You can see the Mona Lisa right now. Pull out your phone. Google it. You'll find a higher resolution image than you'd ever get standing in the Louvre. No crowds. No bulletproof glass. No three-hour queue. You can zoom into the sfumato around her eyes at a level of detail that Leonardo himself never saw.
So why are there still lines?
the museum paradox
The Louvre hit 8.7 million visitors in 2024. 8.7 million people queued for a small painting they could see in better quality on their screens.
The Rijksmuseum took it further. In late 2024, they launched an AI-enabled Collection Online platform. Their entire collection in hyper-resolution. 700,000 objects. Free. From your couch. No ticket, no flight to Amsterdam.
Did attendance drop? They hit 2.5 million physical visitors, their third-best year ever. Only the Rembrandt celebration and the Vermeer exhibition drew more. Online visits reached 7.2 million. Social media reach: 58.5 million. Making the art infinitely accessible didn't replace the in-person experience. It amplified the demand for it.
Now walk past the paintings and count the humans. The docent who notices you lingering at a Vermeer and says exactly the right thing. The conservator whose hands have memorized the viscosity of four-hundred-year-old pigment, painstakingly removing the varnish layer from The Night Watch, a process no algorithm will touch. The curator who placed this piece next to that piece, creating a conversation between two artists who never met.
Every year, museums get better technology. AI-guided tours. Computer vision for authentication. Algorithmic visitor flow optimization. And every year, the humans who remain become more important, not less. The technology handles the volume. The humans handle the moments that matter.
This isn't a museum problem. It's a preview of every industry on earth.
the inversion nobody's talking about
Jack Dorsey cut 4,000 people from Block on February 26. Not because the company is struggling. Gross profit is growing. Customers are increasing. He cut them because "intelligence tools paired with smaller and flatter teams are enabling a new way of working which fundamentally changes what it means to build and run a company."
Not will change. Are changing. Present tense.
The instinct is to see this as subtraction. 4,000 people gone. But the remaining 6,000 aren't doing the same work with fewer hands. They're doing different work. The work that can't be automated was never the filler. It was always the point. It was just buried under ten thousand hours of context-switching, status-updating, ticket-grooming, meeting-about-the-meeting administrative theater.
I see the same pattern across teams I work with. Once AI handles the routine, what remains isn't the scraps. It's the core. The conversations, the judgment calls, the moments where someone has to decide what matters. Engineering teams that adopt AI coding tools don't end up with fewer people doing the same work. They end up with the same people doing more consequential work.
What if the 4,000 jobs that disappeared weren't real work? What if they were the human-shaped duct tape holding together systems that weren't smart enough to run themselves?
the artisan economy is already here
A sommelier doesn't exist because wine is scarce. Wine is everywhere. A sommelier exists because choosing well is scarce. The abundance of options created the demand for taste.
Paul Graham argued in 2002 that taste is not subjective. Designers get measurably better over time. Earlier work is worse, not merely different. Taste is a skill that compounds with experience. Unlike production, it can't be automated to zero.
The same inversion is happening with information work. AI makes production infinite. Code, copy, analysis, designs, all approaching zero marginal cost. When production is infinite, curation becomes priceless. Judgment becomes priceless. The ability to look at twelve technically correct options and know which one actually matters becomes the most valuable skill in the room.
Look at what happened when photography replaced portrait painting. Painters didn't disappear. Painting moved from documentation to art. The camera handled the mechanical reproduction. The painter handled the interpretation. Every time a technology automates the mechanical, the interpretive becomes more valuable.
Anthropic just launched a plugin marketplace. Enterprises are building "dozens, hundreds, or even thousands" of AI mini-apps for every department. ServiceNow stock dropped 23%. Intuit dropped 33%. The tooling layer is commoditizing in real time. The market is pricing it in.
The "AI will replace everyone" crowd models a world where human work is a fixed commodity being eroded. But human work isn't fixed. It's elastic. It expands into whatever territory machines can't reach. And that territory is growing, not shrinking.
the handoff problem
I spent a week in February designing what I called a human-on-demand API. The idea: when an AI agent hits a wall requiring empathy, legal authority, subjective judgment, or trust, it calls an API endpoint. The platform handles context transfer, skill matching, and orchestration.
