In the past, entrepreneurs who wanted to launch an online product would either have to learn to code themselves or outsource that work to developers. But AI agentic coding means that anyone with a subscription to Claude Code, Codex, Cursor, or Replit can now “vibe code” their idea into functional software. At the moment, this has mostly allowed for the emergence of more competitive, small companies and increased the output of developers. But if you could easily generate your own software, why would you—or your company—pay for Adobe, Salesforce, or any other software-as-a-service?
Under the scenario laid out in the Citrini essay, DoorDash and Uber Eats are gutted by crafty entrepreneurs who code their own competing platforms and undercut the dominant food delivery platforms with significantly lower fees. And you could go even further. Many people are already using tools like OpenClaw as personal AI assistants—giving the AI access to their email addresses, credit cards, messaging apps, and so on. At the moment, if you ask an AI agent to order you tacos, it will place the order through DoorDash. But as they get more capable, couldn’t they skip the platform (and its fees) and order directly from the best restaurant?
At market close Monday, DoorDash fell from $176.60 per share to $164.50 per share, and remains down 2 percent from its price before the essay’s release. DoorDash co-founder Andy Fang on Monday tweeted a response to the Citrini essay’s warning of agentic commerce, acknowledging, “we’re gonna need to adapt to it,” but emphasizing that the company was already doing so, focusing on optimizing the platform both for human users and for AI agents shopping on users’ behalf. “That’s why we’re so focused on the agentic commerce tools we build for ourselves and our merchant partners.”
The DoorDash example was among the most controversial online, as its competitive advantage wasn’t its unique code for a food ordering app, but rather its massive partnership network with restaurants and grocery store chains, its driver network, and its customer base. Those relationships and network effects are—to use a term coined by hosts of the business podcast TBPN—“unsloppable”; AI can’t generate them for you. The same is true of various payment rails, of businesses that rely on a physical component, or of businesses that face regulatory constraints or capture. And when you order from DoorDash, you have confidence that you won’t get scammed, that you will get your food, and that issues will be resolved if they do come up.
The code itself is also less easily generated than is popularly believed. AI writing is still generally simple and uninspired, and that’s true whether it’s writing in English or C#. Marco Iansiti—a professor of business administration at Harvard Business School, where he is the co-chair of the technology and operations management unit—told TMD, “You can put in a crappy prompt into [an AI agentic coding platform], and this thing will crank out an application that sort of works and does the things that it does.
But if you want to create something that’s actually complex and well designed, I do think that there is still actually a role to human engineering.” Iansiti explained, “I don’t think you’re going to go and tell the AI to solve all the world’s problems, and all of a sudden they’ll all get solved, right? … There’s an evolution of the tools and technologies and platforms that we’ve had over the years that it’s going to continue to evolve.” But, he added, “Is it going to get rid of engineering? I don’t think so.”
The Citrini essay scenario detailed that sector-wide shifts from human to AI labor cascaded into a “negative feedback loop,” in which companies would lay off staff, deploy AI programs, and use the additional revenue to optimize AI tools to take on further work, allowing companies to fire more workers, and on and on. This would allow firms to produce more for less, subsequently increasing the national gross domestic product (GDP), but the essay calls these gains “Ghost GDP” because it “never circulates through the real economy.”
But this concept defies basic economics, Alexander Arnon, director of policy analysis at the Penn Wharton Budget Model, explained: “I think it just sort of doesn’t really make sense to have this concept of ghost GDP. If GDP is being produced, then someone is getting the income.” While he noted that this could lead to concerns about wealth distribution, he told TMD, “If AI overlords are raking it in all of a sudden, they’re going to be doing something with that money.”
Daron Acemoglu, an economics professor at the Massachusetts Institute of Technology and recipient of the 2024 Nobel Prize in Economics, explained, “AI, as it’s presented to us right now, is an automation technology, meaning that it will save labor costs,” adding that if it can be successfully integrated across business workflows, then many firms will be able to scale back expenses by cutting human workers and investing most of their salaries into AI.
That would result in productivity benefits, but as Acemoglu told TMD, “The question then becomes, what happens to those workers who are laid off or who do not get a job in this company? … There is no guarantee in the market process that good quality, high paying, new jobs will be created to replace those that are taken by automation, whether that automation is led by robots or by AI.” But he added, market actors “also use new technology in order to create new jobs, new tasks, new ways of doing things that workers can fulfill,” so the productive gains may translate into new jobs for some workers, depending on how employers deploy the tech.
James Pethokoukis, a senior fellow at the American Enterprise Institute and author of the Faster, Please! newsletter, noted that, historically, technological innovations that led to automated work have not led to apocalyptic economic consequences. “It’s easy to create destruction, and destruction is a lot easier to imagine than creation, though history suggests there’ll be more creation than destruction,” he told TMD, noting that “there will still be things that machines can’t do, and there’s gonna be things that people aren’t going to want machines to do.”
As a working paper published by the Brookings Institution in November found, “increasing automation moves workers from the intelligence sector to the physical sector under mild conditions.” One of the paper’s co-authors, Ioana Marinescu, a labor economist at the University of Pennsylvania and a research associate at NBER, told TMD that positions at risk of becoming automated are “jobs that you could do 100 percent online.” The work AI won’t be able to substitute are physical sector roles, which require workers to be on-site but don’t necessarily involve manual labor, such as teaching.
In turn, these physical sector jobs “could experience an indirect productivity boost, as in, they’re using better tools, and so they become more productive,” she said. “And therefore wages could very likely rise, which is what we show in our model.” Marinescu added, “If people lose a high-wage job [because of AI automation] and they transition to what today has been a low-wage job, that doesn’t mean that this lower-wage job couldn’t eventually have higher wages after we adjust for the general impact of AI on the economy.”
As more companies choose to integrate AI, some roles will become obsolete, harming workers with skills that AI can perform at a lower price. Iansiti noted that EZ-Pass erased the need for toll booth operators—but because the job largely involved menial tasks and long shifts in both the summer heat and winter cold, few people lamented its demise. “For the most part, I think having the creation of more interesting, more humane jobs is also a good thing that we should be aspiring towards.”
