FTSE 250 Energy Company

Reduce the amount of financial provision (£100M) for ‘bad debt’

Data Discovery:
Half-hour consumption data for 10,000 customers over 100 days
Industry SIC codes chosen to create data model of different business consumption rates
Identification of 100 labelled energy data anomalies (17,500 customers/2 years)
Anomolies ranged from change of tenancy, theft and faults

Agree Priorities

Model Bad Debt Provision

Energy consumption anomaly detection report created using multiple techniques and models
Deep Learning autoregressive Recurrent Neural Network
Model outputs a prediction based on historic data
Predicted usage is compared with the real usage –and identifies anomalies if out of confidence levels
Retrospective analysis across all of the client’s data, carried out
Results enable a prediction of ‘bad debt’ to be calculated

£100M bad debt provision reduced to £3M
£97M reallocated within months to internal investment fund
Further Bad Debt collection optimized by avoiding potential bad debtors being onboarded

FTSE 100 Financial Services Company

Report based on initial Data Analysis
A schematic on test data of one of the analyses, a Sankey diagram analysis of how calls route through from start to finish, can be seen on the right.
Even though this was a test dataset, the client stated that knowing if the fraction of outbound calls ending with customers disconnecting (i.e. hanging up) increases, it is likely a sign of a marketing campaign that is performing poorly.
Another discovery was that a significant amount of issues were caused by UK regional accents, with the word ‘balance’ being detected as ‘balan’ [sic] frequently enough to be an issue.

Improve Natural Language Call Steering (NLCS)

Use Natural Language Processing (NLP) Machine Learning (ML) models on customer spoken intents; model learns most common themes of calls automatically
The time taken to get a first solution for testing ready is shortened considerably, model monitoring is built-in and new commonplace intents that aren’t being captured can be identified

Proactive Incident Response

Monitor call/caller locations, intents and transcripts in near-real-time in aggregate, and compare call sentiment and topics to historic, e.g. same time last year, and different regions
A sudden negative spike all mentioning the same theme might require a dedicated Call Centre and/or company-wide response

Process using LLMs and cosine similarity, and group into clusters in seconds (unsupervised ML)