S.03 / CHATBOT
A chatbot for businesses that answers from your documentation and your data, on your website, on WhatsApp or for your team. It says it’s an AI, shows where each answer comes from and hands the conversation to a person when it should.
02 / PROBLEM
A chatbot that confidently answers something it doesn’t know is worse than no chatbot at all.
Most chatbots disappoint for the same reason: they answer in generalities, make up what they can’t find or trap the customer in a loop with no way out. The problem isn’t the AI model. It’s that nobody decided where the answers come from or what happens when there isn’t one.
A reliable assistant is designed the other way round: first the sources, the limits and the way out to a person. Then the model that writes the answer.
03 / SIGNALS
If the answer is already written down and the question repeats every day, there’s an assistant waiting to be built.
04 / FIT
An assistant amplifies the information you already have. If it isn’t written down, that comes first.
05 / WHAT I BUILD
Six ways to put an assistant to work. The channel and the source change; the underlying design stays the same.
An assistant that answers about products, services, lead times and terms using your own information, and links to the page each answer comes from.
EvidenceRocio.com
The same assistant connected to the official WhatsApp Business Platform: it answers after hours, collects the details of the enquiry and leaves the case ready for your team.
Procedures, template contracts, manuals and team FAQs you can query in natural language, with access by role.
Questions about sales, operations or projects answered with queries the server validates. The model never writes SQL or sees credentials.
EvidenceAtalayaIQ
When a question is out of scope, the customer asks for it or something has consequences, the conversation goes to someone on your team with the full context.
Every conversation logs its cost, latency and user rating, so you know what works and what it costs to run.
EvidenceRocio.com
06 / ARCHITECTURE
Six steps. The AI model only takes part in one: the others are what make it trustworthy.
The user writes on your website, on WhatsApp or in an internal tool. The assistant says it’s an AI.
It finds the pieces of documentation or data that answer that question.
The model writes the answer using only what was retrieved. If there’s no basis, it says so.
The answer links to its source, so the user can check it.
Out of scope, or when the user asks, it passes to a person with the context.
Question, sources, cost, latency and rating are stored for review.
07 / EVIDENCE
One public, over an editorial archive, and one internal, over a company’s operational data.
Editorial platform
A public assistant that answers from its own archive, with sources and with every conversation measured.
STACK Nuxt / Vue / MySQL
Case study: Rocio.com
Business intelligence
An internal assistant that answers questions about business data without writing SQL or seeing anything outside its catalogue.
STACK Nuxt / Vue / MySQL
Case study: AtalayaIQ08 / BUILD OR BUY
It doesn’t always need building. This is the line I use to decide.
09 / QUESTIONS
A conversational assistant that answers customers or staff on behalf of the company. The useful ones don’t improvise: they answer from the company’s own documentation and data, say when they don’t know something and hand the conversation to a person when needed.
Yes, through the official WhatsApp Business Platform, which requires a verified number and approved templates for messages the business initiates. The assistant is the same one as on your website; only the channel changes.
Yes. In the European Union, Article 50 of the AI Act requires that people interacting with an AI system are told so, unless it’s obvious. Beyond compliance, it’s worth explaining what it can and can’t answer.
Your AI assistant has to say it’s an AIIt says so and offers to put the user in touch with a person. An assistant that only answers from what it finds in your sources fails less often, and when it fails, it shows. Unanswered questions are logged: they’re the list of what’s missing from your documentation.
Sometimes it is. For simple questions about public information, a platform is faster and cheaper. A custom chatbot makes sense when it answers from your data, integrates with your systems, or you need to control what information leaves and what each conversation costs.
It depends on conversation volume and the model. That’s why the cost of every conversation is measured from day one, as in the Rocio.com assistant, which logs cost, latency and rating for each one.
10 / SAME SYSTEM
A business problem rarely lives in a single piece. These usually appear in the same project.
Tell me what your customers or your team ask, where the answer lives today and which channel the questions come through. I’ll tell you whether an assistant solves it and how I’d build it.