Conversational AI: Its potential, challenges, and impact on CX

Flitto DataLab
Published in
12 min readMar 22


Part 1 of The European Chatbot and Conversational AI Summit 2023–3rd Edition series

Image source:

In today’s fast-paced world of technology, chatbots and virtual assistants have become essential tools for businesses looking to improve customer service and streamline operations.

Flitto DataLab had the honor of participating as a Gold Partner and engaging in the latest conversation on chatbot and conversational AI technology at the European Chatbot and Conversational AI Summit — 3rd Edition which commenced last week in Edinburgh, Scotland.

Experts who spoke at the event were more than just aware of the current trend. Valuable insights were shared at the summit, and Flitto DataLab would like to briefly share some of the highlights of the conference sessions.

ChatGPT — Strengths, Weaknesses, Opportunities & Threats

Presenter: Cobus Greyling (Chief Evangelist at HumanFirst)

ChatGPT and Large Language Models (LLMs) are becoming increasingly popular for their ability to understand and generate human language. These models have various strengths that make them useful for a wide range of applications.

Image source:

Strengths: Multilingualism, Generative Power, and More

Cobus Greyling began his presentation by listing the strengths of LLMs, including their existing default base knowledge covering a wide range of topics. Additionally, these models are almost always multilingual, which allows for fast and efficient translation and makes them accessible to users across different languages.

Another strength includes the generative power of prompt engineering, a useful tool that can guide the model to generate text that is more concise and coherent.

LLMs also have the capacity to train themselves in real-time. A few examples include its zero or few shot capabilities, which refer to the ability to perform a specific task without being explicitly trained. This allows for greater flexibility and adaptability. Dialog fall-back, which refers to the ability to provide fallback responses when the model is unable to understand or generate a response to a user’s query, is also useful in tasks such as question-answering and information retrieval.

Another strength of LLMs is the ability to leverage synthetic training data for chatbots. While it may not be a foolproof solution, this synthetic data can be artificially generated based on various scenarios, user intents, and conversational patterns, enabling the chatbots to learn how to respond to a wide range of inputs and contexts.

Opportunities: Versatility for Research, Innovation, and Product Development

LLMs provide several opportunities for their continued research and development. As more data becomes available and more sophisticated algorithms are developed, LLMs’ accuracy, speed, and range of applications can continue to evolve to provide better performance and more reliable results over time.

With their impressive language capabilities, LLMs can be leveraged to quickly and efficiently create chatbots that can handle a wide variety of customer inquiries and provide personalized responses. They also provide a foundation for innovating new products and services. With the release of GPT-3, we could see the rise of a range of apps that leverage the power of this model to create innovation and value.

Another area boosted by the LLMs is prompt engineering, which has been giving opportunities to a lot of developers and companies alike to create apps that facilitate the interface between humans and models.

Threats: Data Governance Challenges and Language Availability

However, there are also some threats associated with LLMs. Availability of minority human languages is a threat to the well-functioning of such models, as there still exists room for improvement to train them for accuracy.

Another threat is data governance challenges. LLMs may contain data protected under another company’s governance policies, and the display of such data to the end-user can be problematic for the LLM provider. Various policies, including personal identifible information (PII) policies, must be considered to avoid unwanted consequences.

Cobus Greyling on the threats for ChatGPT and LLMs

Weaknesses: Rigorous Preparation Required for Enterprise Usage

Greyling points out that the default state of LLMs is not suitable for enterprise usage. In line with the threats mentioned above, fine-tuning must be done to duly prepare the model for appropriate domains.

Currently, fine-tuning LLMs requires significant technical expertise and access to large amounts of high-quality training data, making it a challenging task for most developers and businesses. Moreover, the lack of NLU and NLG tools further hinder the data preparation process.

The development of better no-code and low-code fine-tuning tools could be a solution to enable more innovative applications that can improve efficiency and productivity in various fields.

For the time being, continuous monitoring of user inputs as well as appropriate scaling of the models would be crucial in maintaining the quality of services provided by LLMs.

Voice-first Multilingual Bots

Presenter: Melissa Samson (Data Analyst Manager at Cisco Systems — Webex), Kelsey Kraus (Data Scientist at Cisco Systems — Webex)

In this session, it was pointed out that despite the widespread adoption of conversational AI, the majority of chatbots have been optimized for English-speaking users. Even the most advanced language models, such as ChatGPT based on GPT-3.5, which supports over 95 languages, are often trained on American-centered English, making it difficult for users from different linguistic and cultural backgrounds to engage with these technologies.

