## Overview Anthology is an education technology company specializing in higher education solutions, including learning management systems, back office operations, and student success services. Their Student Success division operates essentially as a Business Process Outsourcing (BPO) provider for educational institutions, supporting students outside the classroom with services related to admissions, financial aid, enrollment, and general student support. With over 1,000 employees handling 8 million inbound and outbound interactions annually, Anthology serves as a critical support infrastructure for higher education institutions that lack the scale to manage these operations internally. The case study centers on Richa Batra, Senior Vice President of the Student Success division, who spent over five years advocating for a contact center transformation. The initiative began as a basic lift-and-shift migration from homegrown CRM and on-premise contact center infrastructure to the cloud, but evolved into a comprehensive AI-first transformation once GenAI capabilities became viable. The project was approved in August 2023 and went live in July 2024, deliberately timed before their most critical peak period (the back-to-school summer season). ## Business Context and Drivers The transformation was driven by several interconnected business challenges. First, Anthology faced extreme seasonality, with interaction volumes spiking dramatically during July, August, September, and January when students return to school. This required them to hire over 1,000 seasonal agents annually, effectively doubling their workforce during peak periods. This seasonal hiring pattern was disruptive to innovation and service quality, as much of the organization's energy was devoted to simply "getting through" the peak periods rather than strategic improvements. Second, their legacy infrastructure suffered from significant reliability issues. During their busiest months, they experienced 12 unplanned outages, causing major operational disruptions and requiring manual failover procedures. These outages were particularly problematic because they occurred when students needed support most urgently—for example, when a student needs a financial aid question resolved before they can register for classes. Third, the nature of student inquiries was highly repetitive. Approximately 20 core questions accounted for the vast majority of interactions, with password resets and login assistance being the top request. Richa Batra noted that such seemingly minor issues could have major consequences—a student unable to reset their password might decide to "take a break" from a class and potentially never return. Fourth, consistency and accuracy were ongoing challenges with a heavily human-staffed model. New agent training was resource-intensive, and with high seasonal hiring volumes, ensuring consistent responses across all agents was difficult. Additionally, agent attrition created continuous training burdens. The transformation vision evolved significantly over the five-year planning period. Initially conceived as infrastructure modernization (moving from on-premise to cloud), it ultimately became a business model transformation enabled by AI. Richa presented the business case to three different CEOs, three CFOs, and three different boards before finally receiving approval, noting that the breakthrough came when they repositioned the initiative from a "lift and shift" to a transformative vision focused on business outcomes rather than technology features. ## Solution Architecture and Implementation Approach Anthology selected Amazon Connect as their cloud-based contact center platform after their CEO, who had previous experience with AWS, suggested exploring AWS solutions. The engagement began with discovery workshops led by AWS Solutions Architect Kathy Hofstettler, who worked closely with the Anthology team to understand their current architecture, operational metrics, and desired future state. The implementation strategy was methodical and phased rather than attempting a big-bang deployment. Originally, Anthology planned to go live with all capabilities simultaneously, but the team pivoted to a release-based approach: - **Release 1 (July 1, 2024)**: Core Amazon Connect capabilities went live at the start of peak season - **Release 2 (July 22, 2024)**: Additional AI capabilities introduced - **Subsequent releases**: Continued through the year with Release 4 scheduled for December 2024 This phased approach allowed the team to learn from each release and incorporate newly available AWS features as they were launched. The iterative framework became embedded in Anthology's operational culture, enabling continuous optimization rather than treating the implementation as a one-time project. The technical architecture leverages multiple Amazon Connect capabilities: **AI Virtual Agents for Self-Service**: Students interact with AI-powered conversational agents via voice and chat channels. The AI assistant greets students, outlines available services, and handles verification through student ID and Social Security number validation. The virtual agents can resolve common queries like refund status, password resets, and registration questions. The natural language processing capabilities enable fluid conversations rather than rigid menu-driven interactions. **Seamless Human Handoff**: When AI agents cannot fully resolve a query or when students request human assistance, the interaction seamlessly transfers to a live agent. Critically, the AI interaction context is summarized and provided to the human agent, eliminating the need for students to repeat information. The agent immediately knows why the student is calling and what has already been discussed. **Agent Assist Capabilities**: Human agents receive real-time AI assistance during interactions. The system provides step-by-step guides relevant to the specific inquiry (for example, detailed procedures for handling refund status questions). It also offers AI-generated guidance on how to respond based on the conversation context, displayed in a side panel within the agent interface. This is particularly valuable for new agents—one agent in the demonstration was handling calls on his first day with no prior contact center experience and minimal financial aid knowledge, yet was able to successfully resolve student issues using the AI assistance. **Contact Lens Analytics**: Amazon Connect's Contact Lens provides AI-powered analytics and quality assurance capabilities. It automatically generates interaction summaries, eliminating the need for supervisors to manually listen to call recordings or review chat transcripts. The system performs sentiment analysis tracking how customer sentiment evolves throughout an interaction (for example, from "concern and distress" to satisfaction after resolution). It also