After “AI-First” Comes “AI-Only”

What if AI doesn’t want to be human?

I cofounded an AI-first company, Lemonade, and I think that by the end of the decade, “AI-first” will seem quaint. The “horseless carriage” of artificial intelligence.

Lemonade was early to “AI-first.” Back in 2017, shortly after we launched, my cofounder, Shai Wininger, published a blog titled “Don’t hire more people, build smarter AI.” That directive shows up starkly in our numbers: in the past three years, we nearly tripled revenue, added 1.25 million customers, grew gross profit sixfold — and reduced the size of our team.

But having scaled the “AI-first” mountain, the next peak reveals itself: “AI-only.”

By “AI-only,” I do not mean humanless companies or “solopreneurs.” I mean extensive workflows in which no human sits inside the operating loop. Humans still set the goals, values, constraints, and the conditions for escalation, but execution runs end-to-end on machines. Humans move from participation in the workflow to stewardship of the system.

If AI was once conceived as a “co-pilot” helping the human, and “AI-first” asks whether it can do the work of the human pilot, “AI-only” asks whether the very notion of a “pilot” isn’t an anthropomorphism. An anachronism. A relic from the days when human-centered design made sense.

Think of the F-35 strike fighter jet. It is an extraordinary piece of machinery, and it requires just one human: the pilot. Set against the cost and complexity of the aircraft, an aircrew of one seems almost incidental. It isn’t. Because there is a human in the cockpit, the aircraft needs a cockpit: canopy, oxygen, displays, controls, ejection seat, pressurization, ergonomics, and a safety envelope built around the human body. The plane must accommodate sight, breath, reach, attention, fatigue, hunger, the call of nature, and the mere 9Gs the human body can sustain — even with years of training, anti-G suits, and chronic neck damage.

Removing the pilot isn’t about saving their salary. It is about unleashing the true capabilities of $100m of stealth technology.

Dammed AI

98% of our code at Lemonade is written by AI. The gains are real: 30–50%. But they’re not yet 30x or 50x. Why?

The agent can write the code, but the system still waits for humans to decide what matters, prepare the environment, resolve ambiguity, review the diff, coordinate dependencies, and approve the ship. Intelligence is machine-speed. The workflow is gated by the speed of the last human in the loop.

And what is true in engineering is true across our organization: there is a growing reservoir of intelligence pressing against a dam. And the dam is us.

Human Architecture

In recent years, many CEOs have adopted Shai’s credo, enforcing a strict discipline: no new headcount until it is proven that AI cannot perform the task. The rewards of this “AI-first” posture are tangible, but the mindset remains anchored in the past. It identifies a human-shaped vacancy and asks if a model can fill it, assuming the traditional org chart—the jobs, teams, and handoffs—is essentially correct. It merely asks who should occupy the chair. “AI-only” asks why the chair exists at all.

Dwarkesh Patel has argued that AI won’t replace humans anytime soon, because jobs are fuzzy and demanding. “Every day, you have to do a hundred things that require judgment, situational awareness, and skills & context learned on the job,” he wrote. He is right. But that logic cuts both ways: if a job is a human-shaped bundle of tasks, why preserve the bundle? “AI-only” doesn’t ask if a model can be an employee; it asks what a workflow looks like when it is no longer packaged for one.

We are blind to how much of our corporate architecture is built around human quirks. A queue, a handoff, a manager, a meeting, an expert — these exist because human attention is finite, memory is leaky, training is narrow, and bandwidth is limited. Working hours and approval loops are the “safety envelope” built around our limitations. Remove the final human, and the system is unleashed from human topology and human speed. The resulting productivity isn’t just about saving a salary; it is the sound of the dam breaking. The release comes from discarding everything that salary represents: the delay, the variance, and the overhead. When the architecture itself changes, gains are no longer measured in percentages. They are measured in multiples.

Human Cost

I buy the cold logic I’m selling. If one company can deliver the same service faster, cheaper, and more reliably by removing humans from the ordinary flow, others will be forced to follow. What begins as a managerial choice becomes a condition of competition. This does not happen on a single Tuesday. “AI-only” arrives slice by slice, workflow by workflow, as systems earn trust.

But the direction of travel is inevitable and irreversible. Not because it is necessarily good for our species, but because it is good for business. There is no responsible way to write about “AI-only” without acknowledging the potential for growing and permanent unemployment. The economic case for the transition will be overwhelming for individual firms, while the social cost will land on workers, families, and communities. The firm-level math and the social ledger will not reconcile themselves.

Anyone concerned with that gap, and we should all be, must resist wishing away the economic forces that compel its emergence. Better to name them and prepare accordingly. (For my own efforts, see mosaicmodel.org.)

Alien Intelligence

The first combustion engines had about one horsepower, so the “horseless carriage” frame made sense. We understand new machines by reference to what they replace.

Then the machine outran the metaphor. Tens of horsepower, then hundreds. At some point, the engine was no longer intelligible as a better horse. It had become something else entirely.

AI may be in its own horseless-carriage phase. The phrase “artificial intelligence” makes AI sound like a synthetic human, a drop-in coworker. That works for the AI-first era. It lets us imagine AI as a better, cheaper, tireless human inside our familiar workflows.

But AI’s future is not as a better human. It is as something else entirely. “Alien intelligence” may prove the more useful backronym for what comes next.

That is the move from AI-first to AI-only, and it begins when we stop imagining intelligence in our own image. Ada, the android in After Yang, puts it plainly: “We always assume that other beings would want to be human. What’s so great about being human?”