## Overview So Energy is an independent energy retailer in the UK that was formed approximately a decade ago with a mission to provide 100% renewable electricity to customers at competitive prices without compromising customer service. The company serves around 300,000 customers with approximately half a million meter points for gas and electric. This case study, presented at AWS re:Invent 2025, details their comprehensive transformation of customer experience through the implementation of Amazon Connect's AI-powered contact center platform. The presentation was delivered jointly by Hara Gavliadi, a Customer Success Specialist for Amazon Connect at AWS, and Mohammed Khan, Director of Technology and Product at So Energy. This dual perspective provides both the vendor's strategic view on customer experience enhancement and the customer's practical implementation journey. ## Business Context and Challenges So Energy faced a fundamental challenge that Mohammed Khan articulated clearly: customers don't typically enjoy speaking with their energy provider. Unlike conversations with friends and family, interactions with energy companies occur primarily when something has gone wrong or when customers desperately need help. This means customers arrive at these interactions already frustrated—they're starting from a negative baseline rather than neutral ground. This challenge was significantly amplified by several external factors. The COVID-19 pandemic forced So Energy to transition overnight from a fully in-house operation in a small West London office to a completely remote workforce. This sudden operational shift strained their existing processes and platforms. Following closely on the heels of the pandemic, the UK energy crisis caused customer bills to double or triple overnight, fundamentally changing customer perceptions of energy companies and dramatically increasing contact volumes beyond anticipated levels. Adding further complexity, the electrification of the energy industry was transforming customer needs. Energy was no longer just a monthly bill that customers paid without much thought. Customers began thinking about energy usage on a daily basis, considering EV chargers, time-of-use tariffs, and other complexities that required significantly more support and education from their energy provider. The specific operational problems So Energy faced included: **Fragmented Platforms:** Agents were forced to use one platform for voice communications, another for chat, and yet another for email. When customers contacted the company across multiple channels, they had no continuity of context and were required to retell their entire story each time they switched channels. This created significant frustration for customers who expected seamless experiences. **Repetitive Manual Work:** The fragmentation meant agents had to constantly re-verify customer details, manually take notes, and track customer context across disconnected systems. This placed enormous cognitive load on agents and reduced the time they could spend actually solving customer problems. **Lost Escalations:** When cases were escalated to back-office teams, So Energy's processes weren't adequately tracking these tickets. In some unfortunate instances, escalations would be lost entirely, leading to customers contacting the company again only to find that the organization had no record of where their issue stood. **Compounding Frustration:** All these operational inefficiencies meant that customers who were already frustrated at the point of initial contact became even more frustrated as their interactions progressed, leading to negative sentiment and deteriorating customer relationships. ## Solution Architecture and Implementation So Energy embarked on a platform selection process that evaluated traditional CCAS (Contact Center as a Service) platforms, which offered strong but traditional functionality with AI features that were often repackaged NLP capabilities or basic text summarization. They also considered new market entrants with interesting offerings but deemed them too risky given the critical importance of the contact center to their business. Amazon Connect emerged as the chosen platform because it had evolved from a disruptor to a leader in the Gartner Magic Quadrant while maintaining significant ongoing investment and innovation velocity. The key requirements that drove the platform selection were: **Truly Omnichannel Service:** So Energy wanted to meet customers where they were, not force customers to come to them. This meant supporting voice, chat, email, WhatsApp, social media, and other channels within a single unified context. The platform needed to maintain complete continuity regardless of which channel customers used to initiate or continue their interactions. **Agent Empowerment:** The solution needed to provide a unified workspace that gave agents complete customer context, allowing them to spend less time figuring out what customers were saying and more time solving actual problems. The platform needed to incorporate agent assist capabilities and reduce cognitive load, especially given the increasing complexity of the energy industry. **Dynamic Resourcing:** The company wanted the ability to flex capacity to match demand dynamically, routing customers to the right agents at the right time based on urgency and topic expertise. **Future-Ready Foundation:** Critically, So Energy recognized that the contact center industry was on the precipice of significant AI-driven change. They didn't want to find themselves on a perpetual transformation treadmill, conducting expensive migrations every few years. They needed a platform that would grow organically with advancing technology capabilities. The implementation journey followed a structured timeline: - **POC Phase (December-January):** So Energy initiated a proof-of-concept starting around December-January, working closely with AWS colleagues in London. Notably, they attempted to engage an implementation partner but found it actually slowed them down rather than accelerating progress. Their in-house tech team possessed sufficient context and capability, and AWS provided the necessary upskilling on the Amazon Connect platform specifically. - **Design Phase (Concurrent with POC):** As the POC neared completion, they moved into a design phase where they co-created their solution architecture with AWS architects, incorporating best practices to maximize their return on investment. - **Build and Test (April-May):** During this phase, So Energy deliberately focused on foundations rather than getting distracted by new shiny features. This grounding helped ensure successful implementation and adherence to timelines. - **Change Management and Training (May-June):** The company trained over 250 people across four weeks on the new platform, completing the rollout and exiting their legacy platform in June. The entire transformation took approximately 6-7 months from initiation to completion. Mohammed emphasized that the pace of acceleration depends heavily on ecosystem complexity and, crucially, how quickly you can bring people along on the journey. The human change management aspect typically becomes the slowest part of any transformation. To build excitement and buy-in, So Energy had AWS present at company all-hands meetings and conduct customer experience presentations. This helped create organizational excitement around the transformation and addressed potential concerns early, particularly fears around AI replacing jobs. ## AI and LLMOps Capabilities Deployed The Amazon Connect implementation incorporated several AI-powered capabilities that directly address LLMOps considerations for production environments: **Automatic Identity Verification and Intent Recognition:** Regardless of which channel customers use to contact So Energy, the system automatically performs identity verification and recognizes customer intent. This AI-driven capability eliminates manual verification steps and enables the system to understand what customers need before agents even begin the interaction. **Contact Summarization:** One of the most impactful AI features deployed was contact summarization. When customers have called multiple times or have ongoing interactions, agents can glance at AI-generated summaries to quickly understand the complete context before speaking with the customer. This allows agents to immediately begin working on solving the problem rather than spending valuable time reconstructing the situation from raw notes and transcripts. **Case Management and Task Orchestration:** The platform provides unified case management that maintains complete context across all channels. Whether an interaction starts as a voice call, continues as an email, and concludes via chat, the system maintains a single unified view of the customer case. This eliminates the context-loss problem that plagued their previous fragmented system. **Intelligent Routing:** AI-powered routing ensures customers are connected to the right agent based on their needs, the agent's skills, and current capacity. This intelligent matching improves first-contact resolution rates and reduces customer frustration. **CRM Integration:** So Energy achieved deep integration between Amazon Connect and their internal CRM system. The energy industry has specific quirks and requirements, and the platform needed to adapt to So Energy's processes rather than forcing the company to adapt to the platform. This integration provides agents with a single unified view combining contact center data and CRM customer information. ## Production Operations and Results The transformation from fragmented systems to unified AI-powered platform yielded significant measurable improvements: **Operational Efficiency:** Call wait times decreased by 33% within a few months of rollout, with expectations for continued improvement. Critically, these improvements in voice handling didn't come at the expense of other channels—email response performance remained strong, demonstrating sustainable progress across all channels simultaneously. **Channel Expansion:** Chat volumes increased from less than 1% to 15% of total contact volume, demonstrating that So Energy successfully met customers in their preferred channels rather than forcing customers to use only voice. **Customer Satisfaction:** CSAT scores improved significantly, with Trustpilot ratings approaching 4.5. Mohammed emphasized that they achieved this through focusing on basics: answering customer queries as quickly as possible and getting things right the first time, without implementing fancy AI gimmicks. **Agent Experience:** Agents moved from working across three different screens and platforms to a single unified workspace. The cognitive load reduction was substantial, allowing agents to focus on problem-solving rather than system navigation and manual context reconstruction. **Resource Optimization:** The platform enabled concurrency where agents could handle voice and email simultaneously, with voice taking priority. This meant all agents could take calls, but when call volumes were lower, they would automatically work email queues, dramatically improving resource utilization. **Training Acceleration:** New agents could begin productive work within their first week rather than waiting for four weeks of comprehensive training. The system allowed managers to assign focused work appropriate to agents' current skill levels, accelerating time-to-efficiency. **Operational Visibility:** Management gained complete visibility into customer contacts, whether they were tasks with back-office teams, active with agents, or waiting on customer responses. This end-to-end visibility enabled unified management of customer experience. The implementation did encounter initial teething issues. Performance actually declined slightly in the first weeks after June rollout, which Mohammed noted they had anticipated. Rather than panicking, they systematically addressed concerns from people and process perspectives, made necessary workflow adjustments, and turned performance around within a matter of months. ## LLMOps Considerations and Balanced Assessment This case study demonstrates several important LLMOps principles and considerations for production AI deployments in contact centers: **Foundation-First Approach:** So Energy deliberately avoided getting distracted by cutting-edge AI features during initial implementation, instead focusing on solid foundations. This pragmatic approach ensured successful deployment and created a stable platform for future AI capability expansion. From an LLMOps perspective, this represents wise production engineering—build reliable infrastructure before layering on sophisticated AI. **AI as Enabler, Not Solution:** Mohammed was explicit and refreshingly candid about tempering AI hype. He noted that AI isn't going to solve all problems on its own; rather, it's an enabler. The focus remained squarely on benefits delivered to customers and agents rather than technology for technology's sake. This perspective is crucial for successful LLMOps—understanding that AI capabilities must serve concrete business outcomes rather than being deployed simply because they're available. **Human-Centered AI Deployment:** So Energy made people part of the AI story, deliberately addressing concerns that AI would replace jobs. By positioning AI capabilities as tools that empower agents rather than replace them, they maintained organizational