Developing First Utility's chatbot: 'the smallest mistake can make customers very upset'

Dr Natalia Konstantinova describes the benefits and challenges of creating a tireless customer services operative

Two-and-a-half years ago, independent energy supplier First Utility began rolling out a chatbot, building its own in-house machine learning algorithms in order that its customer services coverage could be extended to 24 hours a day.

This was not easy at first, lead software engineer Dr Natalia Konstantinova told the audience at the Computing Big Data and IoT Summit last week.

"There was little understanding at that time, not many people had heard of chatbots," she said.

First Utility could have taken the easier route of outsourcing, but preferred to keep its IP in house, hand-crafting its own rules to more flexible and responsive and for easier integration with existing back-office systems. Data protection was a consideration too.

To create a seed set of rules for the algorithms to learn from, two or three people had the "really boring work for a couple of months" of annotating the archived customer communications data, adding categories and labels, Konstantinova said. In fact, in the early days the biggest challenge was to get the business to release people and funding for these sorts of tasks.

"It was a completely new greenfield project, and we had no idea what customers would ask. We had to persuade the business it would really bring value", she said.

The chatbot has to deal with four classes of customer enquiry. In ascending order of business value and programmatic difficulty, these are: Q&A, account query, conversation and action.

The first of these, Q&A, covers general enquiries for which the answer is the same for everyone, such as "what is a smart meter?" Account queries are contextual and therefore require knowledge of the customer's record - you can ask about what meter readings are needed, or how much you paid last month, for example.

Conversations are about more complex queries Konstantinova explained.

"The customer might say, 'I think my meter is faulty', so then we can run through a list of diagnostics. If we can solve the problem and save a phonecall then that's really great for us."

Action queries are where the highest value lies. They are transactional queries as a result of which the customer account details are changed. First Utility's chatbot can allow users to change tariff for example.

The company was driven to experiment with chatbots because it is a fast-growing supplier (700 per cent growth in customers in 10 years) that differentiates itself largely on customer experience. It is also adding on new services, most recently telecoms and internet, and therefore requires flexibility in its response.

So far it has proved popular with customers, Konstantinova said, as it enables a 24/7 contact channel. It has reduced the average handling time for problems to be resolved, and it is cheaper than any interaction that involves human contact.

That said, there are still issues to be worked through, such as who owns the chatbot (IT, marketing or customer services) and how to integrate some business processes. There is also the fear among some operatives that it might put them out of work, although Konstantinova insisted it's an additional channel rather than a replacement.

Current efforts revolve around fine-tuning the algorithms, including alerts to indicate when a human operator should take over.

"Most customers call in when they have a problem so they're not very happy, and the smallest mistake can make them really upset," Konstantinova said. "You need a control system where you can override the black box".

Because its algorithms are constantly learning from their interactions, the occasions when they need to be overruled should become fewer all the time.