7 min read
Have you ever seen the film “Imitation Game”? It’s the story of how Alan Turing, a brilliant mathematician, contributed to saving millions of lives by cracking the enigma code during the Second World War leveraging his Turing Machine (TM).
Regardless of this huge achievement, since her birth in 1936, the TM was perceived far from being intelligent and its model of computation turns pale in front of the most recent AI solutions like Natural Language Processes (NLP) and Natural Language Understanding (NLU) techniques, Convolutional Neural Network (CNN), Deep-learning algorithms, etc.
Yet, here at Spixii, we have never been ashamed to say that our Conversational Process Automation (CPA) platform and our insurance chatbots are essentially based on the same foundation of the TM.
This might indeed be a shocking statement for an insurtech company working within the Artificial Intelligence domain, therefore we feel compelled to challenge our industry and justify our upstream perspective.
Is it necessary for chatbot conversations to employ the most up-to-date technologies? How do we define the intelligence of a chatbot?
Let’s proceed from the ground up.
How TM and Spixii CPA are similar
By definition, a TM is a mathematical tool based on
- A limited number of states (state 1, state 2, … state n)
- A writable tape in which every cell contains a symbol fetched from a finite set of symbols (for example, the alphabet of Latin characters or the integer numbers between 0 and 9)
- A deterministic transactional function represented with a matrix that, given the current state of the machine and a symbol read in the current position of the tape, returns a new state, writes a symbol in the pointed cell, and moves to the following or the previous cell on the tape.
Turing Machine with Tape Erasure (source)
Spixii CPA is a powerful platform that allows insurance businesses to map a process into an automated conversation with the end customer. For example, a First Notification Of Loss (FNOL) claims process can be mapped into a friendly chatbot conversation with a policyholder.
Expanding this specific example, we can see how an automated conversation is fulfilling the same requirements of a TM, in particular:
- The states of the machine are the states of the conversational process (eg: identification of the customer, acquisition of the claim’s details, validation of the information, submission of the claim, etc …)
- The writable tape is the chat itself, where every cell is a response either of the machine or of the policyholder
- The transactional matrix is the set of rules that, given the current state of the process and the response of the policyholder, gives back to him/her a new message and moves the machine forward to the next state.
In addition, there is a special state in the TM flagged as the initial state, and at least one state flagged as final, where such states are reached and the computation ends. A conversation follows the same schema: the beginning of a conversation with the first chatbot message and the end with the final message (for example, the claim has been accepted or rejected).
Although Alan Turing centered the definition of Artificial Intelligence upon the ability for a machine to exhibit intelligent behaviour equivalent to, or indistinguishable from, that of a human (the famous Turing test), in a very broad sense we can say AI is about machines which act intelligently.
Now, how can we measure the intelligence of TM and CPA?
The intelligence of a conversational process
The TM is a tool to solve mathematical problems, therefore the definition of intelligence might be tricky as even today math holds many uncovered mysteries.
Generally speaking, a mathematical tool like the TM is considered intelligent if it can solve a number of problem’s categories (for math geeks, we should break down the P, NP, NP-complete and NP-hard classes of the computational complex theory, but it will take us a bit too far).
When it comes to a conversational process, the definition of intelligence is much simpler and its performance easier to measure. A meaningful chatbot conversation is an exchange of information through messages that takes the customer from A to B, or from a specific customer intent to its achievement, therefore a conversational process is intelligent if:
- no conversation is lost between A and B (i.e. uncompleted conversations)
- the number of steps from A to B is minimal (the conversation is effective and sharp)
In other words, a TM (and CPA by extension) is judged on the number of problems that it’s able to solve. The complexity is calculated by answering the question: how similar is this problem to another problem of which the complexity is known? Or within a conversational process: how similar is this customer query to another query of which the complexity is known? This analysis is called the Polynomial and Exponential-time Reduction problem.
If the conversational process were to happen between two machines, its performance would be judged by the number of steps to complete a journey from A to B and the number of times that one of the two machines “gives up” before reaching the status B.
Now, in our specific example, where the process involves a human and a machine within a particular industry, like insurance, the definition of machine intelligence needs to be extended to accommodate a set of human requirements and some industry-specific expertise.
Tuning the machine to enhance human interaction within the insurance domain
Let’s go through all the additional requirements that make the machine intelligent within an insurance context.
1. Outstanding Customer Experience (CX)
Even though the conversation is happening between a human and a machine, we aim to keep it always meaningful, personal and pleasant. Meaningful, because we strive to serve the customer by solving his/her query, not to do small talks. Personal, because we want to always be relevant and contextual and shape the conversation according to the person we are talking to. Pleasant, because we try to move away from clunky and mechanical interactions such as the webforms one. Through a chatbot conversation, we can easily convey the brand personality and identity through a tailored interface and tap emotional power through a considered tone of voice.
All of the above factors can be measured with verbatim feedback and Transactional Net Promoter Score (TNPS) analysis. Spixii customer success stories have extensively demonstrated that chatbots have a much higher collection rate compared to other traditional channels.
Needless to say, the conversation needs to always be available, and again chatbots rise to the challenge with 365/7/24 coverage.
2. Secure and compliant conversations
In the Insurance industry, compliance and data protection play a central role. While a TM doesn’t have any requirements for security and data protection per se, a conversational process has to guarantee the confidentiality of all the information shared, especially when it comes to health insurance.
For insurers, it is key to ensure transparency of the processes whilst protecting policyholders’ data. Mapping out the conversational process in a visual way, clean from data identifiable to an individual or entity becomes an essential feature.
With compliant conversations, we mean the ability to gauge the tone of voice according to the brand guidelines, but also to carefully tune it according to the context. For example, in delicate situations like claims FNOL or funeral insurance, the customer will demand further empathy and professionalism.
3. Informative auditing and analysis
Related to compliance, it’s important that the CPA system learns common patterns, highlights critical conversational dropouts and points of friction to facilitate the generation of insights and suggests contextual optimisations.
This is where auditing and analytic tools come into place: to foster continuous improvement the CPA looks for ways to reduce drops, cover more conversational paths and, back to the concept of “states”, reduce the number of steps to take the end customer from A to B.
4. Easy configuration and maintenance of the conversational process
A huge differential for a chatbot conversation over a website is its dynamism, as it is an organic and continuously improving solution. In order to make the conversational process easy to configure in the first place and maintain after, it is better to have a platform that allows its orchestration with no-code or low-code interface.
In other words, conversational processes easy to configure by non-tech profiles through an intuitive interface means higher speed in improvements and lower costs.
Quicker changes to the conversational flow combined with instant analytics set the basis for a continuous improvement loop fueled by real data. As a side effect, it also generates an evidence-based centre of excellence that increases the intelligence of the system within an iterative environment.
To summarise, a good insurance conversational process machine needs to:
- Optimise the number of steps to complete the process itself
- Have a low rate of uncompleted conversations
- Have an outstanding CX at every level, from the conversational process to the interface and tone of voice
- High-secure technological stack built within a strong data protection framework
- Enable continuous improvement through collection of new conversations paired with insightful analytic tools
- High-level configuration, accessible by non-technical profiles
Under this new light, we move the focus from Artificial Intelligence itself, where huge emphasis is on the machine, to the intelligence of the conversational process. We believe business intelligence is achieved with a smart process orchestrated by humans and augmented by machines.
In the same words of B. F. Skinner:
The real problem is not whether machines think but whether men do.
Therefore to insurers, rather than looking out for the most AI-powered solution, we suggest focusing on the conversational process first and then try to see which technology best fits their business needs.
It is all about applying the right technology and techniques to the relevant industry.
