Being early adopters of Snowpark last year with Snowflake, EDF UK played a huge role in pushing pioneering change in the ML and AI domain with the use case receiving media coverage and promotion on the main stages of AWS Summit, Big Data LDN, and even featuring on the Snowflake Summit Las Vegas Keynote. Now, EDF is taking on all ML and AI challenges of the organisation with platform-as-a product, reusable, and shared capabilities.
Data driving transformation
EDF UK had ambitions for years to use data, analytics and AI to become a customer-centric organisation and help Britain achieve Net Zero. However, a legacy technology stack that was restricted the team.
Last year, the vision of ICE (Intelligent Customer Engine) was formed with three core principles: simplicity, progress over perfection and bringing the code to the data. The EDF team already had Snowflake storing nearly a PB of data across the organisation. Taking data out of Snowflake and putting it into a siloed MLOps platform led to barriers surrounding accessibility, administration, complexity, and cost.
Running the EDF ML and AI product workloads directly over the data with zero movement was logical, which is why EDF chose to be early adopters of Snowpark. The team transitioned all ML and AI workloads to run directly over Snowflake, integrated with AWS technologies such as Sagemaker. This new approach unlocked benefits and recognition such as:
- Capability productionised within three months.
- Data Science products release cycles improved four-fold.
- Data Scientist efficiency improved by 80%.
- Cost reduction of c.£3 million per year.
- Customer lifetime value and financial vulnerability model initial products released, improving everything from marketing campaigns to supporting financially vulnerable customers.
- Introducing real-time decision making with AI:
- Knowledge assistants utilising retrieval augmented generation technology, capable of delivering a response in less than five seconds.
- NLP for contact intelligence analysis, insights, and routing.
- Hosting of finetuned LLM’s specifically for understanding EDF business language
- Enabling a play development area for data scientists to access AI ready data from a centralised platform.
The ICE platform is a unified MLOps environment that integrates AI and ML AWS and Snowflake services. With a sensible list of environments and minimal amounts of infrastructure to be managed, it allowed all development focus from EDF to be on pipelines and models, becoming products.
The team built four environments and created a handful of access personas and roles. There are now only a few infrastructure components to manage rather than over 50, which was the case before the transformation. One major benefit of this flexibility means the data team can ramp up GPU and compute power for when it is needed, optimising costs.
The initial ICE platform gave EDF UK the ML foundations to serve the use cases of the retail business. However, EDF had an AI team ready to take ICE to the next level with a set of generative AI (genAI) use cases to revolutionise the customer business and enable real-time decision making.
A standout innovation for the EDF AI Team is the Ask Agent ENZO – a genAI RAG pipeline hosted in ICE which puts an on-demand knowledge base into the hands of a contact centre agent:
- User asks a question of the knowledge base.
- Safety checks are done, and API send the request to the indexed content find top four chunks of text most relevant to the words used in the request.
- Chunks are sent to the embeddings model to be LLM ready.
- Embeddings are sent to the LLM to generate a reply based off the specific prompt.
- Response is sent back to the user via API.
- Safety checks are done and a generated response is sent back to the user.
Other noteworthy metrics about ICE include:
- It has been finetuned using LLM’s (FlanT5 and Mistral) and genAI analysis utilising Bedrock:
- Agent eNZO offering informative knowledge base content to agents to improve CX.
- AI-infused IVR solution reducing transfer rate time by 20%.
- Priority services register model is a major data for good initiative. By handling vulnerable customer contact to route them to the team will correct competency to manage contact and the average handle time improved five-fold.
- Complaint model proactively identifying complaints in less than six hours, serving 3 million customers across 250,000 contacts a day.
- AI self-serve outcomes have improved the CX that the ‘ease’ measure of our customer journeys has improved from 2.5/5 to 4/5.
- Rebuilding the MHHS and Volume Forecasting product, which is used to forecast customer energy requirements across 5.5 million customer accounts on the £10 billion energy market EDF UK manages. Building complex ML models on over 1 trillion rows of data including industry, weather, and customer consumption data to manage £300 million of risk and tens of millions of transactional risk on hedging positions.
- Debt analytics leading to over £55 million in payment resolutions and support for financially vulnerable customers.
- Data science contributing to a four-fold increase in ML products.
- Integration of marketing tools with CLV, resulting in a 30% increase in net zero product sales and manging 71% of SMART meter bookings.
- EnergyHubApp leveraging consumption data across over 600 billion rows of data annually to aid customers in managing energy usage.
- Billing to settlements process unlocking £75 million in value through reconciliation techniques.
- Adopting true DataOps principles and software engineering infused into data approaches to help 161,000 beat the peak participants and reduce grid demand during peak hours.
EDF are part of the DataIQ membership programme – the trusted global collaboration and intelligence platform for data leaders. Find out more here: https://www.dataiq.global/membership
