KeyReply

Relevant Industry Language Model to be Enhanced with Company-Specific Vocabulary and Jargon

• Recall that one of the AI techniques that make conversational agents intelligent is Natural Language Processing. The sophistication of the NLP models used in the bot makes a big impact on how smart the conversational AI can be.  
• Not all NLP models will be able to act on users’ queries immediately out of the box. In fact, the cold start problem is prevalent, where there is a lack of clean and labelled data to start the training process.   
• In addition, every industry has its own jargon. The most frequently used phrases in finance and banking for instance, will be drastically different from the complex medical terminology in the healthcare industry. Moreover, the vocabulary adopted to respond to customers questions will also be different from company to company. Think how the same product, say a healthcare screening package, can be termed by different private hospitals in different ways. 
• The NLP models therefore need to be pre-trained with industry knowledge bases to a certain extent. It should then be able to ingest a company’s vocabulary through historical chat logs, product or service collaterals and internal documents. 

Computing Hardware and R&D

• Even the best NLP models can fall short if you do not have the right hardware do to specialised training. The amount of computing power used in AI training has been increasing exponentially. OpenAI, a nonprofit research organisation showed that this compute power doubles every 3.4 months.
• Companies planning to adopt AI technologies should look at the long-term implications of this trend. Developing industry ready AI solutions require a lot of experimentation and tweaks to models to achieve the desired level of performance. Many research papers describe experiments that yield good results within a narrow scope and will require a substantial amount of work before it can be applied to solve real world scenarios. 

Dialogue Scripting for Empathy and a Smooth User Experience

• While the NLP models influence how efficiently the bot responds to user queries, a good conversation requires much more. Consider the below interaction between an anxious user and a bot.  

Multi-language Capabilities

• The target audience for any business solution is never a monolith. In today’s globalised multicultural world, customers can vary in their demographics and language preferences. This is especially true if an enterprise is moving into or has its eyes set on new markets. In these cases, the target audience may use a language that is a mix of English and local dialects, or a different language altogether.  
• For example, a U.S firm that wants to expand into Southeast Asia will have to deal with customers who tend to speak Singlish or Manglish, or who prefer Malay or Thai). The conversational AI will need to be trained with data in their target languages to be able to respond to their queries. Machine translation currently may not deliver the level of performance that companies require. 
• English being a high-resource language has large quantities of publicly available data sets that can be used for training. This may not be the case for the local languages. Look for vendors with proprietary NLP models which can handle mixed languages, and who have gathered a large corpus of native language datasets over time to train their language engines.  

Beyond Keywords to Contextual Understanding

• The simplest chatbots rely on keyword matching to extract responses to questions. The user asks a question with certain keywords and the bot looks for literal match in a database to return the corresponding response. The best conversational AI solutions must go beyond that.  
• Specifically, they should be able to perform a semantic search. Semantic search involves understanding the user’s intent and contextual meaning of terms as they appear in the database. For example, consider the following exchange. 

A Unified Experience for Multiple Stakeholders

• Companies today have to manage many more stakeholders than before. Single use case bots that only do one thing or serve one customer type can very easily hit a ceiling for large enterprises. They are also harder to maintain over time. Hence, it is important for the platform to be able to manage multiple bots, and know when to send specific information to each user. 
• In some cases, different users might ask the same question but the responses often vary depending on their characteristics and personal information. For example, members of three different departments in an organisation could ask the same question about the benefits they are entitled to. But the answer could vary based on their pay grade, department, job function and type of role.

Ease of Training by Users

• One of the golden rules of testing a conversational AI solution is to approach it as an ongoing process rather than a one-time activity. This means regular monitoring, review, and update of the training data to improve the bot’s performance.  
• This ongoing training during testing must involve the key stakeholders in the organisation, especially those who are closest to customers. Based on their input, admins will have to upload, edit, and manage intents and the content. Conversational AI solutions should include such an element of flexibility to allow the users to train the bot based on their subject matter knowledge. 
• If each change takes a lot of time or if the company needs to rely on the vendor for all types of changes, it will not be efficient. Businesses today need to be nimble and be ready to update their communication channels often. The interface needs to be easy for non technical users to use. The best software interfaces are those that require minimal training. 

Integration with CRM, Knowledge Base Systems or Backend Systems

• A conversational AI solution that is a standalone system is hardly better than a rule-based bot. If the solution only has access to the user queries and a static database of pre-written answers, it will not get far.  
• Instead, integration into other internal systems like CRM databases and knowledge bases can take the chatbot to a much higher level of utility. In healthcare institutions for example, integrating a conversational agent to the electronic medical records (EMR) allows it to personalise responses based on the users’ medical history, previous treatment, allergies and more.  
• In an ideal world, the conversational AI system will be integrated to all other digital systems. But this may not be feasible, at least in the initial stages of deployment. IT teams have to work with the bot development team to figure out which of they internal systems are open and ready for integration.  

Customisable Analytics for Decision Making

• In the ongoing testing of the bot, the goal should be to iteratively improve performance. Performance can be gauged differently by different functions within an organisation. For example, the customer service team might prioritise customer satisfaction above all else. The marketing and sales functions will be interested in the total leads generated. The operations team on the other hand will want to know the peak hours of the day so they can coordinate resources accordingly. The conversational AI system should be able to provide insights that can inform each of their decisions.  
• To this end, the system should allow admin users should be able to monitor the bot usage and activity. They should be able to study analytics for different time periods to make business decisions. Common analytics metrics that are relevant to track include user statistics and profile, customer satisfaction, most frequently asked questions, and percentage split of answerable to unanswerable questions. A certain level of flexibility should also be available, in order to cater to the changing goals within an organisation. This means giving the organisations the option to customise  the analytics to decide what metrics they want to monitor. 

Seamless Live Agent Takeover

• There may always be scenarios where the conversional AI is unable to respond to a user query effectively. The bot might be in its early days of training, the question may be out of scope or there may not be an easy answer available in the database of responses. In other cases, the customer may simple request to speak to a live agent.  
• In such cases the most appropriate thing to do would be to hand over to human to continue the conversation. The best conversational AI solutions should have this fall-back option available, to seamlessly transfer from the automated chat to a human agent. This means being able to find the right human agent with the skills relevant to the customer’s question and to allocate resources based on the agents’ availability.