I was reading through YC’s Requests for Startups and thinking about new operating systems for businesses and the physical world—an area I have been deeply interested in for some time. I can see the direction clearly: AI will eventually take over a large share of the work involved in running companies and coordinating the physical world.
But I didn’t begin with a specific company or product. I began by writing down ten different situations where I felt something was still strangely manual, and from there I started thinking about what the future should actually look like.
A small B2B trading company may run entirely through WhatsApp and email—finding suppliers, sending RFQs, comparing quotes, preparing invoices, and chasing updates across different threads.
At Bosta, a member of the investigations team may spend days reconstructing a failed delivery by pulling data from systems, CCTV footage, driver calls, app logs, and warehouse records just to answer a simple question: what actually happened?
As a manager, I still spend much of my time reading emails, sitting in meetings, and following up across Slack just to stay closely connected to what is happening across the company.
Even in my personal life, I can spend weeks comparing hotels, apartments, or restaurants—scrolling through reviews, checking availability, and cross-referencing options—when what I really want is a system that already understands my preferences and gives me a few good choices I can trust.
These situations appeared unrelated. But underneath them was the same question:
Why are humans still the glue holding every system together?
In many cases, the human is not providing vision or making an important judgment. The human is coordinating, supervising, investigating, searching, and keeping everything in sync.
That is real and exhausting work. But increasingly, it should not require a human to perform it continuously.
1. AI is already capable of more than we let it do
AI can already take on parts of the work required to keep an organization running—and a growing share of the planning, sequencing, orchestration, investigation, and exception handling involved.
Much of an operations manager’s day is not spent making profound decisions. It is spent finding out what happened, noticing what is falling behind, chasing the next action, resolving inconsistencies, and preventing small failures from becoming expensive ones.
The closest mental model I have for what this could become is FRIDAY from Iron Man.
Tony Stark does not give FRIDAY a series of tiny prompts. He gives it an objective and a direction. FRIDAY performs the calculations, runs the simulations, coordinates the work, reports what it discovers, and returns when it needs his judgment.
The human provides intent. The intelligence accepts responsibility for moving it forward.
That is very different from most AI assistants today. Today, the human still has to operate the assistant: decide every next step, write every prompt, inspect every result, and keep the overall task in their head.
The assistant may perform individual pieces of work, but the human remains the orchestrator.
The real shift will happen when we can give AI an objective—not merely a prompt—and trust it to keep the work moving.
2. But AI cannot be trusted with control yet
The problem is that AI is book smart, not street smart.
It has read an enormous amount about the world. It has seen photographs and videos of almost everything. It can speak confidently about places, jobs, businesses, and experiences.
Life is not the internet. Life is not text, images, and video about life. What happens on the ground is more than what gets written down about it—and the gap between those two things is exactly where operations fail, where judgment is required, and where trust is currently impossible.
AI cannot safely run the real world until it understands what actually happens in the real world—not merely what the internet says about it.
Until then, it should not be given unlimited control. Not because AI is dangerous by nature, but because it has not yet earned that control.
3. It becomes street smart by being on the ground
AI will not develop this understanding through more text alone. It will develop it by participating in the world.
This understanding will come when hundreds of thousands of machines, agents, and instrumented humans are working in the physical world—touching it, acting in it, observing consequences, and failing. It will come from doing the work, getting it wrong, being corrected, and doing it again.
AI will see the difference between the plan and what actually occurred.
It will learn why a delivery that appeared successful in the system still produced an unhappy customer. Why a machine stopped despite showing no obvious error. Why two people can follow the same process and produce different outcomes. Why something that looks excellent online can feel disappointing in reality.
A phrase I keep returning to is:
To follow the path, look to the master. Follow the master. Walk with the master. See through the master. Become the master.
Today, the master is the human operator. The apprentice is the system. The curriculum is the work itself.
The most valuable lesson is often not the successful task. It is the failure and the human correction that follows it.
Capture the episode of real work—including the failure and the human correction—structure it, and the structured record becomes what future intelligence learns from.
That is how knowledge becomes experience.
4. Control should be handed over gradually
Between AI cannot be trusted with this and AI can run this lies a long period—a decade, probably more—in which control is handed over piece by piece.
At first, AI observes what is happening.
Then it investigates failures and reconstructs events.
Then it recommends actions for a human to approve.
Later, it handles clearly defined tasks inside narrow boundaries.
Over time, as its understanding improves and its decisions prove reliable, those boundaries can expand.
This is how a junior employee becomes a manager. They do not receive complete authority on their first day. They observe, perform limited tasks, make mistakes, receive corrections, and gradually earn greater responsibility.
AI should progress in much the same way.
But “humans remain in control” should not mean that a human approves every minor action forever. If every decision returns to a person, the human remains the bottleneck and AI becomes another assistant waiting for instructions.
Human control should operate at a higher level.
Humans should:
- Define objectives and priorities.
- Establish values, policies, and constraints.
- Decide what authority is delegated.
- Determine which decisions require approval.
- Inspect the evidence behind important actions.
- Correct the system when it is wrong.
- Revoke its authority immediately.
- Remain responsible for consequential decisions.
Inside a clearly delegated boundary, AI should be able to operate independently. Outside that boundary, it should stop and ask.
Trust should never be assumed. It should be measurable, reversible, and continuously earned.
5. From human-operated to human-directed
I do not believe the desirable future is one in which humans are removed from businesses or from the physical world. It is one in which humans are no longer the bottleneck holding every process together.
Humans should define goals, values, direction, and judgment. They should focus on relationships, creativity, leadership, and life itself.
They should not have to continuously supervise, coordinate, investigate, search, and keep systems in sync.
AI can increasingly perform that operating work—but only as it develops a grounded understanding of reality and earns the authority it receives.
The future is not simply autonomous machines replacing people. It is humans and intelligent machines learning how to work together, with a clearer division of responsibility.
Humans direct. AI operates.
And between those two roles lies the most important challenge: teaching intelligence how the world really works, then allowing it to earn control one responsibility at a time.