Tyme’s AI Journey — Embracing AI Safely and Securely
The launch of ChatGPT in November 2022 put the world on notice that the AI revolution is here. As businesses try to navigate this transformative period and strive to find their AI strategy, most are left with more uncertainty than answers. While, many companies are embracing this new frontier, particularly the adoption of large language models (LLMs) like ChatGPT, there are those who have a different perspective. They are choosing to limit or restrict the use of LLMs like ChatGPT in a corporate environment.
Tyme’s Approach to AI Adoption
Tyme recognizes the potential productivity gains that can come from using large language models (LLMs) like ChatGPT. However, we also understand the importance of safeguarding sensitive data and having complete control over its usage.
To achieve this, we have chosen to use the paid version of ChatGPT, connecting to OpenAI through private APIs. This allows us to ring-fence and train on internal data, ensuring data privacy and security. By using a private portal, Tyme ensures seamless adoption by our employees, reducing the risk of using the public web version.
Training the ChatGPT Model
Now, we are at the stage where we train the ChatGPT model to understand and respond to questions unique to the context of Tyme, as well as address specific scenarios and use cases relevant to our business. The new chatbot is named Copilot — your AI assistant that handles the tedious tasks, allowing you the freedom to focus on thinking and higher-level work.
The training of the ChatGPT model can be divided into two phases: teaching Copilot and fine-tuning.
Teaching Copilot
To teach Copilot (teaching in the sense of “teaching” it knowledge), we leverage various company documents like our website content and customer service documents as our training corpus. The training corpus is parsed into smaller chunks of text stored in a database. We convert these chunks into special numerical representations called word embedding vectors.
We then score the similarity between each chunk and question by measuring how similar they are in meaning. Finally, we select only those chunks that have a similarity above a certain threshold as relevant context for ChatGPT to provide the most accurate answers. When a user inputs a prompt to Copilot, we use a technique called semantic search to retrieve relevant information from the vector database and use this information as context for Copilot to accurately respond to the user prompt.
Fine-tuning
To continuously improve the intelligence of Copilot, we have implemented a feedback mechanism. Each response generated by Copilot is programmatically evaluated using indicators that measure how effective the response is. This feedback is then used to refine Copilot’s future responses. Currently, there is human oversight involved in reviewing the responses. This ensures that Copilot is functioning correctly and not providing inaccurate or inappropriate answers. It’s important to note that any changes made to the model are permanent and cannot be reversed at this stage.
Domains and Use Cases
With Copilot, we adopt an approach that targets low-stakes applications combined with human oversight, focusing on areas ripe for disruption and having high potential for immediate productivity gains. In the early stages of our AI journey, we’ve identified specific areas where the integration of LLMs can make a significant impact. These are the domains where we anticipate the greatest value and progress:
Marketing — Copywriting and Content Creation
Copilot empowers our marketers to generate ideas and create content more efficiently. With Copilot, we can maintain a consistent brand voice, writing style, and format throughout our content. It also simplifies the process of personalizing messages for various customer segments, geographic locations, and demographics. Additionally, Copilot can serve as a valuable tool for generating social media responses, specifically for platforms like Twitter, making it easier for us to engage with our audience on social channels. It also provides substantial cost savings benefits as we no longer rely on third-party vendors for content creation.
Customer Service
Tyme leverages Copilot to power an in-app chatbot that provides instant and tailored responses to diverse customer queries, regardless of language or location. This enhances interaction quality and efficiency, automating a greater portion of inquiries, thus freeing up human agents for complex issues. In customer service call centers, every service representative is armed with their own Copilot but tailored with CS user-friendly interface, enabling them to assist customers more efficiently and resolve their issues during the initial call. As a result, response times are reduced, and the rate of resolving customer issues during the initial call is significantly increased. The knock-on effects of this will not only be happier customers but happier call service representatives too.
领英推荐
Internal Communications
Copilot will be available to our employees and will offer valuable improvements to our internal communications. It will be the engine that automates meeting notes, making the process streamlined and ensuring consistent documentation.
Additionally, it assists our employee engagement team with quick ideation and content creation. We can use this to streamline new employees onboarding process and facilitate a faster and smoother transition into their roles. Copilot also acts as a language translation tool, fostering better understanding of company-wide communications among our diverse workforce whose first language may not be English. With Copilot, our internal communications become more effective, and inclusive.
Other AI Initiatives
Apart from the integration of LLMs like Copilot, Tyme is exploring AI-powered initiatives in different domains.
Software Engineering
Software engineering presents an incredible opportunity to leverage AI and amplify developer productivity. It’s worth noting that we separated this domain from the others above, as this falls outside the scope of our utilization of ChatGPT. Our objective centers around facilitating swift task completion, while simultaneously preserving developers’ mental energy, enabling them to focus on more gratifying work and ultimately finding increased enjoyment in the coding process.
The most recognized AI powered developer platform in the market today is GitHub’s Copilot (note: not to be confused with Tyme’s chatbot also named Copilot). Github’s Copilot already demonstrates remarkable productivity gains. Notably, in a study conducted by Github, developers utilizing GitHub Copilot accomplished tasks at a staggering pace — 55% faster compared to those without Copilot. Furthermore, the Copilot users exhibited a higher task completion rate of 78%, surpassing the 70% rate of the non-Copilot group.
At this stage, we have decided not to move forward with Github’s Copilot as there are concerns around sharing of source code data. Instead, we have chosen Codeium, which meets our expectations of an AI-powered developer platform that can produce the desired productivity gains mentioned above and the assurance of having complete control over our source code.
Bespoke Models
While the focus of this article has primarily been on AI and its impact on productivity-driven use cases, we are also leveraging AI for business-focused use cases. As a cloud-native digital bank, we utilize our proprietary AI framework to develop bespoke models such as credit scoring models, anti-money laundering (AML) models, user behavioral models and predictive analytics model. By enhancing our understanding of customer behavior and preferences we can apply these insights to optimize conversions and provide a more tailored experience for our customers.
Conclusion
As of the writing of this article, we are still in the early days of AI adoption, with just a few months under our belt. During this time, we have observed promising results in certain use cases, particularly in Marketing. The success of AI adoption depends on various factors, including team size, standardized playbooks, training requirements, and the willingness of the team to embrace new technology. Our approach emphasizes the importance of safeguarding data while leveraging the potential productivity gains of LLMs like Copilot.
Written by Minh Le - General Director, Tyme Vietnam
For more TymeLabs - Articles: