Scrolling through social media today is a mission impossible: ads for AI courses promising miracles with portfolio management, pseudo-influencers who pass themselves off as experts to the untrained eye — they buy their blue tick and profit from companies desperate to spend vast sums on spam — and, most infuriating of all, the cheap clickbait of the vibe coders:
“SaaS is dead, I built this incredible app in 3 minutes without touching a single line of code and it’s already in production thanks to {insert name of company that slaps an agent layer on top of commercial models here}.”
Okay, breathe. I would stop reading at the third word faced with such an insult to human intelligence. Not even the bots inflating their post stats would feel proud if they had any awareness of what they were doing.
Is SaaS Dead in the Face of the Inexorable Advance of LLMs?
Is that hype-intoxicated guy — the one trying to sell with a quick demo — capable of building a product equal to or better than one that has required years of development, expert knowledge, and thousands of iterations with customers to adapt to their needs? Let’s unpack this.
Setting up a flashy frontend — a pretty calendar or a personal finance dashboard whose backend, at best, is a simple MongoDB connection with barely any business logic — is not a SaaS. A visual prototype with minimal business logic is not equivalent to a mature SaaS: it lacks the robustness, functional depth, and operational guarantees that become apparent the moment you try to take vibe coding beyond generating a few lines of code.
There is no AI-powered system today capable of creating, from a three-line prompt, a million-dollar software product. So no, you are not going to take down SAP or Salesforce with your megaprompt of “Build me a SaaS that makes me rich and requires no effort on my part”. The only three lines generating those millions are the ones on Sam Altman’s blog every time he announces AGI for the umpteenth time just before releasing a new GPT that still thinks 3.11 is greater than 3.9. But let’s leave them a little hope so that at least the pirates feeding off those exposed API keys in source code can keep making a living.
Agents: Enablement Tools, Not Replacement Tools
The near future of SaaS, in my opinion, is not that anyone builds their own SaaS in five minutes. It points, rather, towards integration with agentic models. With current technology, agent systems do not have the capacity to replace and democratise software development at such a scale; they fulfil more of an enabling function. With a minimal knowledge base, these agents multiply productivity and facilitate the transition of people from an executor role to a higher-value-added supervisor role.
On the product side, it is not about building something new, but about enhancing and readjusting what already exists. Most SaaS products already have an enormous asset given the current state of technology: an extensive and cohesive set of APIs and endpoints in their backend enabled for their services. With good documentation and technical harmonisation work, those APIs can be exposed to the LLM via tools or MCP servers to operate at a higher layer of abstraction, without the need for deep structural changes or custom development, reusing that logic and those tools that already work — and that have taken so much time and money to build.
This greatly simplifies the user experience, as users no longer need to learn intricate processes or the product’s internal jargon; they simply need to ask in natural language with a much more human, even multimodal interaction for the LLM to carry out the entire process internally, automating and relieving users of low-value tasks and focusing their attention on the key aspects of their business, rather than on the multitude of steps and processes they have to perform in the software to get a result.
The revolution that AI will trigger lies in how we relate to products. To use an understandable analogy: it is as if the current state were using the command line to communicate with a computer and AI represented the leap to graphical user interfaces. It is clear that the leap will not be uniform or equally fast across all solutions and sectors — how many systems are still anchored in COBOL — but one reality is undeniable: adapt or become irrelevant.
There will be sectors in which, for reasons of security, reliability, scalability, or performance, adoption will be postponed, but there is no doubt that sooner or later the push of demand will force the change. Laggards will lose market share to those who integrate contextualised AI; the new barrier to entry will be the accumulated quality of data, process experience, and structured knowledge.
The Interface of the Future: A Blank Canvas
While on the backend this amounts to small adjustments to what already exists to enhance it, the challenge is enormous on the frontend: this is where the key lies, the evolution towards something unknown and volatile.
In my opinion, the transition must go from detailed, modular interfaces loaded with shortcuts and functions, to a blank canvas where, this time, AI builds forms of interaction that adapt to what the user is looking for. We move from that linking and execution function on the backend to a creative and substitutive role in current frontend practice. AI becomes responsible for generating screens on demand, replacing rigid, predefined, deterministic interfaces. Building a dashboard, a loading screen, an editor… whatever it may be: AI is very good at generating components on the fly, in the moment. Although one must always keep in mind that they are built on top of existing backend functionalities and that they must receive precise instructions in the system prompt to preserve uniformity, quality standards, and minimum safeguards (policy engine, component validation).
