Just as 1920s switchboard operators connected calls, AI operators have returned. You request; AI clicks and types. The human role shifts from dialer to director.
How many times a day do you wish something on the internet would "just get done for me"? Hotel booking, food delivery, ticket buying, grocery shopping, restaurant reservations. Different tasks, same structure: open a website, click a button, fill a form, press pay. We do this dozens of times a day. That loop is quietly disappearing. AI has returned as the old switchboard operator.
This essay walks through that shift from start to finish. Even if you've never used an AI agent, you can follow along — we'll go slowly. Today's example is OpenAI Operator, released January 2025. But the principle applies to any tool that comes next. Even if the name changes, the structure of a "middleman returning" does not.
In the 1920s, people were the operators. You picked up the phone and someone answered first. "Which number should I connect you to?" They physically plugged cables to link caller and receiver. A human sat in the middle.
Time passed. Automatic exchanges came. Mobile phones came. We entered the era of dialing our own number. The middleman was replaced by technology. A hundred years later, computing followed the exact same arc. First, you walked into a cafe to order. Then phone orders. Then delivery apps. Now we click buttons on an app screen — directly, with our own fingers.
Strangely, we stopped there. Why are we still pressing buttons one by one? AI supposedly "understands" everything, so why is the hand still mine? Simple answer. AI didn't know how to look at a screen for us. That just got solved.
In January 2025, OpenAI released an AI agent called Operator. Fun name. Operator. The old switchboard workers used the same word. The name itself gives the answer — OpenAI revived "the being that runs things for you."
Here's Operator in action. User types: "Find me a family-friendly campsite near Joshua Tree this weekend." Operator opens Hipcamp on its own. Clicks buttons. Picks dates. Selects options. Brings back results. Only at the final payment does it stop and ask, "Book this one?"
Other demos — reserves a refundable hotel meeting specific conditions, grabs a 7pm table for two on OpenTable, fills a grocery cart on Instacart, picks four basketball tickets under $500 on StubHub. In every case, the user provided one sentence. The rest — AI watched the screen and did the work.
Technically this works because OpenAI shipped a model called CUA (Computer Using Agent). Computer vision reads the screen in real time, figures out which pixels are buttons and input fields, then drives mouse and keyboard directly. No API integration. It sees and clicks like a human. That's the big shift.
Try this analogy — the old neighborhood errand service.
Years ago, services existed where you'd call a local office and say, "Please submit this form at the district office tomorrow morning." The owner would go, pull a number, find the right window, file the papers, and return with a receipt. You stayed home drinking coffee while the task finished.
The AI operator has the same shape.
Old delivery apps were navigation — "here's a map, go find it yourself." Operator is delegation — "I'll go; you wait." Different structure entirely.
Conditions as of January 2025. These will loosen, but right now:
| Item | Status |
|---|---|
| Model | CUA (Computer Using Agent) |
| Availability | US ChatGPT Pro users only |
| Price | $200/month (ChatGPT Pro tier) |
| Stage | Research Preview |
| Supported sites | OpenTable, Instacart, StubHub, a few others |
| Before payment | Always confirms with user |
ChatGPT Plus costs $20/month. Pro costs 10× more. Why so expensive? Because AI has to watch the screen in real time and reason. Screenshot → analyze → click → take another screenshot → analyze again. This loop runs hundreds of times per task. Computing cost is dozens of times a simple text answer.
Here's the first aha.
The real shift of the agent era isn't the price. It's that the human role moves from dialer to director.
Previously, my finger pressed things. Now AI presses for me. So what do I do? Design what, why, and under what conditions the task should happen. I climb from the one who clicks to the one who instructs.
How do you find the entry point for yourself? Ask one question:
"What do I click and repeat every single day?"
That answer is where Operator will arrive first. Three buckets.
Count how many of your current tasks fall in bucket one. That's the area about to automate.
A concrete example. Task: book a 4-person Italian restaurant in Yeonnam-dong for this Saturday dinner.
Old way — by hand.
Open Naver Map → search Yeonnam-dong → filter Italian → sort by rating → compare four spots → find booking links → enter date → enter party size → enter phone → receive confirmation. About 20 minutes.
New way — one sentence.
"Book a 4-person table for Saturday 7pm at an Italian restaurant in Yeonnam-dong. Quiet seating, no wait, under ₩50,000 per person."
Operator opens the map, finds matching spots, navigates to the booking page, fills the form, and stops only at the final confirmation — "Book this one?" I see one confirmation window. That's it.
Perceived time: about 2 minutes. 10× faster than the old way. And the bigger deal — during the 18 minutes I'd saved, I can do something else. We've returned to the moment when, while the switchboard operator connected your call, you could already prep the next one.
How to start today. Three options.
1. ChatGPT Pro ($200/month) — Operator directly (US)
2. Claude Computer Use — API-based automation
3. Mech, Zapier, Make — existing automation + AI agent glue
If you're in Korea, either use VPN or wait for local platforms (Baedal Minjok, Coupang, etc.) to bolt their own AI agents on top. Between 2025-2026, expect Korean services to ship similar features.
Do one thing in advance. Practice requesting with precision. "Book dinner" won't work. "Saturday 7pm, Italian in Yeonnam-dong, 4 people, quiet table, under ₩50k/person" will. AI getting smarter doesn't automatically sharpen your request. To work well with an Operator, you have to train the language of a clear director.
So what did we do today?
In the 1920s, switchboard operators existed. In 2025, the AI operator returned. A century later, the middleman is back. The internet and apps briefly invented a "click-it-yourself era" in between, but the structure has snapped back to middleman. Because humans weren't born to click. Humans were born to plan, to decide, to savor.
Keep one question inside you — "What do I click and repeat every day?" That's the first thing to hand off. Clicks to AI, decisions to you. Once this division becomes a body habit, your time widens tenfold.
The product names will change. OpenAI Operator won't be called Operator forever. Claude Computer Use, Gemini Agent — whatever comes next, the principle of the middleman returning stays the same. Back then, switchboard operators connected phone calls. Now AI operators connect the web. Tools change. The middleman doesn't.
Request. Relay. Confirm.