Challenges in Developing Multilingual Virtual Assistants: Linguistic and Cultural Nuances

Melissa Samson, a data analyst manager at Cisco Systems, highlighted a startling fact — only 26% of web users prefer to navigate in English. This stands in stark contrast to the emphasis placed on English-language training by many companies developing chatbots. Additionally, Samson noted that English-speaking countries only account for 20% of the global economy, revealing the untapped potential for chatbots that can communicate effectively in a variety of languages.

Melissa Samson presenting how multilingual agents can expand access

However, developing a multilingual virtual assistant is not as simple as just translating text. Different languages around the world have unique written representations of words and phrases. These idiosyncrasies must be considered when designing a virtual assistant. For example, German words can often be longer than their English equivalents, and Arabic is written from right to left, which is opposite to English and most European languages. Therefore, adapting a virtual assistant to different languages and cultures requires a deep understanding of linguistic and cultural nuances. The input layer of a multilingual virtual assistant would need to be modified to accommodate different word lengths and directionality to accurately represent the written language.

Samson also pointed out that training the model based only on written documents would not be representative enough to create a sound user experience for all speakers of the language. The model should account for the way different parts of the world speak a language and collect data to track this spectrum of divergences. Incorporating these linguistic and cultural nuances is essential to ensure the virtual assistant accurately understands and responds to users in a natural and authentic way, which is crucial for creating a positive user experience.

Inherent Linguistic Challenges

Many languages have inherent grammatical features that can pose significant challenges to the development of conversational agents. One of the most notable challenges is gender, a common feature in many languages. It can be challenging for developers to formulate a model that respects all users when it comes to this feature.

Another issue is formality. For example, in French, the use of different words and grammatical structures is necessary to convey formality. Deciding how the conversational agent should behave towards the user in these situations is not an easy task, as user satisfaction will depend on personal preferences. These challenges highlight the importance of considering cultural and linguistic diversity in conversational agent development.

Kelsey Kraus, a data scientist also working at Cisco Systems, shed light on another significant problem to take into account: spelling variations across different countries that use the same language. Some languages have specific academic rules that were crafted to help clarify the transposition from spoken to written language. One such language, French, had a spelling reform introduced by the Académie Française in 1990. However, not every country adheres strictly to these rules; and therefore, different forms of spelling arise among many French-speaking countries.

Resource and Technical Challenges

In addition to linguistic and cultural nuances, developing multilingual bots also poses technical challenges. One main issue is the availability of data resources. Labeled data is abundant for only a small number of languages such as English, German, Spanish, Japanese, and so on, while minority languages tend to be very difficult to acquire data from. This means that people who speak languages outside of this spectrum may not be receiving the same quality of service.

Kelsey Kraus on the technical challenges of multilingual bot development

Another technical challenge is dialogue management. This pertains to a tool to ensure that the bots, when in a dialogue with the user, use the right tone, have a nice voice quality, respond with regularity and accuracy, and provide appropriate responses for the user’s social and cultural perspective.

Developing a multilingual dialogue management system is complex, as it requires not only a deep understanding of the linguistic and cultural nuances of different languages, but also advanced technical skills to ensure the system can handle different input formats, languages, and dialects.

Charismatic AI — How to Turn Raw AI Potential Into Customer Excitement

Presenter: Christoph Esslinger (Founder and Managing Director at

As AI technology advances, more and more companies are incorporating automated customer service chatbots into their business strategies. However, the challenge lies in making these chatbots not only functional but also charismatic to attract and retain customers. In this section, we will explore the key points discussed by Christoph Esslinger on how to turn raw AI potential into customer excitement.

Charismatic AI — How to turn raw AI potential into customer excitement

Why Charismatic AI Matters

According to Esslinger, having a charismatic chatbot is crucial because people remember extraordinary experiences. A chatbot with good service and charisma can leave a lasting impression on customers, making them more likely to recommend the service to others. In contrast, an impersonal and frustrating automated system can quickly damage a company’s reputation.

Avoid Scaling Frustration

To Esslinger, one of the biggest mistakes companies make when implementing automated customer service is cutting corners to reduce costs. This can result in a frustrating chatbot experience for customers, leading to negative reviews and a damaged reputation. Therefore, it is important to invest in creating a flawless chatbot service that provides value to customers.

Useful Tools for State-of-the-Art Automated Customer Service

To create a state-of-the-art charismatic chatbot, one of the things companies should consider is adding multiple accessible languages to the bot. Esslinger points out how frustrating it can be to use the regional services in a country where we do not speak the language. Adding more languages to the bot will make it easier for customers to communicate with it and increase satisfaction. Additionally, companies should focus on creating a flawless service to avoid simple questions that can be easily answered by the automated system.

