There’s a specific moment when most generative AI projects hit a snag. It’s not when the technology fails—it almost always works. It’s a few weeks after the kick-off, when the initial enthusiasm gives way to the crucial question: How can we turn AI’s potential into a concrete, shared, and sustainable path forward?
It’s a question of methodology. And the numbers show just how significant it is. According to S&P Global Market Intelligence, which surveyed more than 1,000 companies in North America and Europe in 2025, 42% of organizations have abandoned most of their AI projects in the past year—more than double the figure from 2024.
It’s not a technological problem. It’s a matter of approach. And the answer has a specific name: Agile —the project management methodology introduced in 2001 with the Agile Manifesto, which replaces rigid, sequential planning with short, iterative cycles, continuous feedback, and close collaboration within the team. And, as we’ll see, it’s also a matter of people.
When uncertainty is the rule, not the exception
Generative AI projects have one characteristic that sets them apart from almost everything else: no one knows exactly where they’ll end up. Not because of a lack of expertise, but because the very nature of these projects is exploratory. You start with a hypothesis, test it, gather feedback, and adjust course. Above all, you refine the initial data.
Traditional project management methods—Gantt charts, milestones, and six-month delivery schedules—work well when we already know what we need to do. But when the destination is yet to be discovered, planning everything from the start is like building on sand. In these contexts, the problem isn’t the lack of a plan—it’s placing too much trust in the plan.
Agile was created precisely to fill this gap, and it turns the logic on its head: instead of planning everything first and then executing, you execute in order to learn what to plan.
The Minimum Viable Product as an Organizational Principle
The agile approach works in short, iterative, and measurable increments. Instead of designing the entire system and then delivering it, you build something that works in just a few weeks—the minimum viable product —have real users test it, gather feedback, and decide how to proceed.
This fundamentally changes the way we think about errors. In a traditional project, discovering after six months that the solution doesn’t work is a failure. In an agile project, discovering it after three weeks is a learning opportunity. The error becomes part of the process, not the end of the process. Amazon and Spotify have been doing this for years: they release features, observe user behavior, and make corrections. Not because they don’t know how to do it any other way, but because it’s the most effective way to navigate uncertainty at high speed.
Agile is about teams, even more than it is about processes
However, there is one aspect of Agile that is often overlooked when it is reduced to sprints and boards: Agile is, first and foremost, a way of building a team and fostering teamwork. It is no coincidence that the Agile Manifesto begins by placing individuals and interactions before processes and tools.
The ceremonies that mark each sprint—planning, daily, review, retrospective—are not bureaucratic rituals: they are structured discussions. They serve to bring people with different perspectives together to address the same problem at regular intervals, forcing them to look at what has been built and tell each other what works and what doesn’t. It is from this dialogue that the best ideas emerge and the most profound conclusions are drawn. The method does not create value on its own; it creates the conditions for people to create value together.
AI performs at its best when it becomes part of the team
And here’s the point that directly concerns generative AI. An experiment conducted by researchers from Harvard and Wharton in collaboration with Procter & Gamble— “The Cybernetic Teammate, 2025”—involved 776 professionals tackling real-world product innovation challenges, randomly assigned to work alone or in teams, with or without AI.
The first finding made headlines: individuals using AI matched the performance of teams working without it. This led to the dangerous temptation to conclude that AI makes teams unnecessary. But the most interesting finding is the other one: the truly excellent solutions—those in the top 10 percent—came most frequently from teams working with AI. Not from individuals enhanced by AI, but from groups.
This explanation is consistent with everything Agile has taught us: AI used in isolation tends to bring everyone toward a good average; collaboration among people, amplified by AI, takes us further. The study also observed a second effect: AI breaks down functional silos. With AI, technical staff produced proposals that were more attuned to the market, and sales staff produced proposals that were technically sounder—a common language that enriches, rather than impoverishes, team discussions.
This also addresses one of the most widespread fears about AI. When used as a personal shortcut, AI risks isolating and stifling creativity. Used within a team—like an extra colleague around the table—it does the opposite: it nurtures precisely what is most human about work—dialogue, dissent, and the shared development of ideas. It’s a way of adopting AI that doesn’t marginalize people, but puts them back at the center.
A two-phase process
When looking at how organizations approach generative AI projects, a consistent pattern emerges: the process consists of two distinct phases, which rarely follow the same approach.
The first isstrategic exploration: understanding what you really want from AI, which processes make sense to transform, and which business objectives to pursue. At this stage, the ideas aren’t clear yet—and that’s as it should be. This is the time for Design Thinking and exploring possibilities. Organizations that rush through this phase often find themselves developing technically sound solutions to problems that no one had actually prioritized.
The second phase isimplementation: bringing something concrete to life, observing how it’s actually used, and letting it evolve. And this is where the Agile methodology truly shines. Not only because short cycles and continuous feedback transform change from a disruption to the plan into useful information, but because it brings people together as they work: each iteration teaches the team something that the previous one couldn’t yet know.
The real risk isn’t making mistakes; it’s not learning
The spread of Agile, after all, is now widespread: nearly all organizations claim to adopt its practices in some form. But claiming to do so and truly mastering it are two different things, and the vast majority do not reach a level of maturity sufficient to generate real value from it. The gap isn’t between those who know Agile and those who don’t. It’s between those who use it as a label and those who use it as a method—and as a team culture.
For generative AI projects, this gap comes at a specific cost: investing in the technology without establishing the right methodological and human framework means greatly increasing the likelihood of ending up as another abandoned project.
In a world where technology evolves faster than plans and skills, competitive advantage does not lie in the technology itself. It lies in the speed with which an organization—made up of people who collaborate—learns to use it together.
Twenty-five centuries ago,in *Alcibiades I*, Plato had Socrates say that the soul knows itself only by reflecting on something similar to itself: another soul, contained within another living being.
We never think as clearly as when we engage with someone who is similar enough to us to understand us, yet different enough to surprise us. Agile has turned this insight into a method. And AI—if we choose to use it within the team rather than in place of the team—adds another mirror around the table—as long as we remember that the deepest reflection remains the one we find in another person’s gaze.
