The Challenge

Prex, a leading digital payments company operating across Uruguay, Argentina, Chile, and Peru, needed a customer support solution that could handle high volumes of inquiries efficiently: from general FAQs to account-specific questions, ticket management, and escalation to human agents when necessary. In fintech, support quality is part of the product: slow or inaccurate responses directly erode customer trust.

The Solution

Marvik designed and implemented Prex Atenea, an AI-powered customer support assistant built on a multi-stage architecture. Rather than relying on a single, generic AI interaction, the system distributes intelligence across specialized stages: each handling a specific part of the conversation lifecycle, from understanding customer intent to formatting the final response.

At the core of the system is Claude 3 Haiku (Anthropic), which is one of the core components of the system. Claude 3 Haiku was selected for its strong instruction-following capabilities, low-latency responses suited for real-time conversations, and a cost profile that supports multiple AI calls per customer interaction at scale.

What the System Does

Prex Atenea handles the full support journey:

• Understands and routes customer inquiries to the right workflow
• Retrieves relevant knowledge and information to generate personalized responses
• Manages support ticket creation, retrieval, and follow-up
• Detects when a customer is frustrated and escalates seamlessly to a live agent, transferring full context so the customer never has to repeat themselves
• Maintains consistent tone and brand voice across every response

Results

Prex Atenea was evaluated on a test dataset of 508 real customer questions provided by the Prex team, benchmarking the system against production-representative conditions rather than synthetic scenarios.

• 86% of questions received a response (coverage)
• 83% of those responses were correct or partially correct (accuracy)
• <9 seconds average response time

Prex's support channel handles an estimated 600,000 messages per month. Applying the benchmark results to that volume:

• ~516,000 messages per month would receive an automated response.
• ~428,000 messages per month (≈71% of total volume) would be addressed with a correct or partial answer and no human intervention.
• ~14,000 messages per day would be taken off the human support queue. Agents could then focus on the remaining ~29% of cases, which are the complex, sensitive, or high-value interactions.
• Customers would get answers in under 9 seconds, around the clock, instead of waiting in a human support queue.

These results show that advanced AI can be applied in fintech environments without sacrificing speed, accuracy, or cost control. At Prex's volume, even a modest improvement in automated resolution translates directly into hundreds of thousands of interactions per month handled at a fraction of the cost of human support.

For Marvik and Anthropic, this case reinforces a key principle: effective AI deployments are defined not by a single model call, but by thoughtful orchestration, with the right model assigned to the right task, measured against real-world data at meaningful scale.