When people say they're "bringing AI into their work," the picture in each person's head is different. One imagines asking AI to draft an email. Another imagines AI generating revenue on its own. They use the same phrase—"adopting AI"—but they're actually describing completely different stages.

If you sort it out, you can divide it into four stages, organized around how the relationship between humans and AI changes—and what the real risk is at each one.

Stage 1: AI assists — AI as a tool

This is the most familiar stage. The human leads, and AI helps.

You ask it to draft a piece of writing. You ask it to generate a single function while you're coding. You ask it to summarize some data. The human is the agent of every decision and every action, while AI is a tool that quickly handles discrete tasks.

At this stage, AI is essentially the same as a calculator or a search engine. Ask, and it does; don't ask, and it does nothing. Quality control rests entirely with the human. Most people are right here today.

Stage 2: AI plans, the human approves — Human in the Loop

AI's role rises from "partial execution" to "planning." This is the stage spreading most actively right now. Most AI coding tools and workflow-automation agents fall here.

AI plans an entire task and proposes a way to carry it out. The human reviews that plan and approves or revises it. Using plan mode in Claude Code, or an AI agent building a task list and the human giving the "go ahead" command—that's this stage.

This structure is called "Human in the Loop." It means the human sits inside the decision-making loop. AI proposes, the human decides, AI executes. Every key judgment requires human approval.

The upside is that even when AI makes a mistake, there's plenty of room for the human to catch it. The downside is speed. Because you have to wait for human confirmation at every step, you can't fully exploit AI's processing speed. The bottleneck comes from the human's presence. IBM analyzes this model as one that "improves accuracy and trust but becomes a bottleneck, especially in fast-paced environments."

Stage 3: AI works, the human supervises — Human on the Loop

This is where the decisive shift happens. The human rises from "inside" the loop to "above" it.

In other words, the human no longer approves every step. Instead, they monitor the whole flow as AI judges and acts on its own, stepping in only when something goes wrong. An authorized person can reset the direction of the work, but doesn't get involved in day-to-day execution.

This structure is called "Human on the Loop." The concept was first established in the military. A classic example: an autonomous surveillance drone patrols independently while a remote operator monitors in real time and intervenes when needed.

In a work setting, it looks like this. Ten AI agents perform different tasks simultaneously. They write code, open PRs, run tests, and fix bugs. The human watches a dashboard to track overall progress and only steps in on the work that has veered off course. The way Y Combinator's Garry Tan runs 10 to 15 parallel sprints with gstack is close to this stage.

The core risk here is "automation complacency." Trusting that AI is running fine, you slack off on supervision, and problems pile up until they blow up all at once. As a gap opens between the moment AI executes and the moment a human intervenes, the blast radius of failure grows.

The EU AI Act (Article 14) mandates "effective human oversight of high-risk AI systems" precisely because it recognizes the risk of this stage.

Stage 4: AI generates revenue on its own — Human out of the Loop

This hasn't fully arrived yet, but the direction is clear. AI operates autonomously without human oversight and even handles monetization. It's the dream of AI finding customers, delivering services, managing costs, and generating revenue.

This is called "Human out of the Loop." The human defines goals, constraints, and success criteria, but doesn't intervene in daily operations.

Parts of algorithmic trading already come close to this territory, and the final stage of self-driving (SAE Level 5) maps onto the same concept.

At this stage, the biggest risk isn't technology—it's accountability. When a judgment AI made autonomously causes a loss, who is responsible? When AI delivers a faulty service to a customer, who is the legal party? Even if it becomes technically possible, this is a stage that's hard to realize without supporting law and social consensus.

The stages coexist

In reality, these four stages don't arrive in sequence—they coexist at the same time. The stages get mixed even within the same company, even within the same workflow.

Take how an airline rebooks passengers when a flight is canceled. Simple rebooking for an ordinary passenger is handled autonomously by AI (close to Stage 4). A complex international connection for a first-class passenger gets the context organized by AI, which then requests approval from a senior staff member (Stage 2). A manager monitors the entire rebooking flow, watching for cost anomalies or pattern errors (Stage 3).

Within a single workflow, different stages apply depending on the risk level of each decision. Going forward, the key to designing AI adoption won't be the vague approach of "we're bringing AI into our company," but judging "which stage does this task belong to."

Where we are now, and what's ahead

Most individuals and organizations sit somewhere between Stage 1 and Stage 2. They use AI as an assistive tool, or the agent draws up a plan and the human approves it.

But the move into Stage 3 is accelerating. As with Claude Code or gstack, an approach is spreading in which AI agents autonomously write, test, and deploy code while the human, as a supervisor, makes only the key decisions.

Stage 4 still sounds like a distant story, but in part it has already begun. The areas where AI automatically runs ad campaigns, adjusts prices, and handles customer service are widening. That said, entrusting all of monetization to AI is a matter of institutions and trust more than technology.

As the stages rise, AI's autonomy grows and the human role shifts from "execution" to "design." And the nature of the risk changes from "technical error" to "structural accountability."

What matters is not simply identifying which stage you're in, but staying mindful of what posture we should hold. That's why I want to conclude that what's needed is the attitude of a planner, a project manager. The person who concretely understands and can control the entire process—from planning a product or service to delivering it—may, in the end, be the one who manages from outside the AI loop, as an executive.