From September 16 to 23, 2026, Barcelona became a meeting point for companies, developers, startups and people working hands-on with artificial intelligence as part of AI Week. The closing event and one of the main highlights was the AI Summit Barcelona, held on September 22 and 23 at the World Trade Center Barcelona. More than 200 international speakers were announced, alongside over 50 side events and activities across the city.
The conversations surrounding the event were just as interesting: What already works in everyday business use? What is still at the prototype stage? What does it mean to be found by AI? And which of today’s big visions could actually shape the way we work with software and the internet in a few years?
Table of Contents
The question is no longer whether AI will matter
By now, hardly anyone seriously disputes that artificial intelligence will play an important role over the coming years.
The more interesting question is:
What exactly will that change look like, and how quickly will it happen?
Expectations are certainly enormous. Catalonia alone presented its AI Strategy 2030 at the AI Summit, with planned investments of around one billion euros. At the same time, major technology companies around the world are investing billions in data centers, models and infrastructure.
Capital markets also make it very clear how much is expected from this technology.
What makes this especially interesting is the contradiction.
While investment levels are enormous, many companies operating directly in the AI sector are still not generating sustainable profits, or they continue to invest significantly more than they currently earn.
That does not necessarily mean expectations are exaggerated.
But it does show that we are currently in a phase where expectations about the future and economic reality are still, in some cases, quite far apart.
That is exactly why I find the practical perspective so interesting.
Not every demo will become a business model. Not every prototype will succeed. And not every vision of how we are supposed to work five years from now will actually become reality.
But some of these prototypes already give us a fairly clear idea of where the technology could be heading.
From the first AI agents to systems that can actually take action
A good example is AI agents.
Back in 2023, projects such as AutoGPT and BabyAGI became popular. The idea was already remarkable at the time: instead of simply answering one question, a language model would be given a goal, plan intermediate steps on its own and perform different actions in order to achieve it.
In practice, these early systems were still very limited.
They sometimes got stuck in loops, made mistakes and required a great deal of supervision.
Even so, they introduced a much wider audience to an idea that has since become central to AI development:
AI does not only have to generate content. It can also perform tasks.
A lot has changed since then.
In January 2025, OpenAI introduced Operator, for example, an agent capable of using websites through its own browser environment. It could click, scroll and fill out forms.
In July 2025, these capabilities became part of the ChatGPT agent.
Developers have also gained access to more and more tools that allow agents to combine web search, files, computer control and external functions.
In March 2025, OpenAI released, among other things, its agent tools and the Agents SDK.
Only a few years separate the experimental agents of 2023 from the systems we have today.
That is exactly what makes the current phase so interesting.
Where does the demo end and real-world use begin?
At an AI event, you inevitably see a lot of visions.
Autonomous agents. AI systems that take over entire business processes. Software that uses other software on its own. Assistants that organize appointments, research information, place orders and prepare decisions.
For many of these ideas, working prototypes already exist.
That does not automatically mean these systems will completely transform every company tomorrow.
There is still a long way between a working prototype and a reliable production system.
Issues such as security, error rates, costs — as you would expect in a world shaped by economic logic — but also data protection, interfaces and responsibility all play a role.
That difference is far more interesting to me than the question of which new model scores a few percentage points better on some benchmark.
The real question is:
What actually works under real-world conditions, and for what purpose?
From The New York Times to renewable energy
Some of the conversations I had during the event fit into this picture very well.
One particularly interesting exchange was with a developer working in QA at The New York Times.
We talked about how AI is changing development and testing processes, and where it can already support or accelerate work today.
The interesting part was that the conversation was no longer simply about AI generating code.
It becomes much more relevant when AI becomes part of an existing workflow, processes information from multiple sources and works together with systems that are already in place.
A completely different perspective came from a project manager from Toronto, with whom I discussed the use of AI in the renewable energy sector.
Before the release of Claude Code, he had never developed software himself.
Today, he is working on a platform designed to evaluate information, structure projects and prepare decisions.
This is interesting on two levels when thinking about the impact of AI.
First, it shows how significantly the barrier to building software has fallen, making entirely new projects possible.
