Conversational AI website builder: Speeding up delivery with 2.3x faster milestone completion

Smooth onboarding

Even a platform that democratizes access to conversational AI features can present a steep learning curve for some users. That’s why we had to make getting started with it easy even for business owners who aren’t used to AI-powered features or conversational website builders.

To that end, we carefully removed friction from the registration user flow based on iterative A/B testing. That helped decrease drop-off during this user flow by 14%.

We also designed and implemented a quick platform tour to onboard new users. The tour increased feature adoption by 16% compared to the platform version without it.

No-code AI assistant creation

For business owners, an AI assistant that doesn’t know a thing about the company’s specific policies or products is useless. So, all AI assistants created with the platform use a retrieval-augmented generation (RAG) layer for custom training. OpenAI’s GPT-5, in turn, powers its core capabilities in natural language recognition and generation.

Thanks to the RAG layer, end users can upload their business-specific data to train the AI assistant. We ensured training goes without a hitch by thoroughly testing this functionality. We also brought down the error rate for custom AI assistants to 1.51%.

AI assistants remain highly customizable beyond the training data, too. For example, users can change the greeting message, set the tone, and define specific lead qualification rules.

Conversational website builder

The website builder is the heart and soul of the platform, its most valuable feature. So, it had to work flawlessly. On our end, that meant implementing smooth dynamic rendering, minimizing errors, and enabling customization. Every generated website also had to comply with best practices like mobile responsiveness, design consistency, and fast loading.

Like with any AI tool, ensuring output consistency was a challenge. To overcome it, we expanded the context window, thus helping the chatbot “remember” the conversation and generate context-appropriate output. This meant users could enjoy a more granular control over the generated website’s UI.

At the same time, erroneous output would no doubt frustrate users and lead to churn. So, we tested and fine-tuned the builder until we could guarantee an error rate below 2%. We also introduced numerous guardrails for the model so that it always followed best practices (e.g., mobile responsiveness).

Dynamic rendering and multilingual support

We implemented server-side rendering and dynamic routing to make building a website a smooth and frictionless experience. That brought the average latency to 75 ms, below the target goal. Introducing multilingual support, in turn, expanded the platform's reach beyond English-speaking markets.

Implementing any of these three features could have negatively affected tenant performance. We were well aware of that and, of course, mitigated the impact. As a result, the platform's application performance index (Apdex) score didn't suffer at all.

Multi-tenant hosting

Most small business owners don’t have the time or resources to find the right hosting provider, configure all the settings, deploy the website, and install updates. That’s why out-of-the-box multi-tenant hosting was a must.

Despite its undeniable advantages (e.g., lower infrastructure costs, easier maintenance), multi-tenancy also comes with risks. The “noisy neighbor” effect can cause lag, for one. Plus, security risks are higher; so, every tenant’s data has to be encrypted and protected with access controls.

We addressed these risks with network-level tenant isolation and automatic resource optimization. As a result, response times for end users’ websites remain below 200 ms, even if request rates peak.

Integration with external websites

What if the user already has a website and simply wants to add a conversational AI assistant to it? Of course, this scenario was likely, and we had to account for it.

While adding a script to the HTML code is usually the best option, we realized quickly enough that it might not work for everyone. So, we implemented multiple embedding pathways.

So, the platform can generate a script or an iFrame tag that needs to be pasted into the website’s code. We added precise, clear instructions on how to do it, improving task success by 11% through iterative improvements. We also added public URL embed for edge use cases (e.g., linking to the chatbot on social media) and implemented a single-click integration with popular CMSs (e.g., WordPress).

Future-proof architecture

We built the whole platform as a lean MVP, but that doesn’t mean we paid no mind to our client’s long-term goals. Quite the opposite: we leveraged loose coupling, service-oriented architecture, and horizontal scalability to make sure the platform could easily grow and evolve.

First and foremost, we profiled the initial codebase and removed critical inefficiencies and performance bottlenecks. That alone helped boost performance by 9%, as measured by the Apdex score. Throughout development, we also tested the platform’s scalability and fine-tuned it based on hard data, increasing the scalability limit by 13%.