buy-in and avoided resistance that often derails AI initiatives. This is a critical LLMOps lesson: production AI systems succeed when humans are properly integrated into the design and implementation. **Measuring What Matters:** Rather than tracking vanity metrics, So Energy focused on CSAT and first-time resolution. This disciplined approach to metrics ensures that AI capabilities are evaluated based on actual customer and business value rather than technical sophistication. From an LLMOps perspective, this represents proper evaluation methodology for production systems. **Platform Evaluation Considerations:** It's worth noting that while So Energy achieved strong results with Amazon Connect, their evaluation process revealed that many vendors were repackaging existing NLP capabilities as "AI" or offering only basic summarization features. Organizations evaluating contact center AI should critically assess whether claimed AI capabilities represent genuine advances or marketing rebranding of established technologies. The case study suggests that Amazon Connect offered more substantial AI capabilities, though customers should conduct their own thorough evaluations. **Integration Complexity:** The deep CRM integration So Energy achieved was crucial to their success, but the case study doesn't deeply explore the technical challenges involved. Organizations considering similar implementations should carefully assess integration requirements, particularly for industry-specific CRMs with unique data models and workflows. ## Future Roadmap and Agentic AI So Energy's future plans demonstrate a thoughtful progression of AI capabilities moving from reactive to predictive customer service: **Enhanced Agent Experience:** They plan to leverage agent assist capabilities more extensively, ensuring agents have optimal context for handling customer queries and achieving first-time resolution. They're implementing AI-based performance evaluations (what Amazon calls edge performance or performance evaluations) that will quality-check every single customer interaction and enable AI-powered agent coaching, creating a continuous improvement feedback loop. **Customer Self-Service:** So Energy is expanding their chatbot implementation (which was deployed concurrently with the main platform) to be more context-aware. They plan to progress to voicebots, giving customers additional self-service options for those who prefer to handle matters independently. **Agentic AI Experiments:** The company is experimenting with agentic AI agents that can handle more complex queries autonomously. One example they've already tested involves automatically processing emails for certain intents: if a customer hasn't provided sufficient context, the system automatically replies requesting the necessary details before processing the request. This represents early exploration of autonomous AI agents with deterministic tooling to solve customer queries independently. **Proactive Outreach:** Future plans include using the platform's capabilities to proactively reach customers and resolve potential issues before they become actual problems. ## New Amazon Connect AI Capabilities Announced During the re:Invent 2025 conference where this presentation occurred, Amazon announced approximately 30 new features for Amazon Connect, with three highlighted in the presentation: **Fully Agentic Solution:** Amazon Connect has been transformed into a fully agentic solution where intelligent AI agents can autonomously handle complex customer interactions. These agents leverage comprehensive tools including knowledge bases, external systems, and Amazon Connect cases, allowing them to independently access, create, or update customer information while resolving issues. **Enhanced Self-Service Voice with Agentic Voice:** The automated voice system can now conduct natural, flowing conversations that pick up different accents and emotional cues. Rather than rigid robotic responses, these AI agents understand context, gather appropriate information, and take meaningful actions to resolve issues. **Message Processing for Channels:** Amazon Connect now supports message processing that intercepts and processes messages, emails, and chat messages before reaching end customers. This can automatically detect sensitive information like PII data across multiple languages and support custom processing based on business-specific rules. These announcements represent continued evolution toward more sophisticated AI capabilities, though the case study itself focuses primarily on the foundational implementation completed earlier in 2024. ## Critical Takeaways and Lessons Several key lessons emerge from this case study that are broadly applicable to LLMOps implementations: **Customer Emotional State:** As emphasized by both speakers, customers contacting customer service are already frustrated—they're starting from a negative emotional baseline, not neutral. Any AI implementation must account for this reality and prioritize rapid, accurate resolution over technological sophistication. **Culture and Change Management:** Technology transformation takes less time than cultural transformation. The human aspects—mindset changes, process adjustments, training, and adoption—typically represent the longest pole in the tent for any AI implementation. **Co-Creation with Vendors:** While So Energy didn't need a traditional implementation partner, they benefited enormously from co-creating their solution with AWS architects. Organizations should leverage vendor expertise to avoid known pitfalls while maintaining control over implementation. **Build for Tomorrow, Solve for Today:** The platform selection prioritized future-readiness without sacrificing current problem-solving. This balanced approach avoided both the trap of implementing only for current needs (requiring frequent migrations) and the trap of over-engineering for uncertain futures. **Transformation Economics:** Transformations are expensive and disruptive. Choosing platforms that can grow organically with technological advancement—rather than requiring complete replacement every few years—represents sound economic and operational strategy. This case study provides a grounded, practical view of implementing AI-powered contact center capabilities in production. The balanced discussion of challenges, the candid acknowledgment of initial performance dips, and the thoughtful tempering of AI hype make this a valuable reference for organizations considering similar transformations. The focus on foundational capabilities before advanced AI, the emphasis on human-centered design, and the disciplined approach to metrics demonstrate mature LLMOps thinking appropriate for production customer-facing systems.