Second, many of those new projects themselves include AI.
That creates an additional multiplier effect in the current development.
Things get interesting when AI does more than answer
For me, this is one of the most important aspects of the current development.
ChatGPT and other systems are still often viewed mainly as better search engines or text generators.
In the long term, however, their role is likely to become much broader, including in everyday life.
And that can begin with something completely ordinary.
A user might ask today:
“Which hotel in Barcelona is close to the beach and offers a vegan breakfast?”
An AI gathers the relevant information and suggests suitable options.
The next step would then be:
“Book one of them for Friday.”
At that point, it is no longer enough for an AI to simply read a website.
It has to be able to understand and interact with it.
That is exactly why I find developments such as the Model Context Protocol (MCP) and similar interfaces so interesting.
They provide a technical way for AI systems not only to retrieve information, but also to work with applications and data sources.
The line between searching, answering and acting is beginning to blur.
This also changes the role of websites
That also changes what a good website needs to be able to do.
Of course, design, usability and traditional search engine optimization will remain important.
But an additional layer is emerging.
It is no longer only search engines that need to understand what a website offers.
Information from websites is increasingly being used outside the website itself and in a far more extensive way than was ever the case with a classic search result.
For knowledge databases such as Wikipedia, this already plays a huge role.
But commercial websites should also be able to clearly answer questions such as:
- What exactly does this company offer?
- Who is the product or service intended for?
- Where is it available?
- What requirements or limitations are there?
- Which common questions can be answered?
- How can someone make contact, submit an inquiry or make a booking?
At first glance, this sounds obvious.
In practice, however, there are many websites where even this basic information can only be inferred indirectly from marketing copy, several subpages or scattered pieces of text.
From SEO to an additional AI layer
That is also why I do not believe traditional SEO has suddenly become obsolete.
Many of the fundamentals remain the same.
Clear page structures, understandable content, internal linking, structured data and technically accessible pages help both search engines and AI systems.
Google itself has pointed this out as well.
But an additional question is now becoming more important:
What specific questions can this website actually answer?
This is exactly where I am working with TRAILS
Many of these ideas are directly shaping the development of TRAILS.
TRAILS analyzes websites to determine which information is already available in a clear and understandable form, and where answers are still missing.
This includes, for example:
- content and potential answer gaps
- frequently asked customer questions
- page and heading structures
- structured data
- internal linking
- crawlability
- technical information for AI systems
- ways for AI to interact directly with a website
The final point is becoming especially interesting to me.
My long-term goal is not only for a website to be readable by AI.
It should also be understandable and usable by AI.
An AI could, for example, recognize what service is being offered, retrieve relevant information and prepare an inquiry, appointment or another process.
That is the core idea behind TRAILS.
More about TRAILS:
https://trails.team/
Which parts of this will become the future remains to be seen
Perhaps the most important takeaway for me from Barcelona was the need to distinguish between three things, despite the speed of current developments:
what already works reliably today, what exists as a functioning prototype, and what is still mainly a vision.
The boundaries between these categories are shifting surprisingly quickly.
The first popular autonomous agents appeared in 2023.
By 2025, AI agents were already capable of operating websites on their own.
Today, the focus is increasingly shifting toward connecting agents directly with business software, databases and other services.
How quickly this becomes a new standard is difficult to predict seriously.
But the speed of development and the scale of investment at least show how high expectations are.
It is precisely this gap between technical possibility, economic expectation and practical reality that makes the current development so interesting to me.
Barcelona was more of a snapshot of where we are today
I have already signed up for additional events and specialist conferences around AI and want to continue exchanging ideas with developers, companies and people from other industries.
For TRAILS, this mainly means not chasing every short-term AI trend.
TRAILS is explicitly not meant to contribute to AI slop, but to prove its value through practical use.
For me, it is clear that people will increasingly ask their AI systems directly in the future.
That immediately raises another question:
Who will be able to provide the answer?
This creates a new requirement for websites.
AI has to be able to find and understand the answer accurately.
And if agents continue to develop at the same pace as they have over the past few years, a second requirement will probably follow quite quickly.
AI also needs to be able to do something with that information.