Designing a Charismatic Automated Chatbot

To design a charismatic chatbot, it is essential to understand the foundational pillars of human-to-human charisma. According to Olivia Fox Cabane in her book “The Charisma Myth,” these pillars are power, presence, and warmth. In the context of AI customer service design, the most important pillar is warmth. A charismatic chatbot should make the customer feel valued and put their needs at the center of its attention. As chatbots are Machine Learning algorithms trained on human text, a technical approach would be to bias the prediction process toward more friendly, correct, and concise responses.

Human Substitution?

The possibility of AI replacing human jobs has been a topic of concern for many years. Esslinger believes that AI will replace humans in jobs that require thinking, while humans will engage more in jobs that require emotional intelligence. Therefore, humans should view this as a change of working area and mindset rather than as a replacement.

Designing a charismatic automated chatbot is crucial for attracting and retaining customers. By investing in a flawless service that prioritizes customer needs, adding accessible languages, and focusing on warmth, companies can create a service that provides an extraordinary experience. Additionally, it is important to view the integration of AI in the workforce as a change rather than a replacement.

The Future of Conversational AI: Exploring Ethical and Philosophical Questions

Presenter: Sonia Talati (Senior Manager of Conversation Design at GoDaddy)

With the release of large linguistic models such as GPT-3, the possibilities for conversational AI have expanded beyond consumer-facing chatbots and business tools. However, this new horizon also raises important ethical and philosophical questions about what defines a human being and how much human-likeness is acceptable in a robot.

Image source:

Sonia Talati, a senior manager of conversational design at GoDaddy, believes that we need to have a clear understanding of what distinguishes humans from machines to draw the line when developing new technologies. As machines become more humanized, we risk losing touch with our origins, manners, and nature in general. To illustrate this point, Talati asked participants to choose which of three robots they found least intimidating: a normal chatbot assistant, a robot in human form, or a less humanized robot with greater thinking and response capacity. The vast majority chose the chatbot (first option), suggesting that human beings are very visually guided and uncomfortable with robots that look too human.

However, Talati also asked participants to predict which of the three robots people would become used to in the next three to eight years, and the winner was the robot that was slightly less humanized but more human than the chatbot. This suggests that there is an acceptable range of human-likeness that robots can exhibit without becoming too intimidating or disturbing.

Slide visualizing the three options of robots

Ethical Questions Raised by Conversational AI

The rise of conversational AI, powered by large linguistic models like GPT-3, presents numerous practical applications and opportunities for human advancement. These models represent the collective intelligence of human beings, with vast amounts of human perspectives and insights fed into the training datasets. As a result, these models can collaborate with humans to solve problems and advance technology in unprecedented ways.

The definition of the concept of Human Colossus

However, with this technology comes a series of ethical questions that need to be addressed. One of the most pressing concerns is whether we should be comfortable with AI-generated text that mimics human conversation so closely that it becomes difficult to distinguish from human-generated text. This raises important questions about the use of AI-generated text in human-to-human contexts and the potential consequences of blurring the lines between human- and machine-generated language.

Ultimately, the answer to these ethical questions will depend on the context and the intended use of AI-generated text. As such, it is crucial to have a broader discussion about the implications of this technology and how we can responsibly use it to benefit humanity as a whole.

The Future of AI Development: Machines That Can Think?

As AI continues to push forward, we are already seeing the first steps of its development: training AIs to create what humans usually do, such as generating text, images, and sounds. According to Talati, “The next step will be to train AIs to ask other models to create what humans usually do, and the final step will be to have AIs come up with the ideas of what to create and why.” This would complete the AI development circle, giving rise to machines that could literally think.

To sum up, the future of conversational AI raises important questions about what defines a human being, how much human-likeness is acceptable in a robot, and the role of AI in human society. As AI technology continues to advance, it is important to consider these questions and to develop ethical frameworks that can guide the responsible development and use of these powerful technologies.

Coming up next:

In this post, we have covered few among many notable sessions that took place at the summit.

Flitto DataLab was able to engage with industry leaders and technology enthusiasts at the summit. Needless to say, we are more than enthused to have participated in the event, and we are looking forward to more sessions to come.

Coming up next are highlights on real-life enterprise application scenarios for conversational AIs and panel discussions with an amazing line-up of participants.



Flitto DataLab

Multilingual Data for AI & Integrated Translation Platform