Why This Airline Ditched AI Agent Assist: The Surprising Shift from Push to Pull That Actually Works

Contact Center Experiments with Push vs Pull for Agent Assistance

Despite the vendor-led push toward fully automated agentic enterprises, many companies and their contact centers are still testing out "simpler" generative AI-based assistants. In this exclusive Q&A, Jeremy Hyde, senior director of customer service with Sun Country Airlines and president of the WFH Alliance, shares his company's experience with agent assist technology.

The Initial Deployment: A "Push" Approach

Sun Country Airlines deployed an agent assist capability built into their CCaaS platform. In this first iteration, they identified topics agents might need help with and defined specific information the system would present when those topics came up in customer conversations.

This approach was reminiscent of keyword-driven chatbots from before the generative AI era, only slightly smarter. If a topic was detected, the system presented pre-defined information. All content was managed within the CCaaS platform, separate from their main knowledge base, and appeared in the same window as real-time transcription during calls or SMS conversations.

Why They Deployed Agent Assist

Hyde admits part of the motivation was simply gaining access to the functionality through their vendor's AI suite—a "why not try it" approach rather than solving a specific problem. However, they knew that 12–14% of incoming volume resulted in calls to their internal Q&A line, with many questions having answers already in their knowledge base. The goal was to present helpful information to agents in real-time so they wouldn't have to search for it, potentially reducing internal help requests and improving customer experience.

The Results: A Clear Failure

The results were unambiguous: no one was using it. Manual audits showed agents used the presented information about 1% of the time. Overall feedback from agents was that it was more annoying than helpful.

Understanding Their Environment

Sun Country operates in the airline industry, where common customer contacts include flight changes, new bookings, and general travel questions. Their agent team is 100% remote with an average tenure of 3–4 years.

The Shift to "Pull": A Knowledge Base Chatbot

Now, Sun Country is deploying a knowledge base chatbot for agents that exists alongside the initial agent assist. The key difference: agents initiate the chatbot when they need it. It uses generative AI to understand intent and provides direct answers with links to relevant knowledge base articles.

This represents a fundamental shift from "push" to "pull." In the first iteration, they pushed information to agents based on detected topics, whether they needed it or not. With the chatbot, agents decide when they need help.

The chatbot has access to their full knowledge base—hundreds of articles—rather than a limited set of pre-programmed topics. Through generative AI, it understands intent and can reason through scenarios. Hyde's favorite example: if an agent asks whether a customer can bring a Great Dane on a flight, the tool can connect that to dog carrier size requirements in their knowledge base and explain a Great Dane would likely be too large.

Community Insights and Best Practices

After sharing their experience on LinkedIn, Hyde learned that agent assist can work, but design, delivery, and knowledge management make a huge difference. The consensus was that it needs to be:

• Non-intrusive, showing up at the right time without getting in the agent's way
• Relevant, potentially surfacing customer-specific context like recent purchases or previous support calls
• Connected to primary knowledge bases and/or CRM systems, not separate content libraries

Many agreed that the "pull" versus "push" approach was the way to go.

Moving Forward

For now, Sun Country's focus is on getting the chatbot live to learn from the "pull" model. They'll monitor its use, set up feedback loops, and run small group discussions with agents to understand how it's working. While they could build a more advanced "push" model tapping into their full knowledge base, Hyde wants to hear from the team first on whether that would actually be helpful.

The lesson is clear: successful AI implementation in contact centers requires understanding agent workflows and preferences, not just deploying the latest technology.