A leading online travel company operating across Latin America had separate internal HR chatbots for different countries. The bots supported employees with questions about benefits, policies, and labor-related topics.

While their structural logic was similar, each country had different regulations and content requirements, which had led to inconsistencies, duplicated flows, and unclear responses. The goal was to integrate AI capabilities while improving clarity, accuracy, trust, and scalability.

The existing ecosystem was built around predefined intents and structured conversational flows. At a high level, the bots followed a similar logic across countries — but the content behind those experiences couldn't always be treated the same way, because policies, regulations, and HR information differed by market.

This created several problems:

• Similar experiences were being built and maintained separately.
• Some flows were duplicated across countries.
• Similar or overlapping intents could lead to unclear responses.
• Predefined answers had limited flexibility when employees asked in unexpected ways.
• Ambiguous or mixed queries were hard to handle within a rigid intent-based model.

The system needed to preserve the structure and control required for HR information while becoming more flexible in how it handled real employee questions.

As I mapped the existing conversations, I found that some problems were structural rather than purely conversational. A predefined intent architecture works well when the user's question closely matches an expected pattern — but HR questions don't always arrive that way.

An employee might:

• combine more than one question in a single message;
• phrase a known request differently from the examples the system anticipated;
• ask something that falls between two existing intents;
• need information that exists in the knowledge base but isn't covered by a predefined flow.

As the number of countries, intents, and variations grew, simply adding more flows would make the system increasingly hard to maintain. This revealed an opportunity to use AI not as a replacement for the existing architecture, but as a layer that could extend it.

I led the conversational architecture and AI prompt design across the different countries, focused on a system that could be standardized architecturally while still accounting for local differences.

The architecture had to accommodate different regulatory and content requirements while preserving a consistent experience. So the solution separated the global conversational structure from country-specific knowledge and requirements.

This created a scalable framework that could adapt to local needs without reinventing the entire conversational system for every country.

The final system combined five layers that together kept precision while adding flexibility.

Designing for multiple countries taught me that scalability isn't about making one solution identical everywhere. It's about creating a system flexible enough to respect meaningful differences while keeping a coherent experience.

The project also reinforced something I find particularly important in conversational AI:

Adding AI isn't the same as designing an AI experience

The value came from deciding where AI should be used, what information it should access, how it should respond, and where the structured system should remain in control. That balance between flexibility and structure became a central part of the experience.

The project was developed for a leading direct-sales cookware company in Argentina, operating through a network of independent entrepreneurs who sell its products and grow within a structured business model with different performance levels.

Training and education are an important part of that growth journey. The company had an Academy with more than 25 multimedia courses covering sales skills, leadership, and business development. It already had an FAQ chatbot and wanted to evolve it by integrating AI.

The idea was to go beyond answering frequently asked questions and explore how the chatbot could become a more useful way for entrepreneurs to access the company's learning content. Many were already asking Academy-related questions directly in the chatbot, even though the courses lived on a separate desktop platform — an opportunity to connect the two experiences.

Bringing the Academy into the chatbot wasn't simply a matter of making its content available to an AI. The training content was designed for a structured learning platform, while the chatbot needed to respond in a much more immediate and flexible way.

There was also a personalization challenge. Entrepreneurs were at different levels, and their course progress varied. The assistant needed to avoid recommending content that was irrelevant, too advanced, or already completed.

At the same time, the AI needed clear boundaries around what it could answer — staying focused on Academy content while avoiding overly broad responses, unnecessary loops, and excessive token consumption. So the challenge became less about adding AI to a chatbot, and more about how an AI agent could access, understand, and use an existing learning ecosystem in a meaningful way.

The Academy's multimedia courses needed to be transformed before they could become a reliable conversational resource. I transcribed, condensed, and reorganized the content by course and module, creating structured knowledge bases optimized for retrieval and conversational clarity — the foundation for the agent's responses.

But knowledge was only part of the problem. The assistant also needed to understand who was asking. A recommendation that made sense for one entrepreneur might not fit another, depending on their business level or course progress. Personalization couldn't be left entirely to the generative model: the system needed a defined logic connecting the user's question, the relevant knowledge, and the user's context.

I introduced a Conversational Agent Canvas and facilitated workshops to define the target audience, business objectives, use cases, scope, and personality of the agent. That turned the broad request of “we want AI” into a concrete problem: enabling simple, contextual access to Academy content directly within the chat.

From there, I focused on the knowledge architecture, personalization logic, and conversational behavior. I transformed the 25+ courses into AI-ready knowledge bases and worked with the team to define how the agent would connect the user's question with their business level and course progress. I also contributed to validation, testing, and refinement of the prompt logic — the technical implementation was led by a developer, while I defined and refined the conversational behavior and its boundaries.

The architecture followed a simple principle:

The agent identified the topic, retrieved the appropriate course knowledge, and connected through an API to obtain the entrepreneur's business level and course progress — recommending appropriate next steps while avoiding redundant or misaligned content.

We also set boundaries around what the assistant could do. It was intentionally limited to Academy-related content, with topic-based decision rules, hallucination guardrails, escalation handling, and conversational closure strategies to manage unnecessary loops and token consumption.

The agent launched as a pilot and stayed in continuous optimization. The initial evaluation focused on qualitative feedback: the pilot generated positive stakeholder satisfaction and increased engagement with Academy-related inquiries inside the chatbot.

It also helped identify opportunities to refine the decision rules and established a scalable personalization framework for future AI use cases.

Good AI experiences don't begin with the model

They begin with the quality, structure, and purpose of the information we give it. The interesting part wasn't simply bringing AI into an existing chatbot — it was figuring out how knowledge, context, and conversation could work together to help someone move through an existing learning ecosystem in a different way.