1. Make Empathy A System-Level Metric

One of the biggest misconceptions in AI-powered customer service is treating empathy as a surface-level add-on. In reality, empathy has to be engineered into the system itself, Kopardekar explained. “Empathy is not something that is a top layer; it has to come from each case handling pattern and should be part of the system. That means training agentic AI not just to resolve tasks but to interpret sentiment, detect frustration, and adjust its behavior accordingly. A billing dispute or fraud alert, for example, isn’t just a technical issue; it’s an emotionally charged moment that requires clarity, reassurance, and sometimes escalation.”

At scale, automation naturally drifts toward efficiency unless it’s actively corrected. The solution isn’t to slow systems down, but rather to redefine what success looks like. Instead of measuring only containment or speed, leading teams are introducing new evaluation loops, such as: "Did the customer feel understood?" and "Was the resolution appropriate for the emotional context?" and “Did the interaction build or erode trust?”

“To ensure AI responses closely mirror human agents in case handling, it is essential to train models on historical human agent transcripts, including both high and low CSAT interactions. This helps the AI learn not only what good looks like, but also what to avoid,” she shared. “Empathy training should include clear sentiment classification (e.g., frustration, confusion, urgency) paired with an action framework that defines how the agent should respond in each scenario. For example, frustrated users should trigger acknowledgment and faster resolution paths, while confused users should receive simplified, step-by-step guidance. A strong feedback loop is critical. This can be achieved by shadowing human agents, comparing AI vs. human responses, and identifying what drives better outcomes. These insights should continuously feed back into training and prompt refinement. Additionally, providing a structured, step-by-step view for stakeholders enables effective course correction. This transparency helps refine both empathy and spontaneity in AI responses, ensuring the system evolves toward more human-like interactions over time.”

2. Start With Lower Frequency Skills

Not all customer service interactions are created equal. Some happen constantly with low risk. Others have a higher impact. Frequent use cases, such as basic inquiries or routing, help organizations build confidence in agentic systems while improving overall customer satisfaction at scale. These interactions are typically lower risk but high in volume, making them ideal training grounds. Lower-frequency scenarios, like billing errors, fraud alerts, or complex account issues, are where trust is won or lost. These moments are often driven less by speed and more by clarity and emotional intelligence.“ A lower frequency skill if gotten right at all times, can gain a lot of trust and confidence in the product,” she notes. “The catch? These scenarios require stronger guardrails, better evaluation frameworks, and more intentional escalation paths. Automate the easy wins first, but design to scale up towards more complex cases

To be able to handle more complex cases, Kopardekar suggests to train your agent with the simplest, most occurring use cases, with knowledge bases, FAQ’s, and set standard workflows. Then layer in your multi-step workflows by evaluations, advancing the documentation and feeding the human agent's case handling to the AI agent. Every step should be metric-driven. Only once each step is optimized, then the agent will be able to handle multi-intent use cases where it can go through the cycle of evaluations, error analysis and fine-tuning. Shadow testing can be one more step to address the multi-intent case handling efficiency

3. Design AI That Knows When To Hand Off To Humans

The promise of agentic AI is autonomy but it needs to know when to hand off scenarios back to a human. “AI can act independently, but it has to understand and be smart on when to bring in humans for a certain scenario,” Kopardekar says. “This requires building judgment into the system. For example, AI should be able to: recognize when a customer’s frustration exceeds a certain threshold, identify ambiguity or risk in complex workflows and escalate proactively when emotional or financial stakes are high.

At the same time, organizations need visibility into how these decisions are made. Kopardekar emphasizes the importance of a unified journey view, which she describes as a system that tracks not just outcomes, but the entire decision-making process: tool calls, latency, escalation paths, and adherence to workflows. When a handoff is appropriate is completely dependent on the business use case. “A handoff situation like an emergency for a healthcare company is completely different than the handoff to a human agent for a CPG product. Identifying those metrics and having the ability to set the thresholds is the primary key. Metrics like user sentiments (confusion, frustration, etc.), CSAT, handle times, latencies, abandonment probability, resolution of the case, and speed matter in these decision-making trees. All these metrics, individually and together as a cohort, matter in making decisions for human handoff. They should be customizable for the users in their prompts as well as in their metrics for the utmost use case coverage for human handoff needs,” she further explained.

This level of transparency ensures the AI is behaving as intended and it allows teams to continuously refine prompts, guardrails, and cost efficiency. Because while AI can scale quickly, it doesn’t self-correct without oversight, Kopardekar also noted. The biggest opportunity for agentic AI is enabling more personalized, adaptive customer experiences. That means going beyond one-size-fits-all interactions and tailoring responses based on who the customer is and what they need in the moment. Some customers want speed. Others want reassurance. Some need step-by-step guidance, while others just want the answer fast. “Efficiency alone is not success,” Kopardekar said. “Even if an interaction takes more time or effort, if the customer feels understood, supported, and correctly resolved, it is still a win.”

While agentic AI can’t replace human empathy, it can scale it, reinforce it, and, when designed thoughtfully, even enhance it.