Meet Maisa, The Spanish AI Start-Up Backed By The Titans Of Big ...

Maisa founders David Villalón and Manuel Romero

Artificial intelligence has a trust issue. For all the hype about generative AI and large language models, many enterprises are reluctant to really embrace the technology. They worry about the well-documented hallucinations LLMs are prone to generating and they dislike the opacity of black-box systems where they can’t check how the model arrived at its results.

Spanish start-up Maisa promises it can deal with these problems. And while this is a world where technologists are prone to making big claims for their innovation, the company has won the support of some impressive backers. It has just raised $5 million of funding from Village Global, the venture capital firm backed by the leaders of big tech, Mark Zuckerberg, Eric Schmidt and Jeff Bezos, which is investing alongside NFX, the Sequoia Scout Program, and several other business angels.

“Our view is that the approach of many AI companies just isn’t going to work because they are asking enterprises to blindly trust their models,” says David Villalón, co-founder and CEO of Maisa. “Moreover, their models given answers based on the most probable result, even when there are big gaps in the data, meaning that it is inevitable that they will sometimes hallucinate.”

Enterprises have rapidly become very wary of these flaws, Villalón says. He points to recent research from Amazon, which found that while 25% of businesses had begun experimenting with generative AI, just 6% had put the technology to use in a live production environment. Even within this group, some businesses have subsequently pulled back from using the technology after running into difficulties.

Villalón and co-founder Manuel Romero therefore set out to build something different at Maisa. “The fundamental difference with us is that we do not use AI to give us answers,” Villalón explains. “We use AI to give us a process to get an answer.”

If that sounds rather nuanced, Villalón is fond of talking about students’ experience at school. “Your maths teacher didn’t just want you to give him an answer to the problem he set; he also wanted you to show how you worked it out, so he could be confident you understood what you were doing.”

In practice, Maisa’s technology – dubbed Vinci KPU – uses existing LLMs such as Google Gemini and Anthropic Claude to carry out complex tasks. But, critically, it takes users through its results step-by-step, so they can see where they came from – and check they’re happy with them. The idea is to deliver a system that is more accurate, but also completely transparent; users can trace and audit every stage of the decision-making process, incorporating the technology into their business workflows while maintaining sight of what it is doing.

For many enterprises, this is vital. For example, Maisa is working with a large financial services firm that had previously pulled back from AI adoption because it felt very uncomfortable making recommendations to clients where it could not explain how it had reached them. Another customer, a large international car manufacturer, is using Maisa to analyse risk in its supply chain, where unforeseen problems – and inaccurate AI results – could bring production to a halt. Elsewhere, an oil and gas company is using Maisa to manage its critical compliance responsibilities in an industry where safety issues can have life and death consequences.

These early adopters of Maisa’s technology – the company was only founded at the beginning of the year – give investors in the business confidence that it can commercialise and scale. The company has also begun accumulating accolades from AI researchers, including benchmarking suggesting that it matches OpenAI’s much-praised new reasoning model o1 on the GPQA Diamond tests.

“The Maisa team are building a transformative technology to turn AI agents into actual workers that are capable of reasoning through complex workflows,” says Max Kilberg of Village Global. “We are very excited to see the new benchmarks and enterprise traction.”

Villalón believes that even as LLMs improve, Maisa’s approach to AI will be widely preferred. “It’s possible we’ll reach a stage where people can put more faith in AI, but that raises the more profound question of whether we really want to surrender control to these models,” he says. “AI can be incredibly good at doing things humans do badly, but very bad at some of the things we do well, so the combination of human input and AI is going to be the way to go.”