When people talk about bringing AI into their work, everyone pictures something different. One person imagines asking AI to draft an email; another imagines AI going out and generating revenue on its own. They use the same phrase — "adopting AI" — but they're describing completely different stages. 

Sort it out, and you can draw four distinct stages, organized around how the relationship between humans and AI changes — and what the real risk is at each step. 

Stage 1: AI Assists — AI as a Tool

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

You ask for a first draft when you're writing. You ask for a single function when you're coding. You ask for a summary of your data. The human owns every decision and every action; AI is a tool that knocks out discrete tasks quickly.

At this stage, AI is essentially no different from a calculator or a search engine. It does what it's told, and nothing when it isn't. Quality control rests entirely with the human. Most people are here right now.

Stage 2: AI Plans, Humans Approve — Human in the Loop

AI's role moves up from executing pieces of a task to planning the whole thing. This is the stage spreading fastest today. Most AI coding tools and workflow-automation agents live here.

The AI plans the entire job and proposes an execution path. The human reviews that plan, then approves or revises it. Using plan mode in Claude Code, or an AI agent drawing up a task list and waiting for a human to say "go" — that's this stage.

This structure is called "human in the loop": the human sits inside the decision-making cycle. AI proposes, the human decides, AI executes. Every critical judgment requires human sign-off.

The upside is plenty of room for humans to catch AI's mistakes. The downside is speed. Since every step waits on human confirmation, you never get the full benefit of AI's processing speed — the bottleneck is the human's very presence. IBM has observed that this model "improves accuracy and trust, but becomes a bottleneck, especially in fast-moving environments."

Stage 3: AI Works, Humans Supervise — Human on the Loop

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

In other words, the human no longer approves each step. The AI judges and executes on its own across the entire workflow, while the human monitors and intervenes only when something goes wrong. Someone with authority can redirect the work, but stays out of day-to-day execution.

This structure is called "human on the loop" — a concept first formalized in the military. The classic example is an autonomous surveillance drone patrolling independently while a remote operator monitors in real time, ready to step in when needed.

In a work context, it looks like this: ten AI agents run different tasks simultaneously — writing code, opening pull requests, running tests, fixing bugs. The human watches a dashboard, tracks overall progress, and corrects only the tasks that drift off course. Y Combinator's Garry Tan running 10 to 15 parallel sprints with gstack is close to this stage.

The core risk is automation complacency​: AI seems to be running fine, supervision slackens, and problems quietly accumulate until they erupt all at once. As the gap widens between when AI acts and when humans intervene, the blast radius of failure grows.

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

Stage 4: AI Generates Its Own Revenue — Human out of the Loop

This stage hasn't fully arrived, but the direction is unmistakable: AI operating autonomously without human supervision, all the way through to monetization. Finding customers, delivering services, managing costs, generating revenue — the stuff of dreams.

This is called "human out of the loop." Humans define the goals, constraints, and success criteria, but stay out of day-to-day operations.

Parts of algorithmic trading already come close, and the final stage of autonomous driving — SAE Level 5 — is the same idea.

The biggest risk here isn't technical; it's accountability. When an AI's autonomous decision causes a loss, who answers for it? When an AI delivers a faulty service to a customer, who is the legal party? Even once it's technically feasible, this stage is hard to realize without the law and social consensus to back it up.

The Stages Coexist

In reality, these four stages don't arrive in sequence — they coexist. They mix within the same company, even within the same workflow.

Consider how an airline rebooks passengers after a flight cancellation. Simple rebookings for economy passengers are handled autonomously by AI (close to Stage 4). For a first-class passenger with a complex international connection, the AI assembles the context and asks a senior employee to approve (Stage 2). A manager monitors the entire rebooking flow, watching for cost anomalies and pattern errors (Stage 3).

Within a single workflow, different stages apply depending on how risky each decision is. Going forward, the key to designing AI adoption won't be the blanket question "how do we bring AI into our company?" but the sharper one: "which stage does this particular task belong to?"

Where We Are, and Where This Goes

Most individuals and organizations sit between Stages 1 and 2 — using AI as an assistant, or letting agents draw up plans that humans approve.

But the move toward Stage 3 is accelerating. Approaches like Claude Code and gstack — where AI agents autonomously write, test, and ship code while humans supervise and make only the critical calls — are spreading fast.

Stage 4 still sounds far off, but pieces of it have already begun. AI is running ad campaigns, adjusting prices, and handling customer service across a widening range of domains. Handing AI the entire revenue engine, though, is less a question of technology than of institutions and trust.

As you climb the stages, AI's autonomy grows and the human role shifts from execution to design. The nature of the risk shifts too — from technical error to structural accountability.

What matters isn't simply identifying which stage you're in, but thinking hard about the posture you need to hold. That's why I'd conclude that what's required is the mindset of a planner, a project manager. The person who concretely understands and can steer the entire journey of a product or service — from conception to delivery — is the one who can ultimately manage AI from outside the loop, as its executive.