Building a software product used to require an engineering team and weeks of effort. It can now be done by one person with an AI assistant in an afternoon. The tools driving that shift — Claude Code, Cursor, Replit, and Lovable, to name a few — have given rise to a practice known as vibe coding.
“Vibe coding is using AI, large language models, to do coding for you and interacting with the large language models instead of really spending much time looking at the actual code,” says Professor Mattan Griffel, who teaches the subject at Columbia Business School.
But if the model writes the software, what's left for the manager?
“The key skill here really is management,” Griffel says. Vibe coding, he adds, is “bringing out a new form of management and leadership in most organizations” — one that gives companies “superpowers to build products and technology where before they didn't have the resources to do that.”
What you can — and can’t — hand off to AI
For most of the history of software, three jobs traveled together: writing the code, designing the system it ran on, and deciding which business problem it should solve. Vibe coding pulls them apart. The agents take the typing. The other two jobs land on whoever is directing them — and Griffel argues they get harder, not easier, as the pace picks up.
“Now engineers who used to be individual contributors are essentially managers of five to 10 vibe coding agents that are running simultaneously,” he says. “Non-technical managers can now manage technical AI agents.”
That job demands the managerial instinct of knowing what can be handed off cleanly and what can't. The test is specificity. Vibe coding is best where the objective is clearly defined — where you can name the input, the output, and exactly what the tool needs to do.
When the cost of being wrong is severe, though, the human in the loop becomes important.
“There are some mission-critical things, such as sending a space shuttle to the moon, where you're probably going to want to have a human involved in every step because it really has to work,” Griffel says.
Vibe coding also struggles when given work that hasn't been defined yet. An agent builds exactly what you specify, within the constraints you give it, which makes it a poor collaborator when you don't yet know what you're building.
Within those limits, Griffel thinks the opportunity is nearly universal — and mostly unglamorous. Almost everyone, he says, has five to 10 recurring weekly tasks that vibe coding could make faster or more consistent. Think of the status email someone assembles every Friday by pulling the week's numbers out of a reporting dashboard by hand — a script can fetch the figures and draft the message. Sometimes the win is time saved. Sometimes it's the same time spent to greater effect: “Maybe something that you spent 30 minutes on, you still spend 30 minutes on, but because you're using a vibe coding tool that you built, you're able to accomplish two to three times as much,” he says.
What the payoff looks like differs from job to job and firm to firm. There is no one-size-fits-all answer — which is precisely why the practice has to be learned by doing.
Learning to manage the machines
That conviction is the premise of the School’s vibe coding course. It assumes no technical background and runs as a lab: weekly build sprints, side-by-side tool evaluations, and a class-wide demo day. The point, Griffel says, is to “get reps under your belt.” Students leave with several working products, including web apps and workflow automations.
Students are evaluated on how they worked with the AI to get there, and are held to a risk-aware explanation of the code they ship. It amounts to a management assessment.
The biggest gains, though, are waiting outside tech. Nonprofits, brick-and-mortar retail, consumer packaged goods — slower-moving organizations that tech firms' years of head start haven't reached — stand to benefit most, and have the fewest people equipped to deliver it, according to Griffel.
“There's so much value out there that we need somebody who understands the business problems, can take a leadership role, and knows the technology well enough to actually be able to bridge the gap and unlock that value,” he says. “And I think Columbia Business School students are best positioned to be able to do that.”
