How to reinvent the web experience with a conversational AI agent

The era of web-AI: when AI becomes the heart of new user experiences.

The web is undergoing a profound transformation. Once simple, static showcases, websites are now becoming interactive and intelligent platforms. Users are no longer expected to adapt to the site's architecture; the interface must reconfigure itself to provide a smoother, more immediate, and more precise response thanks to AI.

At Smile, we have chosen to seize this technological opportunity to reinvent our visitors' experiences. Our goal is to transform ourselves as we transform our customers.

I wanted to share with you here my feedback as an AI architect on this project: our technological choices, the difficulties and pitfalls to avoid, but also the challenge of data quality and good governance.

Even though innovation is built into our source code, from our business functions to delivery, it was an intense project that wasn't as simple as we had anticipated. With the deployment of our conversational AI Agent, we're breaking with the past: we're moving from simply distributing content to an era of pure interaction.

Meeting the challenge of generative AI

Having an AI agent in 2026 is, in itself, nothing extraordinary. However, deploying true generative intelligence is not something that can be improvised. The engineering behind such a project is not simply a matter of setting up software infrastructure, checking its integrity, and opening the floodgates of information. The ambition goes far beyond mere infrastructure. Our first major challenge was to move beyond the traditional chatbot and its rigid decision trees. We are entering a new dimension: that of generative AI, which is, by nature, organic, and whose responses are built and evolve according to interactions.

Data quality: the foundation of truth

To harness the fluidity of AI, the need for impeccable source data became essential. This rigorous work on data quality is the only safeguard against inaccuracies and hallucinations. Mastering this technology requires technical expertise coupled with ongoing education for stakeholders. But in concrete terms, how did we structure this truth?

Governance: the key to success.  

This is where the major challenge lies. Feeding an artificial intelligence system is certainly not about dumping all of a company's data onto a server! It requires meticulous governance: validating what constitutes a trade secret and what can be made public, ensuring that the technical documentation of our expertise (like that of Alter Way or SensioLabs ) is up-to-date, and purging obsolete data. It is on this precise point that the most intensive internal training effort is required: making teams understand that the relevance and security of our agent depend exclusively on the quality, sorting, and strict validation of the data it is authorized to ingest.

The RAG (Retrieval-Augmented Generation) mechanic : end-to-end mastery.  

Once this data is cleaned, it must be made intelligible to AI. The intelligence of our digital twin is not based on assumptions, but on absolute theoretical and technical mastery of the RAG (Retrieval-Augmented Generation) architecture. To ensure the agent's performance, we deployed and configured the entire data processing chain:

• Chunking and segmentation: Each ingested resource undergoes a methodical segmentation process. We fragment the information into "chunks" (semantic blocks of text) of optimal size, small enough to be precise, but large enough to retain their context.
• Vectorization (embeddings): These fragments are then transformed into mathematical vectors in our databases. This allows the AI to understand not just simple keywords, but the deeper meaning and context of each sentence.
• The retrieve: When a user asks a question, the system performs an ultra-fast vector similarity search to extract (retrieve) only the most relevant "chunks" in relation to the intent of the query.
• Controlled inference: Finally, these fragments of truth are injected into the context of the language model. The generation step (inference) is then under strict control: the LLM writes its final answer based exclusively on the retrieved data, guaranteeing reliable information that is fully grounded in our reality.

RAG: Our reality, everyone's challenge  

Mastering RAG and the underlying data governance is the absolute strategic challenge for all companies in the coming years, and it starts today. At Smile, our conviction is clear: the only way to be legitimate on these critical topics, to speak about them relevantly, and to support our clients in this transformation, is to have done it ourselves. We don't sell theoretical concepts about artificial intelligence; we sell a working technological reality that we have tested and master on a daily basis.

The art of prompt engineering

The second challenge? Prompt engineering. Maintaining harmony between dozens of sometimes contradictory instructions is like a balancing act. One poorly chosen keyword, and the digital twin can veer off course. We have therefore forged ironclad guidelines for:

• Annihilate hallucinations through anchoring : formal prohibition for AI to respond without having previously queried our research databases, safety net in case of lack of data, proposal of a human contact, and absolute locking of redirection URLs.
• Prioritizing information (dynamic vs static data) : teaching the chatbot to cross-reference the static knowledge of its system with fresh news from the site, always giving priority to the most recent data.
• Restrict the scope : lock the context so that the AI remains strictly focused on the Smile ecosystem, without ever straying into external topics.
• Impose linguistic neutrality : require AI to provide instant translation of internal data (mostly in French) while aligning with strict fidelity to the user's native language.
• Safeguarding AI : deploy a real-time security filter tasked with analyzing and instantly blocking any malicious request, whether it be attempts at manipulation (prompt injection), identity theft or code generation.

Mission: to model the collective "brain" of Smile

Our goal was to create a true digital twin with the company's comprehensive memory. For it to be truly relevant, it couldn't rely solely on knowledge confined to its code. This is where RAG architecture comes in. The challenge was to equip the agent with real-time search and analysis capabilities, enabling it to consistently generate the most accurate response possible to Smile's reality.

The real challenge lay in orchestrating the convergence of two distinct but profoundly complementary data universes:

• The pulse of the company (dynamic data): the agent is constantly connected to the continuous flow of our online ecosystem. They capture the latest news, recent figures, and newly published customer case studies. This flow gives them a true "time awareness": they know that breaking news must always take precedence over older data, thus guaranteeing the user absolutely up-to-date information.
• The wealth of expertise (historical data): In parallel, the AI has access to a rich library of raw internal documents, rigorously validated for public use. These more substantial resources (such as our white papers, technology manifestos, and in-depth case studies) ground its responses in the company's history and culture. They provide the perspective and depth necessary to move beyond simple surface information to a genuine demonstration of expertise.

