Serasa transforms debt renegotiation with generative AI

With millions of consumers accessing its digital channels daily, Serasa needed to modernize the debt renegotiation experience to make it simpler, more fluid and more effective. The previous model was supported by a traditional chatbot, with pre-defined flows and rigid menus, which generated friction throughout the journey and directly impacted the channel's strategic indicators.

Despite good adoption, the experience still had important limitations. The journey was fixed, without adaptation to the user's context, without continuity between stages and without a real understanding of each person's intention. Instead of supporting the resolution of the problem, the flow required the consumer to navigate between multiple options, identify on their own which debt made the most sense to negotiate and follow all the steps until concluding the agreement.

In practice, the user needed to deal with an experience that was inflexible and excessively operational.

It was necessary to go through several paths within the conversation, interpret offers, installments and payment methods without contextual support and move forward in a linear logic, often far from the real need at that moment. This format increased abandonment of the journey, increased transshipment for human service and limited the channel's scaling capacity.

To face this scenario, MadeinWeb developed Agente Negociador Limpa Nome, a cloud-native generative AI solution integrated with Serasa's WhatsApp. The proposal was to replace the rigid logic of the previous model with an intelligent operation, capable of understanding natural language, interpreting context, conducting negotiations actively and closing end-to-end agreements with security and scalability.

The transformation represented a clear change in approach. A model based on rules, pre-defined responses and low adaptability emerges. Enter a result-oriented conversational architecture, in which the journey stops being a rigid sequence of steps and starts to function as a contextual, continuous and more human interaction.

The solution was designed as an ecosystem of collaborative agents, in which each agent assumes a specific role within the journey.

• Orchestration Agent

Conducts the interaction in a natural, empathetic and conversion-oriented way, coordinating the journey from start to finish.

• Authentication Agent

Validates the user in the company's systems Serasa and guarantees secure access to the information necessary for negotiation.

• Debt Negotiation Agent

Conducts the negotiation itself, identifies the most appropriate conditions and supports the closing of the agreement.

• Information Agent

Consults and organizes relevant data about the client, offers, accounts and context of the negotiation to support decision-making decision.

• Shared Memory

It works as a layer of context that preserves continuity, consistency and progression throughout the conversation.

With this structure, the experience is no longer just transactional and becomes oriented towards the effective resolution of the user's problem.

The solution architecture was designed to operate with high scale, low latency, resilience and governance. In the case published by MadeinWeb, the technological base includes Amazon EKS as the main computing layer, Amazon Bedrock for access to multiple language models, LangGraph for conversational orchestration, Limpa Nome APIs for querying and recording negotiations, in addition to SQS, AWS Lambda and Redis for decoupling, state control and robustness operational.

This technological foundation enables a more intelligent experience and at the same time prepared to absorb large volumes of interaction without proportional growth of the operation.

After the launch of the Agente Negociador Limpa Nome, Madeinweb also took over the continuous support of the solution in production, under a managed services model. This work includes monitoring the performance and scalability of the AWS architecture; incident management and corrective improvements; oversight of the integrations with the Limpa Nome APIs; observability of components such as Amazon EKS, Amazon Bedrock, AWS Lambda, Amazon SQS and Redis; as well as the continuous evolution of conversational flows, prompts, guardrails and governance mechanisms of the AI agent.

This operating model allows Serasa to keep the digital negotiation channel running continuously, with proactive monitoring of quality, security, cost, latency, resilience and consumer experience, turning the solution into a living and scalable operation rather than a one-off project delivery.