In early 2025, a team of five researchers from the AI Futures Project published AI 2027: a scenario document tracing how technological progress, governance choices and geopolitical dynamics would interact and compound over the next several years. The final report concluded that by 2027, it would be plausible for A.I. systems to surpass human cognitive performance across a meaningful range of tasks. This is called artificial general intelligence (AGI), and it’s a threshold event the entire A.I. industry has been tracking for years.
One year later, and with one year to go until that prophesied AGI moment, it’s worth revisiting the researchers’ original forecasts to highlight what is missing from the conversation, examine how A.I. capabilities are compounding faster than our institutions are prepared to handle and explore what that means for businesses (and the rest of society) in 2027 and beyond.
How did the AI 2027 authors’ predictions play out?
The authors outlined their five-year path to AGI, with explicit metrics to track progress along the way. They characterized 2025 and 2026 as foundational years, marked by the emergence of A.I. agents that would propel advancements in computational power and model capabilities, setting the stage for AGI in 2027. They extended the scenarios through 2030 by presenting two possible endpoints for a post-AGI world.
In a formal self-assessment published in February 2026, the authors estimated progress toward AGI had reached roughly two-thirds of their expected pace. Reaching AGI will thus take longer than expected, and the authors updated their forecast, pushing the likely arrival of AGI to between 2029 and 2032.
Although the authors may disagree among themselves on timing, that uncertainty does not detract from the pace of observable change. For enterprise leaders, A.I.-driven disruption is not a threshold event scheduled sometime in the future. It is a process already underway, and the gap between the disruption currently in motion and the institutional capacity to absorb it is widening, not closing. Even if all A.I. progress were to stop today—no new models, no new capabilities—the effects of what has already been deployed would take more than a decade to fully work through institutions, labor markets and organizational structures. A.I. is already advancing at a pace that infrastructure cannot match.
What the AI 2027 authors got right: agents are booming
The report’s predictions through 2026 held up well in one specific area: the uneven success of A.I. agents. Commercial rollout has been lackluster, but research and coding capabilities have made enormous leaps.
2025 was the year of A.I. agents. Perplexity and OpenAI launched their agentic browsers, though both have struggled to gain traction. OpenAI’s Instant Checkout, piloted with Walmart, produced underwhelming results. Walmart has since pivoted its agentic strategy, finding success with its in-app chatbot, Sparky. According to its fourth quarter FY26 earnings call, roughly half of all Walmart app users have engaged with the tool, with average order values 35 percent higher than those of non-users. The lesson is one of nativeness: agents succeed when embedded in existing trusted contexts, not when they ask users to adopt new ones.
While A.I. agents have had mixed consumer uptake, they excel in research and coding. AI 2027 described a self-improving loop, where coding agents accelerate research, research accelerates next-generation models and next-generation models become better coding agents. That loop is now structurally visible. In January 2026, Boris Cherny, head of Anthropic’s Claude Code, said that “pretty much 100%” of the company’s code is now A.I.-generated. At Davos, Anthropic CEO Dario Amodei projected that the industry is six to 12 months away from A.I. handling most software engineering end-to-end.
And there are no signs this compounding is slowing down. METR’s task-completion data, updated through April 2026, shows an exponential trend in the complexity of tasks frontier agents can reliably complete, with no evidence of flattening. GPT-5-class agents can now complete, in minutes, tasks that would take a skilled human developer roughly two hours.
What the AI 2027 authors got wrong: the economic impact of AGI
The report framed economic disruption from A.I. systems as something arriving at the end of 2026, on the precipice of the AGI threshold. What was underweighted was the uneven economic impact already accumulating well before that point.
A.I.’s capabilities—strong at codifiable, structured tasks and notably weaker at high-context judgment and relational work—map directly onto entry-level professional roles. As a result, the bottom of the career ladder is being displaced first, because their tasks are precisely the ones A.I. has gotten good at completing quickly and with accuracy. The Federal Reserve Bank of New York reports recent college graduate unemployment at 5.7 percent in the fourth quarter of 2025—higher than overall unemployment at 4.4 percent, and with underemployment at 42.5 percent, its highest rate since 2020. Block cut 40 percent of its workforce, with its CFO, Amrita Ahuja, calling A.I.-driven reductions an “inevitability for companies.”
But labor market headlines obscure a more systemic organizational problem. Companies have historically built capability depth over time by allowing entry-level hires to learn through progressively more complex work. When A.I. handles that entry-level work, future mid- and senior-level talent no longer accumulate the foundational experience required to grow into these roles. Eliminating the bottom rung of the corporate ladder, by choice or by force, produces short-sighted gains, with much larger talent implications in the years ahead. No existing training infrastructure, including higher education, can take someone from zero junior-level experience directly to mid- or senior-level roles.
Meanwhile, a significant and underappreciated gap has opened inside most organizations: the distance between how executive leadership perceives A.I. adoption and how it is experienced at the working level. Boards and C-suites are debating strategy—the domain where A.I. still struggles—while the operational burden of implementation falls to junior and mid-level employees. These workers are navigating workflow disruptions, absorbing the difficult decisions that A.I. routes upward after handling the routine work, managing change without established playbooks. This is fundamentally an organizational design problem, and it is one largely invisible in the standard A.I. returns conversation.
What’s next: A.I. in the physical world
A decade ago, the dominant assumption was that A.I. and automation would handle physical and menial work—logistics, manufacturing, routine labor—leaving humans to do the thinking. Generative A.I. inverted that story. Cognitive work has been disrupted first, while physical tasks remain harder to automate.
But the inversion we’re seeing may be temporary. Researchers Yann LeCun and Fei-Fei Li have both founded billion-dollar research labs focused on physical A.I.: models grounded in how the world works, not just how language describes it. When that type of project matures, both cognitive and physical labor will face disruption simultaneously, and the original thesis reasserts itself. The window in which blue-collar work appears relatively protected from the downward impacts of A.I. advancement may be shorter than it seems.
We’re already seeing early signals of this reversal. DoorDash’s Tasks program pays gig workers to complete physical activities A.I. cannot yet perform autonomously, like photographing environments and restaurant menus, to generate robotic training data. Waymo is reportedly paying DoorDash drivers to close robotaxi doors that passengers leave open. A.I. agents that can’t complete tasks on their own are hiring humans as subcontractors through platforms like TaskRabbit or RentAHuman.ai.
My take: there is no existing legal or regulatory architecture to govern the type of inversion we’re seeing, where A.I. acts as a principal, and humans are the agents. And without those governing principles or bodies, we may soon be entering the real “Wild West” of the A.I. growth story.
Nuance, uncertainty and the reality that’s already here
The honest answer to “when does AGI arrive?” is that we do not know, and the people closest to the problem disagree by years in either direction. That uncertainty is appropriate and should be respected. However, the AGI date is the wrong variable to organize around.
Current systems, built for a pre-A.I. world, are already bending under the load of current A.I. deployment. We are likely only 10 to 15 percent of the way through the total impact that today’s A.I., fully adopted and integrated, will eventually produce. The remaining 85 to 90 percent will arrive not because capabilities improve further, but because organizations, regulations, labor markets and infrastructure slowly catch up.
The question for enterprise leaders is not whether to prepare. It is whether they are preparing for the right disruption—the one already in motion—rather than waiting for a threshold event that may be less significant than the accumulating weight of what is already here.
