Which AI startups will survive the boom?

Governance of intelligent agents, control of costs and private data, as well as knowledge of the customer context are essential to building solid projects that can grow and be sustained beyond the AI ​​’boom’. In the midst of the AI ​​and intelligent agent boom, with thousands of applications, tools, and proposals, many are already wondering which AI companies will manage to survive. And regarding the new ones that are built, what essential elements will they need to incorporate to grow and differentiate themselves from the competition ? Finding the right sector in which to focus the proposal is another vital issue for making the idea a success… and convincing the market . Miguel Planas , founder of the venture builder Iris, is clear about the ingredients of the formula for success: good AI governance, cost control and knowledge of the context . The company, which Planas has led since 2010, identifies opportunities and, from there, helps to build teams and companies mainly in the fields of AI, cybersecurity and data. With the experience of having launched 15 companies , the founder of Iris has acquired a solid understanding of which tools and technologies make the most sense in each sector and time. Currently, faced with the explosion of AI, he doesn’t hesitate to state that ” nine out of ten AI companies aren’t AI companies.” He believes that “they have nothing to do with this technology beyond using a model as an assistant or adding an agent to a product.” According to Planas, the future will not come from proposals for AI used as an assistant to workers, “but from corporate projects as agents that operate the company’s processes, systems and decisions .” Based on this idea of ​​AI created from within corporations, Planas points to corporate governance and cost control as two essential factors for new projects to work. A success story highlights Axium , one of the projects built on Iris, which has focused on a proposal to measure and control how much an agent consumes in each area of ​​the company, assigning a cost to whoever generates it. “Its technology sets limits and automatically cuts off usage the moment it spikes, in addition to governing employee use of AI to prevent information leaks and its use in the shadows,” Planas points out. For Sonia Fernández , a partner at the Kibo Ventures fund , costs and governance are also essential issues. The investor is pessimistic, as she believes that “there is still great uncertainty about business models and real margins.” Corporations, Fernández maintains, “are starting to take it seriously, but in many of them, the agents are being deployed faster than the mechanisms to supervise them. An agent acts, makes a chain of decisions, and its cost can skyrocket without anyone noticing until the bill arrives.” Thus, governance cannot be an added layer at the end of processes; it must be integrated from the design stage. The investor from Kibo Ventures sees this as an interesting opportunity for new startups that emerge and specialize in observability, cost control, and agent security. The power lies in the data. For González, another key to founding successful projects lies in knowing how to focus your efforts: “Today, the value is not so much in building generalist models that only large companies can finance, but in applying AI to specific problems where there are complex workflows, proprietary data, and clients willing to pay for specific use cases that generate measurable results.” Particularly interesting are the “companies that can add value by owning their data, curating and enriching this data to train models and providing differentiation in their own information.” Corporate governance and cost are key areas of focus for NeuralTrust , a company that provides security infrastructure to guarantee the protection and control of AI agents. “Awareness of the need to control AI agents is growing, but unevenly,” emphasizes Joan Vendrell , co-founder of the company. “Concern about cost has come earlier, often as a result of unexpected bills. An agent can make dozens of queries to the model to complete a single task, and that consumption is difficult to predict if it isn’t measured from the beginning,” argues the founder, who previously led the deployment and security of AI at Mango. At the textile group, Vendrell saw the existing shortcomings in this area and the need to promote greater security in the deployment of AI and agents in companies than was available at the time. This concern motivated him to found NeuralTrust in 2023 with his partners Víctor García—who was part of his team at Mango—and Alejandro Domingo. The company’s clients include companies such as Air Europa, Abanca, Iberia, and Sabadell . In governance, awareness is lower, and this is what most concerns companies like NeuralTrust. Vendrell points out that “an agent queries data, operates on internal systems, and executes actions on behalf of the company, and many organizations still don’t know how many agents they have in operation or what permissions they have.” There have been several news reports in recent months about incidents involving AI agents gaining unauthorized access to systems and data belonging to companies and public administrations. Even with this ever-present risk, the founder emphasizes, “there is still a significant gap between the level of concern and the controls that are actually in place.” In his opinion, “ governing an agent requires technical controls that intervene every time it performs an action, with limited permissions, a record of everything it does, and the ability to stop it at any time.” To avoid risks, NeuralTrust recommends using an AI Gateway: “A layer through which all communication between agents, models, and tools passes. It allows for cost control measures and, above all, provides complete visibility into how AI is used.” The value of context Understanding how to value and utilize context is another essential element for an AI project to succeed. “Enterprise AI only works if it understands