Sean Maddison

Omnichannel & AI-enabled evolution

Building an all-channel, intelligent and personalised conversational foundation for one of the UK's largest electricity and gas energy suppliers. Making sure customers can easily manage their account and get answers to common questions, freeing up human agents to spend time on those with tricky problems or who need extra support.

Project overview

Background

A large retail energy supplier, taking over 200,000 voice calls and 60,000 chats every week—along with a significant presence on social media. There is some traditional IVR functionality, with no automation in any other contact channels. They also have some native website and mobile app self-service capabilities, but extending & adding to these is often costly and takes a very long time to delivery.

Problem

Firstly: we wanted to grow our digital contact channels but had no intelligence or automation capabilities to offer, with all contacts going straight into a queue for a human agent.

Also, we had identified over 60 use cases for new bots to handle core customer transactions. Using traditional IVR and DTMF menus, there was no way to practically deliver more than a handful of services without creating unwieldy menus that would cause customer frustration.

Finally, as a very large organisation many customer contacts ended up with the wrong agent and had to be transferred—wasting both the agent & customer's time.

Solution

We ran a rapid, agile infrastructure project to give us the ability to route digital contacts through our automation layer (Genesys Intelligent Automation), and to make Natural Language Understanding (NLU via Google Dialogflow) available across all of our channels.

As well as the technical delivery, another key outcome of this project was to train a tight team of experts on how to build conversational omnichannel bots, and maintain our NLU models.

Design approach

Our base contact centre platform supported delivering automation through multiple available channels, so this stage was a relatively simple task of altering our routing strategies to accomodate this. We extended our existing infrastructure to route webchat, mobile app messages, SMS text messaging, and Facebook Messages through our automation orchestration solution.

The main decision and integration work was in selecting an NLU provider, and making sure this worked effectively in each channel. To support this, I built several prototypes in the leading vendors solutions and assessed them all on a number of criteria. I took the 5 main NLU products on the market at the time and scored them on the following criteria:

• Ease of use by trainers and admins
• Ease of integration into our existing stack
• Cost per transaction
• Support for UK English

We scored against these based on trainer & integration team feedback, product documentation and our own trialling of each product. I eventually selected Google Dialogflow as best meeting our needs.

Roll out

To support testing and live roll-out we had to build out a small NLU model of a dozen well-known intents, and create a simple omnichannel bot that would work in voice, rich-media chat, and plain-text chat (for social media and SMS).

Continuous improvement

Once the NLU roll-out was underway, I started a continuous Agile programme of change to identify improvements to existing bots, new bots to develop, and other solutions to improve the customer journey. We had many sources of user stories for fixes, improvements, and new bot concepts—a few examples:

• My team's own ideas
• Insight from raw customer input into NLU routing
• External customer research
• Insights from our business reporting team
• Competitor intelligence
• Feedback from our own customer service agents

NLU training strategy

We wanted to quickly build out our NLU model to start getting immediate benefits by improved contact routing and reduced internal transfers. To achieve this, I implemented a live pilot "burst training" model with my new team of administrators.

1. Select an inbound call or chat line
2. For X minutes each day, put Y% of customers through the new NLU journey to capture their raw input and route to a default destination
3. Once the capture time is finished, bot trainers then analysed the raw input to categorise NLU intent
4. Repeat over four weeks for all primary business areas

Ongoing development

By the end of this project we had a new suite of technical capabilities, and began training our new team to start really getting some value out of this.

With the technology & team in place we ran a continuous development agile programme to implement new bots, develop new APIs into our backend systems, improve our NLU model, and much more.

Benefits

This project was a technical foundation that let us go on to spend 2+ years building out over 50 conversational omnichannel services on top of a very comprehensive NLU model, which ultimately saved the business over £15m a year in operational costs.

We did see an immediate benefit in improved performance of our traditional IVR system, just as a result of replacing our legacy voice ASR system with Google's AI-based speech recognition service. This boosted existing application success by around 10%