Not everyone uses AI, and thankfully there are roles and circumstances in life that operate in reality—the physical kind. But most people have at least tried it, and they use AI as a question box. Instead of searching on Google, they type into ChatGPT, "give me three dessert recipes without cream." It works. You get answers faster than before. A bit like how all new cars have GPS.
The automotive industry has a somewhat fuzzy but still systematic way of describing the path toward fully autonomous and self-driving cars. It's called SAE levels and goes from 0 to 5. Level 0 is no automation—you do everything yourself. Level 1 is cruise control. Level 2 is what Tesla calls Autopilot: the car maintains lane and distance, but you must keep your hands on the wheel. Level 3 is conditional automation—the car drives itself in certain situations, but you take over when it asks. Level 4 is high automation—you enter an address and the car takes you there. It could be your own car or one owned by Waymo or Volvo, but some human is keeping watch, either remotely or in the vehicle. Level 5 is full autonomy—no steering wheel needed; you take a nap, watch a movie, and get a notification when you arrive.
Actually, level 0 is already magical. For anyone who remembers the map that came in the phone book, that's the real starting point. At our level 0, you don't read the map yourself—you get real-time updated instructions. But you're behind the wheel, stepping on pedals, turning and signaling (unless you're driving a BMW or Audi, where that apparently isn't necessary). The system gives you information, but you do the work.
The interesting thing is that most people using AI today are right there. They have a Ferrari in the garage but are using it to buy milk.
Level 0 is where most people are: you type a question in a chat window. Wait. Get an answer. Type the next question. Wait again. The AI is your GPS, telling you where to turn, but you're driving. The system pauses between each instruction. One conversation at a time, synchronous. You ask what the difference is between OKR and KPI, read the answer, think, write a follow-up question.
You level up to Level 1 and pick up speed. You write a question, and while the AI is answering, you're already formulating the next one. You upload documents, create projects, build knowledge bases. The AI understands the context and gives better answers. But you still have your hands on the wheel. You guide every step. You ask for a document summary, and before the answer is even complete, you've already written "make a bullet list of the key insights" and "suggest three next steps."
When we move to Level 2, things start happening. You hire someone to drive for you. You describe the destination, not the route. "Analyze Spinout AB as a potential partner. Review their annual report, product offerings, customer reviews, and news articles. Summarize strengths, weaknesses, and risks in a memo." The agent breaks down the task, executes it, evaluates the result. You take over when it gets stuck or needs a decision, but you're no longer keeping your hands on the wheel the whole time. You define the task, go do something else, come back to a finished result.
At Level 3, you orchestrate a team of capable workers doing the job. You have three, four, five AI agents running in parallel. Each agent has its own task and its own track. One creates a competitor analysis, one answers customer emails, one prepares a board meeting. The difference from before is that they all drive themselves. You coordinate; you don't drive.
Up a level—now we take it a step further. You put multiple agents on the same problem and they collaborate without you overseeing each step. You need to produce a market analysis. Agent 1 analyzes competitors' pricing. Agent 2 goes through industry reports. Agent 3 compiles customer surveys. Agent 4 writes the final report based on the others' results. A task that would take a week is done in an hour.
Level 5 is where the fastest companies are right now. These are primarily companies operating in or near the already massive AI industry—like OpenAI, Google, Meta, Cursor, Lovable. Here you take the full step and give a team of agents a large, complex task. "Build a browser," "Create a new feature called Cowork," "Build a new runtime for Java." The agents break down the task and then create their own sub-agents to handle each part. What might this sound like at a bank? "Do due diligence on Spinout AB for a potential acquisition. Include financial analysis, legal review, market position, technical infrastructure, and cultural fit." The agent creates sub-agents, one for each area. They work for hours, sometimes days. They ask questions when stuck, iterate on their solutions, and eventually deliver a complete foundation for decision-making. You kick off the process Monday morning and have a thorough decision brief by Tuesday. Like calling McKinsey and saying "fix this," and getting delivery for the cost of the first meeting.
What's happening in engineering shows the way for the rest of us. In January 2026, a shift occurred. Developers discovered that if you put AI in a loop that keeps trying until the tests pass, it can solve problems that previously required weeks of human work. They call it the "Ralph method." Persistence as strategy. Then came sub-agents. Instead of one long conversation, the AI now spawns specialized mini-agents. They don't know about each other and don't interfere with each other's work. Problems that previously required project management are now handled by the system itself. What's happening in engineering today will soon reach sales, finance, and HR. The only question is when.
To understand why this is a shift, you need to understand how agents keep track of what they're doing. It's not magic. It's three techniques that together create something new.
Every AI has a "context window"—how much it can hold in its head at once. Think of it as a desk. The bigger the desk, the more papers you can have out. In 2023, the desk was small. Maybe 8,000 words. Now it's 200,000 words or more. But it's still limited. Just like working memory in a human.
Instead of trying to hold everything in their heads, agents use the file system as external storage. The Ralph method works like this: The agent works, saves the result to a file, clears its short-term memory, reads the file back in, and continues. It's like you writing notes, going to bed, waking up and reading the notes to continue where you left off. The agent doesn't need to remember—it just needs to know where the information is.
The third technique is externalizing the plan. Instead of keeping the entire project plan in its head, the agent writes it down. "Next step: analyze file 3. Then: summarize results. Then: create report." The agent doesn't know what it did an hour ago. But it knows exactly what the next step is.
Previously, AI was limited by how much it could hold in its head. That set a ceiling on complexity. Now AI is limited by how well you can structure the task. With the right structure, an agent can work for days and deliver results that previously required an entire team. Just like a Level 5 car can drive a thousand miles without the driver needing to be awake.
When AI can produce unlimited material, value shifts. What has value is the ability to specify—to describe the end result so clearly that an agent can work for days without going off track. That's harder than it sounds. It's like programming the destination; if you're unclear, the car ends up in the wrong place. And the ability to verify—to determine whether the result is good enough. Building tests, not just for code, but for quality, tone, correctness. Like checking that the car actually delivered the right package to the right address.
Klarna is an interesting example. They automated customer service at scale—700 employees replaced, wait time down from eleven to two minutes. But after a while, they started hiring again. Pure cost focus had led to quality deterioration. The lesson: AI takes over execution, but someone still needs to define what quality means. Even a Level 5 car needs someone to decide where it should go.
This isn't about technology. It's about what you do with your time.
Twenty years ago, we spent hours formatting documents. Then came templates. Ten years ago, we spent hours on research. Then came search engines. Now we spend hours producing, writing, analyzing, compiling. That time is disappearing too.
The question isn't whether you can keep up. The question is what you do with the time that's freed up.
Those who get stuck at Map Reader will use AI to do the same things a little faster. Like buying a modern car and only using cruise control. Those who find agentic AI will do entirely different things. Things that weren't possible before. Not because they work harder, but because they understand what the tools can actually do.
I talk to super-smart people every week who say "we don't have time to learn this." They're wrong. They don't have time not to. But the beautiful thing is that it doesn't require a big project. It doesn't require a strategy or a budget. It requires you to test. To give an agent a real task and see what happens. To let go of the wheel and feel what it's like.
Once you've done that, you understand why this is a shift and not a feature. And then no one needs to convince you of anything.
Sources
- Klarna has gone from 5,000 to 3,000 employees with the help of AI, Ny Teknik
- After the AI push, Klarna starts hiring again, EFN
- AI could take over 300,000 jobs, Google/Swedish JobTech report
- SAE Levels of Driving Automation, SAE International