The critical insight wasn't the matching. It was the context pipeline. The agent doesn't just dump context at handoff time. It streams context snapshots throughout the workflow. Background on the situation. What's been tried. What's at stake. By the time a human is called in, they arrive warm. Not reading through fifty messages trying to figure out what happened. Pre-briefed. Ready. Effective in thirty seconds instead of thirty minutes.
Think customer success: the agent handles the standard workflow, but when the conversation turns emotional or the customer threatens to churn, the agent calls for a human with specific de-escalation skills. HR: the AI processes routine parts of a performance cycle, but hands off to a human for the actual conversation. Legal: the agent drafts the contract, a human signs off on the risk.
I thought I'd designed something new. Then I found HumanLayer. 9.6k GitHub stars. An open-source human-in-the-loop SDK. The same architecture: agents call an API when they need human input. Context transfer built in.
They pivoted away from it. The team now builds CodeLayer, an IDE for orchestrating AI coding agents. The original SDK documentation is marked "Legacy."
They saw more money in helping humans orchestrate more AI than in helping AI orchestrate more humans. That tells you where the market's attention is. And where the opportunity is hiding.
The patterns work in production. A voice AI system for insurance applications already uses trigger-based escalation: detecting when a user struggles (same form field edited five times), pulling in a human who arrives with full context. 97 tests. Real code. The infrastructure is being built, quietly, while the market chases the next coding agent.
the price of presence
There's a concept in Japanese aesthetics called wabi-sabi: the beauty of imperfection, the value of the handmade mark. A tea bowl with a slight asymmetry isn't flawed. The asymmetry is the art. It's evidence that a human was here.
The same principle applies to work.
When every company can spin up a fleet of AI agents that produce technically correct outputs at machine speed, the question stops being "can you do the work?" and starts being "can you be present for the work?" Can you read a room? Can you know when the data says one thing but the situation demands another? Can you sit with ambiguity long enough to find the answer that doesn't exist in any training set?
This is what I mean by the last handoff. Not the last time a human touches the work. The last moment that requires a human. The moment of presence, judgment, trust. Everything before it can be automated. This moment cannot.
The best doctor isn't the one who diagnoses fastest. That's a search problem AI already solves. The best doctor is the one who looks you in the eye and helps you understand what the diagnosis means for your life. The best lawyer isn't the one who finds the precedent. The best lawyer knows when to fight and when to settle, and has the judgment to tell the difference.
the economics of irreplaceability
A human customer success manager spends roughly 40-60% of their time on context: reading threads, catching up on history, sitting in meetings to stay informed. With proper agent infrastructure, that drops to near zero. The agent maintains context continuously. The human is called in only when needed, arriving pre-briefed.
I modeled this out. If context transfer time goes from 10-30 minutes (reading threads, asking questions, getting up to speed) to 30 seconds (structured briefing from the agent), that's a 20-60x improvement in the handoff alone. Multiply that by 8-10 handoffs per day and the math changes the entire role.
Utilization jumps from 40% to 85%+. The human does more of what only humans can do, less of what was always a waste of their time.
This isn't replacing the human. It's concentrating the human. Distilling away the noise until only the essential remains. A chef doesn't spend hours grinding wheat. They spend that time deciding how this particular dish should taste on this particular evening for this particular table.
When a human's real-work output is concentrated 3-5x, their value per hour doesn't just increase. It category-shifts. From "labor cost" to "strategic asset."
what this means for builders
If you're building products, the implication is clear: design for the handoff, not the replacement.
The companies that win the next decade aren't the ones automating humans out of the loop. They're the ones building the infrastructure that makes human moments happen at scale, with full context, at the right time.
The museum didn't fire the docents when it got audio guides. It freed them from giving the same introductory speech four hundred times a day so they could do what they actually do best: notice the person standing alone in front of a painting, and say something that changes how they see the world.
Every industry is about to face the same decision Block just made. The question isn't whether to restructure around AI. The question is what you're restructuring toward.
Are you optimizing for fewer humans? Or for better human moments?
The answer will determine which companies build something lasting and which ones discover, too late, that they automated away the only thing that made them worth choosing.
The machines are getting very good at production. The humans who master presence will be the most valuable people in any room.
Built in Porto. Where machines should end and humans should begin.