The art of this system lies in the intersection of these two worlds. Before uttering a single sentence, our agent conducts an investigation: he compares the immediacy of web data with the depth of our historical documents.

Thanks to this dual engine, Smile.AI doesn't just guess or generate random text. It delivers a hybrid, sourced, and perfectly calibrated response that accurately reflects Smile's technological, geographical, and commercial reality at any given moment.

Behind the scenes of the project

Building a chatbot doesn't require an army, but a synergy of expertise. This project was led by a team of five people, including a project manager, front-end and back-end developers, an AI architect, and a UX designer, who brought Smile.AI to life in less than two months.

The project is based on Drupal 's historical foundation and web standards (HTML, CSS, JS). Intelligence and security are powered by GCP (Google Cloud Platform) , guaranteeing computing power while protecting the architecture from external manipulation attempts.

Why did we build this first iteration on GCP (Vertex AI)? As architects, the realities of the field often require a strict trade-off between theoretical ideals and time-to-market. Faced with a critical deadline, we opted for Google's integrated solution, which allowed us to move from concept to production in record time. Vertex AI offered immediate robustness and guaranteed the native language performance essential for our international needs (particularly English and Ukrainian), whereas deploying a 100% sovereign solution from scratch would have required an integration and fine-tuning phase incompatible with our timeline.

And what about sovereignty ? That's precisely why we've already started Phase 2. After testing our RAG architecture on this initial scope, we are now taking the time to scale it up by beginning our transition to a sovereign version powered by the Mistral model.

While this development will benefit the website, its true playing field will be internal use. The goal is to infuse this sovereign AI into all of our IT services company's vital processes, transforming our employees into "super-consultants":

• Commercial performance: automated analysis of complex requests for proposals (RFPs) and pre-drafting of commercial proposals based on our database of historical responses.
• Staffing & HR: intelligent matching between the skills required by a mission and the semantic analysis of the CVs of our thousands of experts.
• Legal & administrative: instant compliance analysis of contracts (NDAs, partnerships) and accelerated search in our internal policies.
• Delivery & project management: automatic meeting summaries, report generation and risk detection in project reports.
• Tech capitalisation: creation of an internal search engine of the type "private StackOverflow" allowing our developers to instantly find code snippets validated by our CTO.

LLMOps and Monitoring: Log analysis as a strategic gold mine

Launching a Digital Twin into production is not an end in itself; it's the true starting point. Once our "feline" is released into the website's ecosystem, leaving it unattended is out of the question. This is where LLMOps (Large Language Model Operations) comes into play. Continuous analysis of conversation logs within our GCP architecture isn't just about monitoring server health; it provides us with critical visibility into three major areas:

• Proactive security (GuardRails in action): Our logs allow us to observe the effectiveness of our security filter in real time. By tracking requests classified as "TRIGGER" versus those validated as "OK," we can analyze the behavior of malicious users. We precisely monitor prompt injection attempts , identity theft attempts, and requests aimed at forcing the generation of computer code. This monitoring loop allows us to continuously strengthen the robustness of our system against emerging attacks.
• Business and competitive intelligence: Semantic analysis of queries provides real-time market research. By reading what users ask our AI directly, we capture weak signals and industry trends. Which technologies from our seven core areas of expertise are most in demand? Which business problems arise most frequently? These logs fuel our self-learning : they help us identify future market needs even before they are officially articulated by our clients.
• Continuous improvement of the RAG engine: Logs allow us to immediately identify "holes in the net" of our data governance.

In short, in this era of Web-AI, chat logs are no longer just lines of code for developers. They have become the direct voice of our users, a strategic management tool that informs both our cybersecurity experts and our sales teams.

Results: effectiveness and relevance

The result? A virtual assistant capable of generating structured responses: needs assessment, expert recommendations, a client case study, and access to white papers. We successfully stabilized the AI's behavior while meticulously respecting Smile's geographical and technological realities. Human-machine interaction is no longer limited to clickable navigation; it becomes a high-value conversation.

Perspectives: towards hyper-personalization of the web

This digital twin is just the first building block of the Web-AI era. Tomorrow, powered by the technical mastery of our RAG architecture, the prospects for evolution are immense:

• Hyper-personalization and commercial offers: The interconnection of our RAG engine with real-time data feeds and internal applications will allow the agent to become a truly proactive business generator, capable of automating tasks and commercial offers.
• Talk to Data (structured/unstructured hybrid): the next step is to connect this engine with our complex relational databases within the same ecosystem. Enabling an LLM to orchestrate these simultaneous queries in natural language raises significant challenges in terms of governance, pipeline unification, and access security.
• Self-learning: Careful analysis of queries will help us identify future market needs to continuously refine AI training.
• Internal application: Leveraging the power of the Vertex AI ecosystem paves the way for the creation of internal conversational agents, dedicated to increasing the productivity of our own employees.
• Augmented search engines (ASEs) as a lever for sovereignty: What we have built for ourselves demonstrates our ability to deploy ASEs on hybrid architectures. This directly supports our pragmatic vision of digital sovereignty : it is not about blindly shutting ourselves off "at home," but rather about systematically having the choice to rely on the best solutions on the market, whether open source or proprietary.

By transforming information into conversation, the Smile teams prove once again that innovation is an integral part of our source code. Better still, they demonstrate our deepest conviction: artificial intelligence is not something to be discussed, it's something to be experienced .