provides full transcripts with metadata including call duration, agent identity, and timestamps. **Automated and Hybrid Evaluations**: The platform enables 100% automated agent performance evaluations using AI analysis of interactions, compared to the previous manual approach that could only review 1% of interactions. The system supports both fully automated evaluations and hybrid evaluations where supervisors can add their own observations or modify AI-generated scorecards, combining AI efficiency with human expertise. ## LLMOps and Production Considerations Several aspects of this implementation demonstrate mature LLMOps practices and considerations for running AI in production contact center environments: **Unified Infrastructure**: A notable architectural decision is that the same infrastructure used to build AI agents for customer self-service is also leveraged to provide agent assist capabilities. This unified approach simplifies development, deployment, and measurement processes, as the team uses consistent tools and frameworks across different use cases. **Model Selection and Optimization**: During implementation, AWS ProServe team recommended specific ML models based on Anthology's business objectives. This consultative approach to model selection rather than one-size-fits-all deployment was credited with helping achieve the 14-point accuracy improvement. The case suggests ongoing model evaluation and optimization as new releases incorporate different or updated models. **Accuracy as a Key Metric**: The team set a goal of over 90% accuracy from user acceptance testing (UAT) through production. They measure accuracy of AI agent responses and track improvements over time. The 14-point accuracy increase achieved represents a significant operational improvement—incorrect responses that would previously require coaching and retraining of human agents are now systematically addressed through AI refinement. **Real-Time Performance Monitoring**: The system provides hourly and daily visibility into key performance indicators including wait times, handle times, average speed to answer, containment rates (what percentage of interactions are resolved by AI without human escalation), and accuracy metrics. This operational observability is essential for running AI agents in production. **Proactive Issue Detection**: Contact Lens enables pattern detection across all interactions. In one example, when institutions Anthology supports were being targeted by threats, they used Contact Lens to identify patterns across chat and voice interactions and provide information to local authorities within hours—work that would have taken weeks manually. This demonstrates how the analytics capabilities support both quality assurance and security/risk management. **Data-Driven Optimization**: The team uses historical data analysis to inform ongoing improvements. Kathy Hofstettler described how they compare pre-migration data with post-migration performance to understand the student journey and optimize containment rates. This data-driven approach to iteration is fundamental to improving AI agent performance over time. **Scalability and Reliability**: The migration to cloud infrastructure addressed critical production concerns around uptime and ability to handle variable load. Reducing unplanned outages from 12 to 2 during peak months demonstrates improved reliability. The system successfully handled the extreme seasonal spike when going live July 1st with no downtime—described as "flipping a switch" with immediate operational readiness. **Change Management Integration**: The case emphasizes that technology deployment alone is insufficient without organizational change management. Richa conducted weekly calls with clients and internal teams throughout the implementation to communicate goals, share performance metrics, and allow time for questions. Recognizing that different stakeholders move through the change curve at different paces, particularly around AI concerns about job elimination, was treated as integral to successful production deployment. **Phased Rollout Strategy**: The decision to move from big-bang deployment to phased releases represents a mature production deployment approach. Each release provides learning opportunities and allows incorporation of newly available AWS features. This reduces risk while maximizing impact and ensuring adoption. **Integration of Professional Services**: The implementation leveraged AWS Professional Services and partner organizations using an agile delivery framework. This "one team" approach brought together Amazon Connect specialists for technical deep dives, the Amazon Connect service team, and solutions architects working collaboratively with Anthology's team. This demonstrates the importance of expertise and support for complex AI deployments. ## Business Outcomes and Impact The transformation delivered measurable improvements across multiple dimensions: **Wait Time Reduction**: 50% reduction in wait times during the busiest month of the year, with zero client escalations regarding wait times since go-live. This directly addresses a major source of student frustration and potential dropout risk. **Accuracy Improvement**: 14-point increase in response accuracy compared to human-only operations. The AI consistency advantage means errors that would require individualized coaching and retraining are systematically eliminated. **Attrition Reduction**: 10% reduction in agent attrition immediately following go-live. Lower attrition reduces training costs and improves service consistency, as more experienced agents remain with the organization. **Reliability Improvement**: Unplanned outages during peak months reduced from 12 to 2, dramatically reducing operational disruption and eliminating the need for manual failover procedures. **Quality Assurance Scaling**: Capability to review 100% of interactions compared to previous 1% review rate. This enables comprehensive quality management, trend identification, and proactive issue resolution. **Workforce Planning**: Early indicators suggest potential to reduce reliance on seasonal hiring, moving toward a more stable workforce model. This would improve service consistency and reduce the organizational disruption of managing massive hiring and training waves. **Strategic Capacity**: By automating reactive work, the leadership team and agents gained "mind space" to be strategic and proactive rather than constantly firefighting operational challenges. This cultural shift enables continuous improvement and innovation. Richa Batra noted that these results exceeded even optimistic expectations, describing them as "beyond the baseline of what we even expected" and enabling the team to "think bigger" about future capabilities. ## Critical Assessment and Balanced Perspective While the case study presents impressive results, several