I was on a farm in the Ambalá-Calambeo corridor collecting field data — counts of flora, coordinates for the trap map, behavior notes in hives. The phone had no signal, the laptop was put away. Six hours of clean work.
When I returned to the motorcycle and regained connection, there were three WhatsApp messages about portfolio services. Questions that the site should have been able to answer. It didn’t. The three contacts had already continued searching.
That day I started documenting how many times the same thing happened. The result in four weeks: eight lost inquiries that my page could not attend to because my page was a static file waiting for me to come and work.
The Diagnosis: A Digital Brochure with Great Presentation
The previous portfolio had good design, well-written texts, and a logical structure. It was also completely useless without me.
If a visitor arrived at midnight with a specific question about the automation service — what it included, how long it took, if it was suitable for their type of operation — the site showed them text. The visitor had to read, interpret, deduce, and if they still had doubts, wait for me to respond. In that time, 70% made another decision.
The problem was not the design. It was the underlying architecture: built to showcase, not to operate.
Most business websites have the same problem. It’s not an issue of aesthetics or copy — it’s a system problem. The site exists to satisfy the owner when they see it on the computer, not to resolve the visitor's doubts when the owner is not present.
What Replaced the Portfolio: Five Operational Principles
1. The site organizes itself based on who arrives
I have three lines of work that don’t have much in common for someone from the outside: automation and data, beekeeping and rural projects, carpentry and space design. If I show the hives to someone who arrived looking for an ERP, I lose them in ten seconds. If I show code to someone looking for furniture design, the same.
From the first screen, the site asks the visitor where they are coming from. By choosing their profile — a company seeking operational efficiency, an investor evaluating projects, a self-taught person looking for resources — the system hides what does not concern them and brings to the forefront what their brain came to find.
It’s not decorative personalization. It’s an information architecture decision that reduced the time visitors spend searching for what to do on the site.
2. Data instead of adjectives
"Experts in sustainability" doesn’t tell anyone anything. "30% reduction in water consumption in a technified irrigation project, 12-hectare farm, Tolima" tells something specific to someone who has that problem.
The projects section does not have narrative text about my capabilities. It has panels with real metrics: percentages of savings, number of installed nodes, documented returns, implementation times. If the data is not measurable or I cannot verify it, I do not publish it.
This has a cost: there are projects I cannot show because the client did not authorize the numbers. In those cases, the sheet says exactly that — "data under confidentiality agreement" — instead of inventing a vague description.
3. An assistant that doesn’t invent answers
I integrated a virtual assistant directly connected to my databases of services, prices, and availability. The difference with a generic AI chatbot is that this one does not have access to general knowledge — only to the real, updated catalog.
If someone asks how much the automation service with n8n costs, it extracts the current price, the estimated implementation time, and the conditions. If they ask something not in the database, it says it doesn’t have that information and suggests the direct contact channel.
The second part was difficult to calibrate. The first prototypes of the assistant answered questions outside its knowledge base with well-written generic information. It seemed useful but was risky: if the assistant invented a price or a condition, I ended up committed to something I hadn’t agreed upon. The current system has an explicit domain limit — it only answers what it knows for sure.
4. Behavioral nudges — reducing friction at the decision point
There is a specific moment when visitors abandon a service site: when they perceive that the next step is difficult, costly, or uncertain. In my case, it was the large projects section — someone interested in business automation saw the scope and concluded that it was out of their budget without even asking.
The site detects that behavior pattern and offers a shortcut: not the complete project, but a modular entry version — a diagnosis, a pilot flow, a downloadable resource. The visitor does not leave frustrated; they find a lower-friction door to the same destination.
This is not manipulation — it’s eliminating the gap between interest and action. The nudge does not push towards something the visitor does not want; it removes the obstacle that prevented them from reaching where they already wanted to go.
5. What this system does not solve
This is the most important principle and the one that makes people uncomfortable when they read it.
Automation and AI do not perform miracles with disorganized processes. If your customer service process is chaotic, the system automates that chaos — faster and on a larger scale. If your business model has holes in financial logic, no dashboard will cover them; it will only make them more visible.
I make this explicit on the site because I learned the difference at my own expense. I entered projects where the client expected technology to solve problems that were process or organizational culture issues. Those projects end badly for both sides.
Before any diagnosis, the site asks the visitor three concrete things about how their operation works today. If the answers indicate that there is a process problem to solve first, I say so before quoting.
The Result After Six Months in Production
Three metrics I decided to track from the beginning:
- Inquiries answered without direct intervention: I went from 0% to approximately 65% of initial questions attended by the assistant without my intervention.
- Time between inquiry and first response: from an average of 4 hours (the time I checked my phone) to less than 2 minutes.
- Quality of leads that reach diagnosis: prospects who schedule a call now arrive with context — they know what I offer, have seen the reference prices, understand the process. The diagnostic call takes 30 minutes instead of 90.
What did not improve: the conversion rate of cold visitors to prospects. That number still depends on how traffic arrives at the site and whether the problem they have corresponds to what I offer. The system does not generate demand where none exists — it filters and attends to what already exists.
What I Would Do Differently If I Started Today
The assistant took three iterations to have a clear domain limit. In the first versions, the model responded well — too well — beyond its knowledge base. I would start with the restrictive limit from day one and expand it with evidence, not the other way around.
The filtering system by profile also assumed that visitors know which of the three profiles they belong to. In practice, a relevant percentage does not identify with any clear category — they are hybrids. That required a fourth option that was not in the original design: "I am still exploring."
What This Implies for Your Business
A direct question before closing: can your website answer the five most frequently asked questions you receive via WhatsApp, at 10 PM on a Tuesday?
If the answer is no, you have a system that depends on you to function. That is not necessarily bad — there are businesses where human presence is the differentiator and automating it would be a mistake. But if what prevents you from responding at 10 PM is not a strategic decision but simply that you are not available, that is an architectural problem that has a solution.
Do you have a process that you have already mapped and that could be automated? That’s where I would start the conversation.
Related Nodes:
- How the assistant connected to real databases works (without hallucinations)
- AI orchestration protocol — the stack that runs behind the site
- Nudge design: the five behavioral patterns I used and which ones worked
Por qué destruí mi portafolio para construir un "Ecosistema Vivo"