at Boost Holdings Sdn Bhd

ChatGPT acquired one million users just five days after launching in November 2022. By comparison, it took Instagram approximately 2.5 months, and Netflix had to wait about 3.5 years to reach 1 million users. ChatGPT's euphoria was unlike any other AI models (even other applications) we have ever seen. Besides chatbots, various GPT models were introduced for image and video creations, benefiting individual users, freelancers, and small businesses. Various articles mentioned that Generative AI could help to grow the business exponentially and increase productivity. Suddenly, everyone hopped onto the bandwagon, fearing missing out (FOMO).

From a corporate organization perspective, the biggest impact of the GPT model is that it has democratized chatbots, where conversational chatbots can now be built faster with a much smaller team with lesser skill sets. Due to security and privacy concerns, most organizations started off with building and deploying a knowledge-based chatbot for internal users. Only after that are they considering extending the knowledge-based chatbot to external users to serve their customers with any queries about the company and products. However, most are still stuck on internal use cases.

Based on feedback from various organisations, after exploring various Generative AI use cases for almost a year, they are now faced with the familiar questions about all AI or Machine Learning models that they have ever deployed, which were “What is the impact on the top-line or bottom-line of the company?” or “Is it bringing any needle-moving impact?” Suddenly Generative AI is at a crossroad, needing to justify its existence in the organization.

Still, Internal Use-Case Focused

Due to the strict regulatory and data security requirements, the safest and fastest way is to build a knowledge-based GPT model and keep it for internal users, not limited to just chatbot, but also for marketing copywriting and policy summarizer. That way, integration to the user database is avoided, and sensitive data is secured. Problem solved, right? Not really. We must start feeding the GPT models with user data without compromising data security to expand to more business-impacting use cases.

Intangible Business Impact

Since the use case is still internal, the tangible impact is somewhat minuscule. It can only provide theoretical cost savings but not a clear impact on the bottom line. For example, an HR chatbot is positioned to save manpower from manually replying to employees’ email queries, but this does not reflect the actual manpower cost reduction. Most importantly, if it does not help increase sales or drive product take-up, the impact will not be seen as needle-moving.

Hallucinations

One of the biggest concerns about extending Generative AI to external users is hallucinations. This is why the use cases are still internal focused, because hallucinations can be damaging to external users. Using Generative AI LLM to power customer service chatbots servicing external customers is risky if guardrails are not built. Even with the latest LLM, the probabilistic chances of hallucinations, though small, can still damage a brand with unexpected responses.

"Cloud service providers are driving hard to rival Microsoft Co-pilot capabilities because I believe Co-pilot will have a similar impact to what ChatGPT did to the masses, this time making AI accessible to the corporate population"

I still believe Generative AI is not just hype. Perhaps, if you are at this exploration stage, you could focus on a few things to keep Generative AI use cases alive and relevant to the organization.

The Second Coming of Generative AI

Cloud service providers are driving hard to rival Microsoft Co-pilot capabilities because I believe Co-pilot will have a similar impact to what ChatGPT did to the masses, this time making AI accessible to the corporate population. Imagine that slide makers who spend days preparing for a presentation deck and MS Excel experts who created numerous formulas linking various sources can now complete their tasks in simple English by just using natural language. Not only will the decks be ready, or formulas will be produced automatically, but the relevant data will also be readily extracted (without having to ask the poor data analysts to retrieve them). Is this solution ready? Not unless the legal and security issues are addressed and, most importantly, the datasets in the data lake are properly set up to provide seamless data retrieval.

Taking a Leap of Faith

Hallucinations may not be fully resolved, but various mitigation measures are being deployed to further minimize the occurrence, such as narrowing the context using a semantic router or enhancing the language translators. Most importantly, it is to manage expectations of senior leadership stakeholders of the potential shortcomings. Data security and privacy must be addressed without compromising the data so that use cases involving database integration can start

Is the honeymoon period over for Generative AI? Maybe not yet…