considerations warrant balanced assessment: The presentation occurs at an AWS conference (re:Invent) and is explicitly promotional for Amazon Connect. The speakers include AWS employees and a customer who has obviously had a positive experience, so the narrative is inherently favorable. Independent validation of the claimed metrics would strengthen confidence in the results. The timeline from approval (August 2023) to go-live (July 2024) is relatively short—approximately 11 months—for such a significant transformation. While presented as a success, this aggressive timeline also represents risk. The fact that they went live immediately before their highest-volume period (rather than during a quieter period for safer testing) was bold but potentially exposed the organization to significant risk if issues had emerged. The case study focuses heavily on operational metrics (wait times, accuracy, attrition) but provides less detail on student satisfaction or educational outcomes. While reduced wait times presumably improve student experience, direct measurement of student satisfaction or whether the transformation actually improves enrollment and retention (Anthology's stated mission) is not provided. The specific accuracy measurement methodology is not detailed. What constitutes "accuracy" in this context—correct information provided, successful resolution of the inquiry, or something else? How is this measured—through subsequent escalations, quality reviews, or student feedback? More transparency on measurement methodology would strengthen the case. The 14-point accuracy improvement is compared to "human agents," but the baseline is not specified. If pre-migration accuracy was already very high (say 85%), improving to near 100% is impressive. If baseline accuracy was lower, the improvement is less remarkable. Similarly, the specific containment rates (what percentage of interactions are fully resolved by AI) are not provided, making it difficult to assess how much human agent workload was actually reduced. The presentation mentions that accuracy goals were "over 90% from UAT," and they saw a "14 point increase," which could imply they achieved 104% (impossible) or that the 90% was an aspirational target and the 14-point increase was measured from a different baseline. This ambiguity suggests potential confusion in how metrics are being communicated or measured. The case emphasizes using "the same infrastructure" for both self-service AI agents and agent assist, positioned as an advantage. While consistency is valuable, this could also mean missing opportunities to use specialized tools optimized for different use cases. The presentation doesn't address whether they evaluated alternative approaches for agent assist specifically. Change management is appropriately emphasized, but the presentation focuses on leadership's perspective (Richa's weekly calls, explaining the vision) rather than agent or student perspectives. How do agents feel about working alongside AI? How do students perceive AI interactions versus human support? These stakeholder perspectives would provide a more complete picture. The cost dimension is notably absent. While the presentation mentions Amazon Connect's usage-based pricing model as advantageous, actual cost comparisons to their previous infrastructure are not provided. Given that cost management was one of the original drivers (reducing seasonal hiring), understanding whether the financial case was realized would be valuable. The presentation mentions that some results were "unplanned" or "bonus" outcomes (like the attrition reduction), which suggests the initial business case may not have fully anticipated all benefits. While discovering unexpected positive outcomes is good, it also raises questions about how thoroughly the business case was analyzed during the five-year planning period. ## Future Direction and Strategic Vision Looking forward, Richa Batra articulated a vision of proactive student engagement rather than reactive support. The goal is to leverage data to reach out to students before they need help—for example, proactively informing a student that their financial aid application is due in 60 days and providing the three things they need to do. This represents a significant shift from transactional support to anticipatory guidance. This proactive vision aligns well with Anthology's mission to increase enrollment and retention. Research in higher education consistently shows that proactive intervention improves student outcomes, particularly for first-generation and at-risk students. By combining their knowledge of critical milestones (semester start dates, financial aid deadlines, registration periods) with student-specific data, they could theoretically reduce friction points that lead to student attrition. The ongoing release schedule (with Release 4 in December 2024 and presumably more to follow) demonstrates a commitment to continuous improvement rather than treating the implementation as complete. This iterative approach is essential for AI systems that improve through ongoing refinement based on production data and feedback. The case also hints at broader organizational transformation beyond just technology. Richa's comment about moving from reactive to proactive culture, with teams now having the capacity to be strategic, suggests the transformation is enabling organizational evolution beyond operational efficiency. ## Conclusion This case study represents a substantive example of deploying conversational AI and LLM-based capabilities in a production contact center environment with significant scale (8 million interactions annually) and high stakes (student education outcomes). The implementation demonstrates several LLMOps best practices including phased deployment, unified infrastructure, data-driven optimization, comprehensive monitoring, and integration of human expertise alongside AI capabilities. The measurable results—particularly the 50% wait time reduction and ability to scale quality assurance from 1% to 100% of interactions—indicate genuine operational improvements. The successful go-live during peak season without downtime represents a significant technical accomplishment and speaks to the maturity of both the technology and implementation approach. However, readers should interpret the case with appropriate context regarding its promotional nature and the limitations in independently verifiable details. The transformation's long-term success will ultimately be measured not just by operational metrics but by whether it achieves Anthology's core mission of improving student enrollment and retention—an outcome that will take longer to fully assess. The vision of proactive student engagement represents an ambitious next chapter that could demonstrate whether AI in contact centers can move beyond efficiency to truly transform service delivery models.