The email challenge: easing the pain with artificial intelligence

Artificial intelligence (AI) and machine-learning are talked about a lot in the media and in business in terms of their potential and benefits. But for many people, there is a certain hesitancy about AI, and an element of the unknown. Will it really do what it is promising? Will it take away jobs? Who takes accountability? And there is often simply a lack of understanding of what it can actually do and where it can help.

So when it comes to showcasing what AI can achieve, it is important to shed light onto the magic of AI, helping to demystify it, and increase the awareness and understanding of what it can and cannot do. We also need to make it very clear, in plain English, exactly what the tangible impact/benefits will look like. And ultimately ensure that it creates some level of comfort and understanding and, with that, a desire to develop it further into broader areas.

There is also the issue of scalability, with a desire to ultimately be able to apply the technology to as many business units and teams as possible, creating the best economies of scale.

One way of tackling such projects is to take a ‘small and fast’ approach – a case that addresses an issue that virtually every unit faces, built in the simplest and fastest way possible, beginning with small teams and then scaling it to as many viable units in the company as possible.

As Armin Schaefer, head of digital and new technology commercial insurance international programmes, explains: “It is about making the pieces smaller, as well as faster, to be able to show what it does, the impact it can have, and get a tangible understanding of it, in order to allay some of the fears and concerns around new technology such as AI. An important element is the concept of ‘seeing is believing’ – a project like this is often a step into the unknown and once teams start to see the practicalities and benefits of the project, then trust starts to build and it can be rolled out further.”

Taking the pain out of handling unstructured communication

AI-driven automation can have an immediate impact by taking on non-value-added time-consuming administrative processes and automating them, thus speeding up the whole end-to-end experience.

This generally involves taking unstructured data and turning this into a structured form that can be processed more easily by humans or systems. One particular area where teams across the organization have an insurmountable amount of unstructured data that they need to pre-process into a structured form: emails.

For example, while in commercial insurance digital marketplace and e-placement models are clearly a market trend, the reality today is that the vast majority of service requests are received almost exclusively as hundreds of thousands of emails with unstructured data.

Tobias Wild, lead digital business analyst at digital and new technology, explains: “The email challenge is across many different aspects of our business – it is in the pre-buying scenario where intermediaries send email submissions with multiple attachments containing various pieces of information that need to be processed in order to be able to quote, through to the servicing aspect once we have bound the business, when we need to respond to a claim or make a change to a policy or international programme structure.”

Dealing with emails and long attachments is undoubtedly time-consuming and frustrating for teams, with manual pre-processing required, often duplication of effort, and a painful experience for the carrier, customers and brokers with regard to speed, transparency and ease of doing business.

Against this backdrop, the automated processing of emails using AI and machine-learning is clearly a relevant use case with high impact potential. It is a simple AI case with a pain point that is felt across all units on a global basis, and provides the opportunity to build a globally reusable solution module.

“Taking away administrative work that no one really likes to do helps to make the project an easier ‘sell’ to employees,” says Gero Gunkel, chief operating officer, ZACM. “People will welcome something that really makes life easier for them and removes the time-consuming and frustrating type of work.”

Solving the email challenge

Within the commercial insurance unit at Zurich, an automated intake solution has been developed. The aim of the technology is to automatically read and process unstructured emails and attachments. The algorithms turn the unstructured data into a structured format to enable a (semi-)automated case completion, speeding up the process for the underwriting team and ultimately for the customer.

The process has two stages, according to Mr Gunkel: “First, we feed the algorithm with lots of documents – a dictionary, novels, etc – so that the machine understands sentences and word order and how verbs and nouns relate to each other, and then it is fine-tuned, taking emails from the actual business process so that the machine can learn the task at hand and picks up the language used in business and the insurance jargon.

“The machine uses a traffic-light system that classifies the output and indicates to what degree it believes it has all that it requires,” continues Mr Gunkel. Green means the machine is very confident and has found all the information it needs to process the email and extract relevant information for downstream processing in various systems. Amber means the machine is not very sure what the email is about but has found the relevant information, so a human will interact and confirm whether this is the correct procedure or not. And red is where something is missing or where it is necessary to ask the sender of the email to provide more information.

In addition, Mr Schaefer explains: “There is also a feedback loop mechanism that allows the human to challenge the output and throw it back to the machine, which will then learn for the future to do it better next time around. The machine indicates how good it thinks it has done its job and the human reconfirms or rectifies, feeding input back into the machine. This is what is called ‘supervised learning’.”

Since the implementation of the solution, the feedback from users has been excellent. They have a machine that picks up about 65% of cases in an automated manner, with the rest requiring confirmation or suggestion of a change. Most importantly, the machine has performed at least as well as a human, perhaps better, with an overall accuracy rate close to 98%.

Originally published on Commercial Risk Online on September 25 2020