The Company Must Change: Knowledge Is No Longer Just for Humans
For the agentic system to gain autonomy and capability, the corporate philosophy must change. Functional analyses, development documents, and almost any other information must not be conceived solely for human reading: everything must be oriented so that it is readable and processable by AI. Format, references, the storage system, and coherence all matter. For AI to interact correctly with users, it must be supplied with an enormous amount of information and an effort must be made to capture the knowledge held in the company’s human and social capital.
This is a change that involves the entire company, not just technical teams: from customer support to management, through sales, product, and operations.
The Key with LLMs Is Obvious: CONTEXT
It is essential to condense, structure, and expose knowledge to models via well-defined tools so that they incorporate it when generating responses or actions. We tend to overvalue the “power” of the model and underestimate context, when what is decisive for obtaining reliable results and reducing hallucinations is precisely the context. In complex products with multiple facets and tacit knowledge dispersed in undocumented human and social capital, it is difficult for a generative AI to execute functional or technical product tasks in a coordinated, scalable, and reliable manner. Furthermore, current limitations — context window and attention capacity — impose an operational limit on the size and complexity that can be addressed.
Hence the need to structure that knowledge and promote a change in the philosophy governing how the entire organisation acts and what it aims for, and why it is not so easy to build a SaaS that meets your needs with a simple prompt.
Agents in the Development Cycle: From Meeting to Git
The change does not only apply to the product side — the what — but also to the methods — the how. Agents or more complex agent systems, with good context and instructions, can play a crucial role in the productivity, resource optimisation, and streamlining of development processes to ship new features or maintain existing ones.
Let me present a use case inspired by an episode of one of my favourite podcasts, which I highly recommend if you enjoy these topics: El Test de Turing. It is an example that Víctor shared about Claude Code, while being bombarded with questions by Arnau about the differences with Cursor as Álvaro tried to hold him back — it made me rethink the potential of multi-agent systems in software development.
Instead of the traditional cycle (requirements gathering, brainstorming, documentation, meetings with the development team, validation, architecture, task assignment, implementation, supervision, testing, and documentation), we can start from the recording of a meeting or a requirements and architecture document to activate an agent system that works in a coordinated manner on versioned Git branches. The development team’s role shifts to final review and spot adjustments, delegating also the technical and functional documentation. The process is as follows — an agent system with the following phases or actors:
- Ingestion of requirements and architecture information (audio / transcript / document).
- Requirements extraction agent.
- Normalisation, coordination, and dependency detection agent.
- Task generation agent.
- Triage agent that assigns those tasks to specialist agents in very specific niches and technologies (because having one backend agent and one frontend agent is not enough — as if it were 2015).
- Code generation, testing, and technical documentation by each of these agents.
- Static validation, policy engine, and security by the coordinator agent.
- Generation of executive summaries and changelogs.
- Final human review and merge.
- Deploy and feedback to the context repository.
Result: the steps of the traditional development cycle are simplified and shortened, making it possible to deploy new features in hours or days.
2025: The Year of Agents… Where Are They?
2025 was supposed to be the year of agents and, although advances in model capabilities have been astonishing when viewed in perspective, we are still only beginning to glimpse part of their true potential. Tools such as Claude Code or Gemini CLI are starting to show the direction the technology seems to be taking, but there is not yet a sufficient level of maturity to guarantee business-grade consistency and quality. Even so, at a time when the slope of the exponential curve is steepening, it is difficult to anticipate even next month’s situation, let alone project it one or several years ahead.
Everyone Talks About AI… But Few Actually Adopt It
For now there is a lot of talk about AI — on LinkedIn or X it seems to be a thematic monopoly — but few companies incorporate it structurally. Adopting AI is not about setting up a RAG chatbot fed by three QA PDFs that only appears useful, like those old banking assistants that got in the way and nobody understood. Being a general-purpose technology, it will sooner or later spread across the productive fabric just as the cloud or the internet did, but for now the data are clear on this point: the degree of adoption is low and there is considerable room for improvement and opportunity. I also think we will witness a return of the Renaissance ideal in the field of human resources: transversal profiles capable of integrating business and technology, with critical thinking and a holistic vision, will start to be valued more highly than pure specialists or executors.
This Is Not the End of SaaS. It Is a Transition
Therefore, no: we are not witnessing the death of SaaS. We are in a transition comparable to what the internet represented for commerce. Even in the face of AGI or ASI, these systems would still be, at bottom, similar to humans: they need context and knowledge to carry out their tasks well.
That expertise, customer fit, and niche development will be the real engine of SaaS companies — far more than pure technological capability.
LLMs do not eliminate SaaS: they shift value towards context quality, orchestration, and dynamic experience. Those who win will be the ones who structure knowledge before the adoption curve accelerates.