Max Life Insurance coverage has an energetic base of 4 million clients with ~ 8 lakh new insurance policies offered annually and the shopper base is rising at a excessive fee of ~ 20% yr over yr. A lot of our energetic clients write to us within the type of emails on subjects reminiscent of contract-related inquiries, coverage clarifications, coverage issuance standing, delivery standing, updates of coverage data and new service requests. On common, Max Life receives round 60,000 emails per thirty days and these are answered by a devoted electronic mail response workforce (electronic mail desk) that’s a part of the customer support division.

With a big quantity of electronic mail site visitors, the present electronic mail response workforce requires a workforce of 50-75 devoted individuals who evaluate, queue, and reply to emails based mostly on the guide evaluate of every electronic mail. Moreover, the response course of itself turns into advanced, requiring queuing, planning, data validation, and processing in back-end workflow methods to make sure an correct and well timed response based mostly on the complexity of the request made by the shopper. This full course of, with a devoted electronic mail response workforce, has a excessive operational price. Because of this, we expertise elevated response TATs with rising electronic mail quantity, which has a adverse affect on buyer satisfaction, which is measured by way of the NPS (Internet Promoter Rating).

The answer

To deal with the above downside, we developed an AI (Core-Converse) answer that makes use of an built-in NLP engine that mechanically understands the shopper request by studying the emails, figuring out the intentions and the shopper request upfront solves. This AI platform is an end-to-end integration into Max Life’s core methods, reminiscent of B. Again-end coverage purposes, web sites, and different customer-facing chatbots. This allows the platform to mechanically determine the shopper, the dialog historical past, to validate buyer data, to extract the related data required for the shopper inquiry and to ship it again to the shopper as an e-mail response. Along with answering buyer inquiries, the platform additionally manages your entire e-mail workflow by planning the solutions and placing them in a queue to make sure the shortest potential processing time for answering these inquiries.

At the moment, the Core Converse AI platform processes 50% of buyer electronic mail inquiries mechanically with an correct response fee of 90% (10% of responses are checked manually). various kinds of buyer inquiries that clients write us for. For the remaining 25% of requests, that are fairly advanced, reminiscent of For instance, in reference to a coverage withdrawal or a mixture of a number of requests made by a buyer in a single electronic mail, the platform creates a response that’s at present being manually checked after which sends it to the shopper. With steady enhancements, we need to make the NLP engine sharper in order that greater than 90% of buyer inquiries may be processed mechanically by way of the platform.

This has enabled us to chop electronic mail desk prices by practically 30%, lower common TAT by 90%, and enhance our clients’ NPS by ~ 1%.

The way it works

The next diagram offers an summary of the method dealt with by the core converse platform.

C:  Users  vshom4069  Desktop  email bot.png

The method begins with receiving the e-mail despatched by the shopper. As soon as the e-mail is obtained, the validation engine program identifies and matches the shopper’s electronic mail ID by wanting by way of it within the registered buyer data. If the e-mail id is acknowledged as one of many present clients, the e-mail goes by way of the core NLP engine which identifies the intent of the e-mail. (What data is the shopper on the lookout for).

To resolve the shopper request, the engine extracts related data from the e-mail textual content, reminiscent of: B. coverage quantity, buyer quantity and different required data inputs, e.g. B. when a buyer is intent on on the lookout for a fee receipt for the coverage renewal. These inputs are then delivered to Max Life’s core purposes to acquire the data wanted to resolve the shopper request by way of varied built-in API calls. As soon as the response data is obtained, the platform integrates it right into a personalized response electronic mail template and sends it again to the shopper. The central Converse platform performs these steps mechanically to deal with the end-to-end electronic mail receive-reply course of.

The total platform is deployed within the cloud and leverages each the AI ​​and the deployment software stack of AWS and Google cloud parts. The next determine exhibits the technical structure of the answer.

C:  Users  vshom4069  AppData  Local  Microsoft  Windows  INetCache  Content.Outlook  CEY4UAJI  Picture2.png

Architecturally, your entire platform is split into two important elements – the e-mail workflow system and the core AI engine.

Electronic mail workflow system –

The primary goal of the e-mail workflow system is to determine and validate the sender by pulling data from varied supply methods. Utilizing the inputs from the core AI engine, it automates your entire electronic mail course of. As soon as the e-mail arrives, the core NLP engine extracts data reminiscent of the e-mail ID, area identify, coverage quantity, and intent of the e-mail.

See additionally

The workflow system is related to MLI supply methods by way of APIs with the intention to validate the id and to name up all processing-related information. Utilizing inputs from the core NLP engine, it validated the shopper identification by way of the coverage quantity and registered electronic mail id, fetched the required buyer data / information and delivered it again to the core NLP engine to generate the response.

In the course of the validation course of, the workflow part additionally detects whether or not the shopper has repeatedly despatched emails for a similar downside, whether or not the e-mail is tagged with different particular electronic mail IDs that point out an escalation, and whether or not the sender has basic Info reminiscent of workplace tackle, workplace wants contact quantity, and so on. Such sophisticated data makes it potential to correctly plan and prioritize the response within the e-mail response queue.

Core AI (NLP) engine –

The core part of the platform is the NLP-controlled AI engine. The first goal of that is to grasp the e-mail content material by way of NLU and to extract buyer intentions and different related data. The engine makes use of Dialogflow as its core NLP engine. It is packaged with a customized layer of pre-processing engine deployed by way of AWS Lambda that works as a tightly knit unit to seize buyer intentions from the e-mail physique and return to the CoreConverse workflow with the recognized intentions to reply. To beat the restrictions on the variety of characters that Dialog Stream can course of, there’s a customized skilled mannequin to interrupt the massive emails into smaller chunks in order that the entire emails may be processed. Earlier than the core engine identifies the intentions, this specifically skilled mannequin breaks the e-mail down into contextual snippets and sends every of these snippets to the core NLP engine, which replies again with the shopper’s intentions, that are re-consolidated and cleaned as much as determine all a number of intentions and take away all repetitive intent. Throughout this extraction course of, the mannequin performs topic modeling by figuring out a part of speech, lemmatization, spelling correction, phrase similarity, and frequency evaluation.

The platform makes use of an in depth tech stack. Along with the Google Dialog Stream NLP engine, the customized mannequin is built-in into Python. The platform entrance finish was developed by way of ReactJS and quite a few handlers, reporting providers, workflow process determination providers, pooling and routing providers have been used for deployment within the AWS cloud. As well as, Netflix persistence was used to maintain the portion of the workflow going.

Subscribe to our publication

Get the most recent updates and related gives by sharing your electronic mail.

Be part of our Telegram group. Turn out to be a part of a devoted neighborhood


Please enter your comment!
Please enter your name here