the company: its data, its systems, and the decisions made,” Planas points out. With this knowledge, startups that know how to organize, structure, and contextualize corporate information will have a bright future. The example of Velorum —another project from the Iris incubator—is representative: an AI startup that connects to companies’ systems (their code, databases, tools, etc.) and builds a “living map” of everything within that organization and how it all relates. This map answers questions in seconds and anticipates when there’s a failure, providing context for agents to work with real-world knowledge of the companies. Oriol Juncosa , co-founder of the Plus Partners fund , also values ​​the importance of context, stating that “AI companies must be highly relevant to the customer and tailored to their needs by providing a lot of context about their sector.” The context and adjustment of the agents to the reality of the company and its workflows must be done with the so-called forward deployed engineers, as Juncosa points out: “They are engineers who spend weeks in the client’s offices to fully understand the operations and internal processes of the organization.” In this sense, the investor sees HappyRobot as a clear example of how a project, before implementing its AI assistants in companies, understands and studies the needs and processes of each organization and, from that starting point, carries out the deployment of the product. Juncosa also sees survival opportunities in AI companies that can “bridge the gap between the physical and virtual worlds, which are known as physical AI projects.” These are proposals, the investor points out, that build applications and agents on top of large language models and aim to solve problems in sectors such as construction, engineering, real estate, and industry. Knowing how to connect Mario Garcés is the founder of The Mindkind , a deep tech company that develops generative AI based on understanding, modeling and enhancing the human mind by integrating neuroscience, artificial intelligence and behavioral science. From this perspective, the entrepreneur believes that ” the projects built around AI that will endure will be those that bring new technology that competitors do not have, because they know how to connect and create new things at a technological level in a very distinctive way .” Data ownership is another major asset for Garcés: “There’s an opportunity for projects that have private data, data that isn’t accessible to other competitors, and that know how to process it, even with the same technology as everyone else.” He emphasizes that data “has intrinsic value and can be leveraged to great advantage.” The ability to monetize and offer something different with unconventional models, adapted to new market dynamics, will be another key element. Solving difficult problems The eVoost project , led by Cristian Pastrana and Koh Onozawa from Abu Dhabi, has found its purpose in offering an AI and predictive analytics tool to market homes completely automatically and according to the needs of each buyer. It is precisely in the application of vertical AI in specific sectors, such as real estate and many others, where Iñaki Fuentes , the company’s technology director, sees good opportunities to build projects with a secure future. Fuentes believes that ” it makes more sense to talk about companies that use AI to solve difficult problems than about AI startups .” He doesn’t consider technology alone to be a sufficient business category. The chief technology officer of eVoost sees the application of AI as particularly interesting in areas with a lot of specific knowledge, fragmented data, still manual processes, and decisions with a significant economic impact, such as health, financial services, legal, industry, logistics, travel, construction… But Fuentes clarifies that this is not because these areas need another chatbot, “but because there is an opportunity to transform scattered information into actionable decisions.” At eVoost, they are carrying out this transformation in the residential sector. “There is a lot of data on the market, prices, inventory, and buyer behavior. Having this data doesn’t necessarily mean knowing what to build and at what price. The opportunity lies in moving from having information to being able to make better decisions with it,” he emphasizes. For him, there’s a very simple test: ” If we remove the words ‘artificial intelligence’ from the presentation, is there still an interesting company? If the answer is no, we have a problem .” Navigating between laws Lexroom ‘s proposal has found its particular field of action in law. This generative AI platform , created for legal professionals, optimizes legal research and the analysis of large volumes of documents . Martina Domenicali , co-founder of the company, insists on the idea that “the best opportunities are found in solutions designed to solve specific problems in a sector, rather than offering generalist tools.” The entrepreneur believes that the projects that will survive “ must have a product-market fit .” Another key aspect is that “you have to build a competitive advantage that is difficult to replicate, based on quality data, specialized knowledge, and technology that is truly integrated with customers.” The real challenge for Domenicali is not launching an AI startup, but building a solid company even when the enthusiasm for this technology has subsided. Joaquín Cuenca , co-founder and CEO of the image and video generation platform Magnific (formerly Freepik ), argues that for a project to succeed, “you have to choose the right battleground: find a large and interesting market that isn’t the target of the big players. Google has been making tools like Word or Excel for millions of people for years, but it has never tried to make a Photoshop.” Cuenca adds another essential factor for success: “ It is important to have clients who pay for reliability, because in the professional market you compete on trust, support and long-term relationships .”