Utility Week

UTILITY WEEK | FEBRUARY 2021 | 39 Operational Excellence Included in Utility Week's latest report is an interview with Mark Wilkinson, head of income at Northumbrian Water Group, in which he explains how use of advanced AI and machine learning gave it a head start on debt management in these uncertain times. Here's a sneak preview: When did you start to look at using AI for debt man- agement? We did some work at a high level a few years ago, trying to use machine learning to predict our debt paths and make the system more reactive to what was actually hap- pening "on the ground". When you build a tool for debt management, it is valid at the time you build it but not necessarily six months down the line when there's a new customer seg- ment or a segment that behaves slightly di€ erently,‚etc. We wanted to get to a point where our system could learn as the circumstances change – a particular con- cern given all the current uncertainties around Covid19 – however, at the time the AI was new technology and quite hard to adapt. Fast forward to 2019 and we were introduced to the AI/ML specialist Inawisdom, whose advanced cloud- based system pro mised to give us much greater insights into our collections data, improve our engagement with customers and optimise the debt recovery process. What's the quickest win you've achieved in terms of better debt management? Crunching the data provided insights into aspects we hadn't even considered. Previously we sent customers text messages when a payment was due, but Inawisdom's analysis of historic data on our customer behaviour and the timing of messages revealed that if we sent SMS's to a speci' c cus- tomer group on a speci' c day of the week, for example on a Friday instead of a Wednesday, we would be more likely to receive payments. It's a di€ erent perspective, we had never really gone into the timing of things. Were you already using automated processes? We use an automated rules-based system that focuses on debt paths for speci' c customer groups, but the prototype developed and implemented by Inawisdom allowed us to re' ne the way those groups are segmented and change the order of the debt path for each customer group to achieve the best outcomes. We found that we could reduce the number of steps involved in collecting a customer's debt, cutting out steps identi' ed as unnecessary (some can't be missed due to regulatory requirements), in the process reducing our costs and improving the overall customer experi- ence. We've also been able to fast track certain stages, and reduce the time between triggering debt recovery and getting the money back. What customer data is fed into the tool? A plethora of data goes in if it's relevant, including cus- tomer payment behaviour, previous debt recovery activ- ity, complaints data, demographic data, socio-economic data, credit scores, data from the deprivation index, a model we previously developed to predict water poverty was another source. Obviously there is a data protection angle to all of this, particularly since the introduction of the General Data Protection Regulation. We anonymise any customer data before it's uploaded to the cloud so Inawisdom can't connect it to individual customers, then when we get the results back we reconnect it to their accounts. Did you have to overhaul your IT infrastructure? No, our statistical so˜ ware links into the Amazon Web Services (AWS) cloud platform where Inawisdom runs its Rapid Analytics and Machine learning Platform (RAMP). Inawisdom is a full stack services provider on AWS and using a cloud-based system has bene' ts in terms of data storage and security. Did you use the AI to estimate longer term debt provision? We carried out some modelling around the longer term impacts of the pandemic and what that might mean for provisioning and levels of bad debt risk, pulling together models for furlough, other ' nancial support, unemploy- ment changes and the impact on particular sectors. Provisioning models normally calculate bad debt risk based on previous bills, but this was more forward looking and predicted the increase in provision needed in coming years through the lens of Covid. Unemployment seemed to be the main driver for bad debt, the extension of furlough and initiatives like Eat Out to Help Out weren't in play when we did that model- ling, but the model is adaptable so we have the option to update parameters and create new forecasts. The tool could come in useful – our ' ve-year business plan means we have ' xed income over that period, so any shocks need to be planned out in advance. What are the limitations of AI? The biggest limitation is our systems and ensuring we can make the most of the learning from Inawisdom to implement e€ ective changes to our debt collection strategies. They built a collections model that essen- tially looks to optimise every element of the output, but the constraints we currently have to work to limit its e€ ectiveness. Q&A Mark Wilkinson, head of income at Northumbrian Water Group "We wanted to get to a point where our system could learn as the circumstances change – a particular concern given all the current uncertainties around Covid19" Download the full report Doubling down on debt – AI and Machine Learning in Action at: https://util- ityweek.co.uk/doubling- down-on-debt-ai-and- machine-learning-in- action/ in association with

• Cover