The Separation: Most People Talk to AI. The 7.7% Who Put It to Work Are Building What Comes Next

In June 2025, Wix paid $80 million for a software company called Base44. It was six months old. It had never raised outside funding. And for most of its life it had exactly one employee: Maor Shlomo, a solo founder who reached $1 million in annual recurring revenue three weeks after launch—by directing AI systems to do work that once required a full engineering team. (TechCrunch)

Stories like this are usually filed under startup lore. I think they belong somewhere else: as early evidence of a structural divide I have come to call the Separation.

The Separation is not the divide between people who use AI and people who don't. That gap is closing on its own—a majority of working-age Americans now use generative AI in some form. (St. Louis Fed) The divide that matters is inside the user base: between those who converse with AI—asking questions, drafting emails—and the much smaller group that directs it, delegating real work to autonomous systems that research, build, monitor, and execute while their operator attends to something else.

Directing sounds technical. In practice, it looks like handing off the busywork. For a business owner: an agent that screens the week's leads, chases the overdue invoices, watches competitors' pricing overnight, and has the Monday report waiting before the coffee is poured. For anyone else: one that plans the trip, tracks the family budget, or quietly handles the errands that used to eat the weekend. Same tools, different posture—work delegated rather than questions asked.

How Narrow the Second Group Really Is

The AI Barometer, a June 2026 survey of 7,675 U.S. adults conducted with Prosper Insights & Analytics, measured the agentic population directly: 7.7 percent of Americans currently use agentic AI. Two-thirds—67.6 percent—have never heard the term. Forty percent say they would delegate a daily task to an agent; barely one in five of them actually has. (Channel V Media)

The same survey found that executives use agentic AI at 2.7 times the rate of employees. The separation is not only happening—it is happening from the top down.

And the window in which it plays out is compressing. Economists Alexander Bick, Adam Blandin, and David Deming found that generative AI reached 39.4 percent of American adults within two years of ChatGPT's launch—roughly double the internet's pace at the same age, and faster than the personal computer. (NBER) Divides that once took a generation to form, and a generation to close, may now form in a handful of years.

The Scholars Are Circling the Same Idea

Researchers are converging on the same idea from several directions, and the evidence is worth walking through.

Stanford economist Erik Brynjolfsson has warned of a "Turing Trap": when AI is built to replace human effort rather than augment it, the gains concentrate with whoever controls the systems, while everyone else loses bargaining power. (Brynjolfsson, 2022) The Separation is, in part, the Turing Trap observed from the individual's side: the person directing the system sits on the winning side of that ledger.

A landmark field experiment by Harvard's Fabrizio Dell'Acqua, Wharton's Ethan Mollick, and colleagues put GPT-4 in the hands of 758 Boston Consulting Group consultants. Those using AI completed 12 percent more tasks, 25 percent faster, at 40 percent higher quality—and the lowest-ranked performers improved the most. (SSRN) Mollick has since documented "secret cyborgs"—employees quietly multiplying their output with AI while concealing it from their managers. (One Useful Thing) The capability gap, in other words, is already inside organizations; it is simply not yet visible on the org chart.

Most striking, a 2025 paper by researchers at Oxford and Google DeepMind formalized the concept of agentic inequality: "disparities in power, opportunity, and outcomes arising from unequal access to, and capabilities of, AI agents." Their key observation is that agents differ from every previous technology because they act as "autonomous delegates rather than tools"—which means advantages scale through delegation, and agents ultimately compete against other agents on their operators' behalf. (arXiv)

The digital-divide tradition described who could access technology, and later who had the skills to use it well. What this newer work describes—and what the Separation names plainly—is a third level: who is willing and able to delegate.

The Evidence on the Ground

Outside the laboratory, the case studies are accumulating quickly.

