Improving staff productivity at Enel using Amazon Bedrock

Enel is a leading integrated electric utility with a presence across 32 countries and an 82-GW generation capacity. The company also assumes a pivotal role as a major electric grid operator, supplying a vast network of 76 million customers and managing 46.5 million smart meters. Since 2014, Enel has been investing heavily in artificial intelligence (AI) by developing a strong in-house know-how that has facilitated the company’s adoption of generative AI using Amazon Web Services (AWS). This technological advancement has made it possible to seamlessly automate tasks that were previously carried out manually.

Context

Enel operates an advanced IT service desk based on ServiceNow to allow users to create and track tickets that encompass various needs, including support for business applications. Service desk tickets related to business applications cover different topics, from user account creation and permissions management to application errors or performance issues. These issues are managed by dedicated Application Management Service (AMS) teams.

Over the years, Enel has automated ticket management tasks where automation was readily achievable with the required level of quality and customer satisfaction, specifically for standard and trivial issues. However, many nonstandard IT service desk requests for business applications are repetitive and are related to documented procedures, and their fulfillment can be automated. Usually, these tickets are manually analyzed and processed by AMS teams, requiring staff effort that could be optimized or allocated to more complex issues. In addition, during periods of peak workload, low-priority requests get queued, resulting in higher resolution time and end-user dissatisfaction.

Enel identified the opportunity to use generative AI to boost IT service desk efficiency by extending automation to nontrivial tasks through basic troubleshooting, providing resolution steps and ticket routing without human involvement. For Enel, the goal of this project was to automate the first-level management of business application tickets, either by generating instructions for the requestor, thereby automatically fulfilling the request, or routing the request and work notes to the most appropriate resolver group. At launch, the initial scope was limited to three business applications for a total volume of 2,000 tickets per month. Moving forward, the goal is to expand this solution to the entire business application portfolio. By automating first-level management, Enel can reduce the workload on the AMS support team and speed up issue resolution while focusing the AMS staff on more complex and high-value tasks.

Solution

The solution is designed around a retrieval-augmented generation architecture and uses Amazon Bedrock, a fully managed service that offers a choice of high-performing foundation models (FMs) from leading AI companies along with a broad set of capabilities needed to build generative AI applications with security, privacy, and responsible AI.

The solution uses Amazon Titan, a family of models exclusive to Amazon Bedrock. Specifically, it uses the Amazon Titan Text Embeddings model to generate embeddings (vectors capturing semantics of text) from Enel’s knowledge base, which consists of a series of runbooks containing incidents classes, preconditions, root causes, resolutions steps, and operations information related to the applications. Embeddings are computed and persisted in a vector database instance (MongoDB Atlas Vector Search), which supports similarity search.

The solution then uses Anthropic Claude to generate the most appropriate ticket response based on the ticket description and the additional context provided by matches in the vector database. Possible responses include providing instructions to solve the issue, requesting more information, generating work notes and routing the ticket to the correct second-level support team, closing the ticket, or informing the user of system unavailability. To strengthen the solution, Enel has implemented comprehensive health checks to ascertain the operational status of the systems hosting the applications.

Because Enel operates across multiple countries and languages, the solution uses generative AI translation capabilities to interact with requestors using the same language that was used to create the case.

Enel developed the solution through microservices integrated in the Enel Digital Platform (EDP), which uses Amazon Elastic Kubernetes Service (Amazon EKS), a managed Kubernetes service, for containers orchestration. These microservices cover two major procedures: a set of batch processes (indexing process) and a transactional user interactions process (resolution process). The batch processes, described in Figure 1 below, retrieve domain-specific data from the Enel knowledge base (runbooks) hosted on Atlassian Confluence and tokenize the retrieved data, produce embeddings, and store and/or update embeddings in a vector database instance. An orchestration microservice coordinates all the different activities for the indexing batch process.

Figure 1: Indexing batch process on the Enel Digital Platform

The transactional user interactions process is described in Figure 2. The process uses an integration that Enel built using ServiceNow, so as soon as a new ticket is created, it triggers the resolution process to first convert tokenized ticket text into embeddings using Amazon Titan Text Embeddings through Amazon Bedrock. As said before, in this initial phase, the solution performs system health checks for the business application. If the business application is properly working, then the process retrieves relevant context from the vector database and generates the resolution response using Anthropic Claude through Amazon Bedrock. The text generated is directly added to the ticket in ServiceNow, and the ticket is either closed, returned to user, or routed to the right resolver group, depending on the specific resolution steps listed in the runbook. The orchestration microservice coordinates all the different components of the interactive resolution process, including microservices and health checks. This process is further complicated when the ticket request text is not in English; in this case, the ticket text is translated as soon as the resolution process is triggered, and the resolution response is generated in the ticket requestor language.

Figure 2. Interactive resolution process on the Enel Digital Platform

Business outcomes

Following the successful implementation of generative AI–managed ticket resolution for IT service desk tickets , Enel observed a reduction in time required to resolve cases, down from 1 day to less than 2 minutes.1 In addition, the implemented solution resolved about 15 percent of the tickets automatically and without human intervention, thereby reducing the time needed to provide a first response to the user from 9 hours to 1 minute.2 These results are in line with Enel expectations and are encouraging the company to progress with project improvements that aim at both refining and expanding the solution to a broader portfolio of business applications.

Long-term goals

Looking to the future, Enel intends to expand the use of FMs and large language models (LLMs) into existing and future products by implementing a deeper integration of Amazon Bedrock into its platform.

“At Enel, the pursuit of efficiency starts with streamlining processes and enhancing staff productivity. In alignment with this commitment, we harnessed the power of Amazon Bedrock to test the integration of generative AI–assisted features into our service support platform,” says Fabio Veronese, head of ICT Industrial Delivery at Enel. “Preliminary results indicate a promising trajectory toward a significant reduction in the effort needed to resolve cases.”

Conclusion

In this post, we saw how Enel, a leading player in the energy industry, is using generative AI and Amazon Bedrock to optimize an existing and well-established manual process and is improving key performance indicators while reducing human effort. To learn more about how you can optimize your processes and transform existing application with Generative AI, see Amazon Bedrock.

Notes
1. This result is only for tickets completely resolved by the generative AI.
2. These results have been computed on the first subset of tickets for the three chosen applications.