The Work That's Leaving

Jack Dorsey just cut Block from 10,000 to under 6,000 people. Stock up 20%. Klarna went from 7,000 to 3,000. Duolingo declared itself “AI-first.” Amazon cut 30,000 corporate roles in two rounds.

Dorsey’s message to other CEOs: “I think most companies are late. Within the next year, I believe the majority of companies will reach the same conclusion.”

Ok, but what’s next?

The headlines are stuck on the symptom. This isn’t about layoffs. It’s about which work is disappearing, why, and what comes after.

The work that was only human because nothing else could do it

For decades, we hired people for tasks that were too complex for rule-based systems. That’s changing fast. Models get more capable every quarter. Infrastructure gets better. Tooling matures.

The first tasks to go are the ones where reliability is provable and blast radius is contained.

Coding is a natural starting point. LLMs are trained on massive amounts of code, so reliability is high. Blast radius stays contained through pipelines and tests: if the code doesn’t work, you catch it before it ships. Karpathy called December 2025 the moment AI coding agents crossed from unreliable to functional. Microsoft says 30% of its code is now AI-written. METR’s benchmark shows the task length where AI reaches 50% success rate has been doubling every seven months.

Call centres follow the same logic. Customers often face similar challenges, and there’s a wealth of data on previous questions. Reliability comes from limiting what the agent can answer on and passing off to a human for anything outside that scope. The model doesn’t know when it’s wrong, so you design the workflow so it doesn’t have to. Klarna’s AI assistant now handles the workload of 700 full-time employees.

Web design is next. Most websites follow common patterns: landing pages, dashboards, e-commerce layouts. The building blocks are well-documented, the output is visual and easy to verify. Graphic design follows the same trajectory: generating styled artwork for social media, presentations, marketing materials. When the patterns are common and the margins are already tight, good enough wins.

These aren’t random starting points. They’re the use cases where the dimensions line up: high reliability, contained blast radius, clear business value. As models improve and infrastructure matures, agents will move into higher-stakes territory. The frontier keeps moving.

There is no shortage of work

This is the part that gets lost in the layoff headlines. The volume of valuable work is not shrinking. If anything, it’s growing. But it needs different offerings. Different skills. Different organisational shapes.

When agents handle the predictable work, your people can focus on what actually requires human judgement: strategy, relationships, creative problem-solving, navigating ambiguity. The work that machines are still bad at and will be for a while.

But “human in the loop” is not a stable position if all you’re doing is overlooking agent output. That’s a formality, and formalities get automated. The human role that lasts is at the root of the intent: defining what should happen, making the calls that require judgement, owning the decisions an agent can’t.

Oversight without agency is just a transition state on the way out.

The transition doesn’t happen by itself either. You don’t just remove the automatable work and hope the rest fills in. You have to redesign: which processes agents take over, how humans work alongside them, what new offerings become possible.

We can’t keep training for the factory

Our education system was designed for Prussia: produce uniform workers for uniform tasks. That model carried into the knowledge economy. Train people in a specialism, slot them into a role, repeat for 40 years.

That breaks down when your specialism gets automated overnight. We need people who understand fundamentals deeply enough to adapt when the tools change. Creativity, critical thinking, the ability to work with ambiguity. Not as soft skills on a CV, but as the core of what makes someone valuable when the predictable work is gone.

And we need a fallback. When entire categories of work get automated faster than anyone expected, people need a way to transition. Not retraining programmes that take three years to catch up with a landscape that moves every quarter. Something faster, more modular, more connected to what’s actually happening.

What is a company, anyway

Maybe this is the moment to step back further. A company is a group of people organised to deliver value. For decades, that meant hiring across every function: marketing, design, development, accounting, HR, legal. Each role existed because the work couldn’t be done any other way.

If agents can handle significant parts of those functions, what does a company actually look like? Maybe much leaner. Maybe a five-person team with agents delivers what used to take fifty. Maybe the default is no longer to hire for every function, but to ask which functions still need a human in the loop.

And the agents don’t have to be yours. Another company might offer you the sharpest marketeer you’ve ever worked with, as a service. The people who used to be staff become independent, packaging their expertise with agents and selling it to many clients instead of one employer. The line between employee and company blurs.

In the end, work is transactional. Someone needs something done, someone else does it. We’ve wrapped layers of structure around that basic exchange. Some of those layers exist for good reason. Others exist because we never had an alternative. Now we do.

Start ahead, start structured

This is incredibly disruptive. And we’re boiling frogs. The temperature has been rising for months, but the layoff headlines still feel like someone else’s problem. They won’t for long.

Things will change regardless. The question is whether you have an answer for your team when what they do gets commoditised.

The companies that start their agentic transformation now get to redesign around it. The ones that wait will be explaining to their people why the work disappeared and there’s no plan.

The ones getting this right aren’t the ones making dramatic cuts. They’re the ones mapping the shift before it’s forced on them. Identifying what’s automatable, understanding the blast radius, and building the infrastructure to make the transition work.

The opportunity is in what comes after: the work that agents unlock, the offerings that weren’t viable before, the human capacity that was buried under tasks a machine can now handle.

Start the transition before your competitors force it. And do it in a structured way.

The question we’re not asking

We’re freeing up enormous amounts of human time. That should be good news. But as a society, we’re not good at valuing what doesn’t come with an invoice.

We have lonely elders. Children growing up barely seeing their parents. Hours lost in traffic every day. There is no shortage of meaningful things to do with freed-up human time. We just don’t value them economically. If we’re about to produce more with fewer people, the question isn’t how to keep everyone busy. It’s how to let the surplus flow to the things that actually matter.

And we need a society that’s graceful about the transition. Not everyone will be on top of everything, and it’s unreasonable to demand it. The productivity gains flow to the companies that automate, not to the people who get displaced. Until that changes, we’re capturing the output and losing the people.

We need children who grow up understanding what to build agents for and how they can be implemented. Who see these systems as tools to direct, not forces to fear. Who can reason about what’s worth automating and what isn’t. That’s not a curriculum update. It’s a different starting point entirely.