The global online learning platform transformed its AI Agent, Cora, into a powerful engine of insight across product, support and prioritization.
Most companies measure a chatbot’s success by what it absorbs. Coursera, a leading global online learning platform, measures it by what it reveals. After hitting strong deflection numbers since initially launching its AI Agent, Cora, on the boost.ai platform in 2023, the team made a quieter but more consequential move: treating every conversation as a signal. By classifying conversations by learner type and feeding that data back to the rest of the business, Coursera’s AI Agent has become more than a basic answer-bot, instead acting as a type of early warning system for platform issues, a feedback loop to product teams, and a personalization engine that respects the different types of customers that Coursera serves.
Coursera’s results at a glance…
-
53% deflection rate
-
5 distinct learner types identified and routed by the AI Agent
-
90%+ knowledge base coverage across help center topics
Built for a diverse global learner base
Coursera serves a wide variety of global learners from different backgrounds - including consumer learners, enterprise learners, and degree learners. Each group engages with Coursera for different reasons. A consumer learner might be building a new skill for a career change or for professional growth. An enterprise learner is working through structured training set up by their employer. A degree learner is enrolled in a formal university program with rules and obligations that come with it. Coursera’s support channel sees them all, and serving each one well means treating them differently.
When Coursera launched Cora, the goal was straightforward: scalability through deflection. The initial version of the AI Agent succeeded at this by absorbing a meaningful share of incoming support volume and letting the company grow without proportionally growing its support team. But Bee Strother, Senior Help Content Strategist at Coursera and the project’s initial lead, saw a limit. An AI Agent that treats every learner the same can only do so much. To improve the experience for the people who matter most to specific business outcomes - and to gather data that could inform decisions beyond the support team - Coursera needed Cora to know who it was talking to.
“In order to further improve our performance and health metrics, we worked with boost.ai to develop more advanced user-type filtering” - Bee Strother, Senior Help Content Strategist, Coursera
From deflection to segmentation
The team’s rationale for adding user-type filtering had three parts: building out more sophisticated escalation logic, enabling personalized flows by learner type and unlocking analytics that could be sliced by segment.
To make it work, Coursera built out multiple API integrations against its own account data, enabling the AI Agent to identify five distinct learner types in real time: degree learners, paid consumer learners, unpaid consumer learners, paid enterprise learners and unpaid enterprise learners.
The result is an AI Agent that adjusts its behavior the moment a conversation begins. “Whenever a user interacts with us through Cora, that information is being captured as soon as we say, ‘Hi,’” explains Kean Ramirez, AI Bot Strategist at Coursera. “And that information will dictate how the rest of the conversation goes.”
The mechanism enables a hybrid approach to conversation design. Cora uses Generative Action - boost.ai’s framework for safely bringing generative AI into customer conversations, with built-in guardrails to keep responses on-topic for complex scenarios that need explanation or personalization.
The company’s refund eligibility process is one example as it has many conditions and lots of learner-specific variables, making it well-suited to a Generative Action use case. Subscription history lookups are the opposite: strict and rules-based, where a predefined flow gives the team more control. “I appreciate that boost.ai allows me to use a mixture of both generative and non-generative conversational flows,” Kean says. “It lets the choice be contextual and gives us precision where it matters, and flexibility where it helps.”
What the data unlocks
With learner-type filtering in place, the AI Agent moved from being solely a deflection tool and started becoming a source of strategic insight in several key ways:
It became an early warning system. The AI Agent is often the first place a platform issue shows up. By monitoring conversation trends, the team has flagged dozens of bugs through AI Agent data before they surfaced in other channels, allowing them to prepare human support agents and file engineering tickets more quickly. The same analytics are also used internally to report on the user impact of bugs and outages.
It became a feedback loop. When the company worked to improve Coursera Labs, a type of technical assessment built into some of its courses, the team built out dozens of intents (or topics) within the boost.ai platform to handle the questions learners were asking. The patterns were tracked and fed back to Coursera’s Solutions Engineering team, which used the direct learner feedback to refine the Labs experience. This allowed Cora to essentially become a structured channel for helping to improve Coursera Labs, in addition to being able to answer questions about them.
It got its own metric. In 2024, Coursera developed an internal measure called Chatbot Escalation Impact: the percentage of human agent support volume that originated as a conversation with the AI Agent. It’s treated internally as a health indicator and is currently sitting at ~30%. Coursera uses it to track how often Cora is the starting point for a learner’s support journey, including for the conversations that ultimately need a human.
It became a personalization engine. Cora tells degree learners about late penalties on overdue assignments, directs enterprise learners to their organization’s curated catalog, and routes consumer learners through the standard subscription and refund flows - ultimately shaping what each learner sees.
Beyond the strategic shift, the impact shows in the numbers. Cora maintains a 53% deflection rate, and the knowledge base behind it now covers more than 90% of all help center topics.
Using AI to reshape support
By allowing learners to handle self-serviceable inquiries themselves, Cora is freeing up the company’s human support agents to specialize in the harder, more nuanced cases that genuinely need personalized attention. “Our agents have become more specialized in the support topics that require human assistance,” Bee says. “This has allowed them to resolve the most challenging issues more quickly, since self-servable topics are frequently deflected by Cora.”
The handoff itself is smoother, too. When a conversation does escalate, the support agent inherits the AI Agent’s transcript and can continue the thread without making the learner start over. “I can see that it is saving our support agents time,” Kean says. “They can quickly skim through the conversation and just say, ‘I see that you were asking about this specific issue,’ and pick up from there.”
Working in the boost.ai platform
For Kean, who came onto the project more recently with a background managing chatbots on a different platform, the boost.ai environment marked a clear step up. “The boost.ai platform is very intuitive,” he says. “It’s clearly built so that anyone with limited technical knowledge would be able to create great customer experiences.” he singles out the platform’s in-app conversation review and filtering tools, and the visibility he gets into API integrations, as key differentiators from his previous tools.
He’s also a fan of the onboarding materials. “It’s chef’s kiss,” he says. “The best I’ve seen.”
What’s next for Cora?
Three years on, Coursera’s current focus is exploring the Boost Agent Orchestrator, boost.ai’s newer functionality for coordinating multiple AI agents under a single conversation. Kean is interested in how it could expand the range of questions Cora can handle - particularly the kinds of course-specific questions that fall outside its current scope.
The broader trajectory is consistent: keep the AI Agent evolving, keep it useful and keep the data flowing back to the business.
“boost.ai really cares about delivering products that are not only useful but also cutting-edge - helping people stay up to date and delivering quality interactions that are five, six, seven times ahead.” - Kean Ramirez, AI Bot Strategist, Coursera
