Epilot Reduces Email Processing Time by 87% Using Amazon Bedrock

Based in Cologne, Germany, Epilot provides an XRM solution for the energy industry. The solution goes beyond customer relationship management, handling assets, sales, services, and other processes. Built using a serverless architecture on AWS, Epilot’s XRM can tackle complex use cases for more than 170 customers—energy suppliers, municipal utilities, grid operators, and solution providers.

Communication accounts for a central part of the workload in sales and services in the energy sector, for both energy solutions and the commodity business. Email chains quickly become lengthy and complex, costing employees time as they read through messages. Epilot sought to accelerate the process, using generative AI to analyze entire email threads and summarize them in a few key points so that sales or service employees can immediately recognize any action needed.

Using funding from AWS, Epilot began experimenting with generative AI in December 2023, conducting user research to determine which features its customers needed most. Working closely alongside AWS Solutions Architects with expertise in generative AI, Epilot launched a minimum viable product in 2 months to summarize long email threads using generative AI.

Epilot built the XRM’s serverless architecture, which seamlessly incorporates the new generative AI capabilities, on AWS Lambda, a compute service that runs code in response to events and automatically manages the compute resources. An API sends the user’s request to summarize an email chain to Amazon Simple Queue Service (Amazon SQS), fully managed message queuing for microservices, distributed systems, and serverless applications. The event in the queue initiates an AWS Lambda function, which orchestrates the generative AI task of summarization in Amazon Bedrock.

Built on this scalable, event-driven, and serverless architecture on AWS, Epilot’s XRM can efficiently handle large volumes of tasks. Engineers don’t have to worry about provisioning compute resources, thereby having more time to work on feature development.

Before the email summarization feature moved into production, Epilot turned to Amazon Bedrock Evaluations to conduct internal testing and define success criteria. To pinpoint the right model for the use case, the engineering team worked with the product team to use human-based Amazon Bedrock Evaluations and the out-of-the-box UI for rating different prompt versions across various email threads. After a thorough evaluation, Epilot chose Anthropic’s Claude in Amazon Bedrock.

Using Claude Sonnet for its AI summarization feature, Epilot generates 55,000 email summaries each month with a negligible failure rate. About 80 percent of Epilot’s users say that the feature simplifies their work. “Using AI summaries, our users save 87 percent of their time,” says Kerem Nalbant, senior product engineer at Epilot. “As an early adopter of Amazon Bedrock, we’ve seen firsthand how it delivers cutting-edge AI capabilities with the security and support we need.”

Epilot is developing additional features for customers in about 2 months from concept to production. A feature called suggested actions permits users to automatically update information rather than copying it from an email chain—thus saving time while minimize human errors, such as mistyping. For example, if a customer sends a request to change their address, Epilot can revise all the customer’s information with one click. The AI agent automatically sorts through complex attributes such as postal codes and street names, extracts the data from the email chain, and uses Epilot’s Entity API to update the data. It also keeps humans in the loop to verify automatic actions. “Our goal is to improve the efficiency of the users of our solution,” says Viljami Kuosmanen, chief product engineer at Epilot.

Furthermore, the company is fostering a culture of creativity by encouraging employees to think of generative AI use cases, holding regular hackathons for engineering teams, and hosting tech exchange sessions. For example, an Epilot employee outside of the AI team developed AI-driven entity filters, a feature that helps customers use natural language to quickly find information. And the AI team stepped in only to take the feature to production. “We have greater collaboration among departments and engineering teams,” says Nalbant. “AI has expanded our sense of what’s technically possible. Our team is coming up with creative ideas and solutions that would have seemed impossible few years ago.”

Additionally, Epilot’s customers benefit from Amazon Bedrock Security and Privacy, which provides full control over the data that organizations use to customize the foundation models for their generative AI applications. By default, Amazon Bedrock offers a zero-retention policy—that is, it doesn’t share logs, prompts, model outputs, or personal data with third parties. Moreover, all processed data remains within a European Union AWS Region. “We are doing our best to build the trust in AI so that our users can make the most out of it,” says Nalbant.

Epilot continually releases iterations of its generative AI features based on user feedback. It is also experimenting with models through Amazon Nova, state-of-the-art foundation models that deliver frontier intelligence and industry-leading price performance, available exclusively on Amazon Bedrock.

“The pace of innovation in this space is incredible,” says Nalbant. “We’re building a strong road map that aligns with our vision of bringing the right solutions to our users.”