In this blog post, we’ll share our views on how businesses could bring ChatGPT-like capabilities to the enterprise in a reliable and effective way.
Ingredient #1 of ChatGPT: Emergent Capabilities
Since Cresta started deploying AI in 2017 with customers, the landscape of natural language processing has been constantly changing. Every two to three years, there’s a significant paradigm shift: from deep transfer learning (2018) to large language models (2020) to ChatGPT-like capabilities (2022).
In 2018, we were one of the first companies to deploy an in-house generative model based on GPT. It powered our Suggested Responses and Smart Compose features. Generative AI, the new craze in technology, was unpopular back then for production. Getting it to work accurately is hard. Doing it in real time is harder. This was the era before GPT-3 and OpenAI API. But we believed it was the future. Our engineers developed techniques to finetune generative models based on customer needs as well as serve them with low latency. The result is a better conversation-level outcome for our chat customers.
GPT stands for Generative Pre-trained Transformer. The concept of pretraining in layman’s terms is: to become an expert in different tasks (e.g. English sales conversations), one must pretrain the model to be great at English. But what drove the technological breakthrough was a unique discovery. As researchers pretrain larger and larger language models with more data and computation resources, the models start to have “emergent capabilities”, such as more fluent text generation, better reasoning, and question-answering capabilities. This is the first ingredient of ChatGPT.
Cresta’s Generative AI leverages the same underlying technology with such emergent capabilities. The bespoke nature of our AI system combined with the tailwind of increasingly powerful language models has proven to be a successful paradigm. This allows our AI models to have high accuracy in real-time to power our product features and capabilities such as Real-Time Coaching, Suggestions, Smart Compose, Auto Summarize, and AutoQA.
Ingredient #2 of ChatGPT: Alignment with Preferences
Companies have applied large language models to virtual agents. For example, Google’s LaMDA. But these virtual agents, while being trained based on human conversations, tend to generate “chit-chat” responses. Why is ChatGPT different?
The second key ingredient underlying ChatGPT and Cresta’s Generative AI is alignment with preferences. Given a conversation’s context, there could be different ways to respond. For example, when the customer is complaining about a product issue, a contact center agent should respond with empathy. Here, empathetic responses are preferred over chit-chatting or simply jumping to the solution.
ChatGPT is specifically trained to generate responses that are aligned with human preferences, using a technique called Reinforcement Learning with Human Feedback (RLHF). During its training, human labelers are asked to rank model responses to ensure the preferred responses have better ratings.
In a similar spirit, Cresta’s AI models every single customer conversation by understanding which responses are preferred over others and ultimately lead to better conversation-level outcomes. The difference is that ChatGPT embodies universal human preferences. Just like there’s a gap between the average human and an expert, learnings within the contact center domain allow our model to understand expert behaviors. Our outcome insights diagnose key factors for conversational success and are used to align Generative AI to drive better business results.
ChatGPT’s Challenges in Enterprise
We deeply believe in the potential of ChatGPT’s underlying technology to transform customer interactions. But According to Sam Altman, CEO of OpenAI:
“ChatGPT is incredibly limited, but good enough at some things to create a misleading impression of greatness.”
Our experience working with customers for the past four years suggests the following challenges and opportunities in Enterprise.
- Understanding the business’s own processes and internal knowledge.
ChatGPT is trained on open internet data. So it doesn’t deeply understand each business’s unique internal processes and knowledge. The AI would need to integrate into various systems of record, such as a company’s CRM, and build a unified view of products and customers. It should have the real-time intelligence of surfacing relevant information at the right moment.
- ChatGPT sometimes “hallucinates”.
It doesn’t always generate factually correct responses. Even more dangerously, it sounds confidently incorrect when doing so. Our experience with Generative AI suggests that this could break users’ trust.
- Integration with key enterprise workflows.
For Generative AI to drive real value for the enterprise, it shouldn’t stop just at generating responses. It should integrate into different workflows such as CRM entries, coaching software, order management, etc. In other words, it should drive actions across systems. Like how our no-code AI platform enables our models to connect with different systems. As a result, it can turn conversation insights into actions that drive key business outcomes.
- Continuous learning.
ChatGPT operates as a one-size-fits-all model. When the model makes a mistake, the only way to correct it would be to change the input text prompt. Our enterprise customers benefit from bespoke models which are finetuned based on user feedback.
Generative AI in Enterprise
For the past four years, we’ve been on an incredible journey to bring cutting-edge AI capabilities to our Fortune 500 customers: from understanding conversation insights to next-level virtual agents. Based on our experience, we think ChatGPT in its current form is not ready to be deployed in the enterprise.
At Cresta, our current team of industry experts is dedicated to producing future-forward and future-proof capabilities for the enterprise. Cresta was founded by Stanford and early OpenAI researchers, and many members of the team previously worked on Google’s generative dialogue system LaMDA.
Our comprehensive solution set leverages the same underlying technologies to help businesses drive better outcomes. The most forward-looking companies on the customer journey work with us to deploy virtual agents that overcome some of the challenges facing ChatGPT.
As the world of Generative AI experiences an increased pace of paradigm shift, we are committed to keeping our customers up to date with the most advanced technologies. So, stay tuned for exciting developments from Cresta as we continue to push the boundaries of AI in the enterprise.