Klarna reported that its AI assistant performed the work of roughly 700 customer-service agents within its first month of operation. (Klarna) Solo-founded startups—historically discouraged as underpowered—now represent more than a third of new ventures, up from under a quarter in 2019, with founders explicitly citing agent stacks as their substitute for early headcount. (Forbes) OpenAI's Sam Altman has described a running bet among technology CEOs on the year the first one-person billion-dollar company appears; most guessed 2028. (Fello AI)

The labor market is pricing the divide in real time. PwC's AI Jobs Barometer puts the wage premium for workers with AI skills at roughly 56 percent over otherwise similar peers—a premium that more than doubled in two years. (PwC)

The mechanism connecting these data points is compounding. A conversational user saves minutes. An agentic operator multiplies capacity—and each automated workflow frees time to build the next one. Advantages of that kind do not add. They stack.

The RiskForward View: What This Means for Business

Here is the part most coverage of AI misses: the Separation is not only happening between companies. It is happening quietly inside them. The "secret cyborg" research above suggests that in many organizations, the agentic minority already exists—unmanaged, unmeasured, and invisible to leadership. That is both an opportunity and an exposure.

The opportunity is that becoming an agentic company no longer requires an engineering department. The current generation of mainstream AI platforms ships with agent capabilities built in: assistants that can be handed a multi-step task and left to complete it, browser agents that operate software the way an employee would, and scheduled routines that run on their own clock—overnight research, morning briefings, weekly reports that assemble themselves. The on-ramp has moved from code to plain English. What remains scarce is not access. It is leaders who treat these capabilities as an operating model rather than a curiosity.

For business owners and executives, that suggests a near-term agenda:

• Bring the cyborgs into the open. Replace quiet, unsanctioned use with sanctioned tools, clear guardrails, and permission to experiment—the 40 percent who are willing are already on your payroll.
• Set delegation expectations deliberately. Define what agents may touch, where human sign-off is required, and how output gets verified—before habits form on their own.
• Reward directed output, not hours. When one employee can supervise the work of several systems, measuring effort by time at the desk misses what is actually being produced.
• Treat delegated authority as delegated liability. Every task handed to an autonomous system is also a risk transferred to it—a coverage and accountability question most insurance frameworks were not built for, and one I examined in The Surging Blind Spot.

There is a larger structural question looming behind all of this. I suspect the Separation will play out between companies the same way it plays out between people: the organizations most entrenched in legacy systems—decades of historic data, layered processes, technology built for a slower era—may find they can only run agents in a sandbox, while agent-native companies, built around delegation from day one, compound past them. If the last decade's disruption story was digital-born companies outrunning analog incumbents, the next one may be agent-born companies outrunning digital ones. That theory deserves its own analysis, and it will get one in a coming article.

What the Separation Is Not

Precision matters, because the dramatic version of this argument overstates it.

It is not a claim that the other 92 percent will be "left behind" by AI generally—conversational tools are diffusing broadly and will keep doing so. It is not permanent; 7.7 percent is a snapshot of an early moment, not a caste system, and willingness already runs at 40 percent. And it is not a prophecy that agentic users will "run the world." It is a narrower, better-supported observation: early compounding advantages tend to shape the institutions everyone else later inherits, and the group compounding right now is unusually small, unusually senior, and moving on an unusually fast curve.

The academic literature is careful to note the other possibility as well: broadly distributed agents could narrow gaps, as the BCG experiment hinted when the weakest performers gained the most. Which way it breaks may depend less on the technology than on how quickly the majority chooses to cross.

Crossing Over

Which returns to Maor Shlomo. Nothing about his story required rare genius or rare capital. It required orientation: the decision to treat AI not as a better search box but as a workforce to be directed—then the discipline to supervise it, learn from it, and reinvest every hour it returned.

That orientation is available to anyone reading this. Today it is held by fewer than one American in twelve—and the tools it requires are not on some distant roadmap. They are already on the device this article is being read on, waiting for instructions.

Every era hands out a brief window in which the extraordinary is simply available—before it becomes expected, priced in, and crowded. The window for learning to direct AI is open now, while 7.7 percent feels early rather than late. Individuals who step through it compound their own capacity. Companies that step through it together may find themselves growing at the pace of the technology instead of the pace of their org chart.

The Separation, then, is less a wall than a door that most people have not yet noticed is open. The 7.7 percent are not a different kind of person. They are simply the ones who walked through early—while the walking was easy.