The Last Generation of Routers

Most of what we call management was always information routing. AI is now eliminating that layer, on the order of tens of thousands of positions per quarter. What is left of the job, and who is left to do it, is being decided right now.

Something is happening to the management profession that has not happened in fifty years. The job is being rebuilt under our feet, and most managers can feel it before they can describe it.

Consider the numbers.

In 2013, the average American manager had 8.2 people reporting to them. By 2024 that was 10.9. By 2025 it was 12.1. Thirteen percent of managers now run twenty-five or more direct reports, a span that any management textbook from the last century would have called impossible. Ninety-seven percent of managers are now doing individual contributor work alongside leading their teams.

The middle has been hollowed out at the same time. Job openings for middle-management roles have fallen by more than 40% since 2022. Forty-one percent of employees say their company has cut middle management in the past year. Gartner projects that by the end of this year, one in five organisations will use AI to remove at least half of their management layers.

The named cases are loud. Block has cut 40% of its staff, about 4,000 people. Amazon has eliminated roughly 30,000 corporate roles since October 2025. Meta is planning a 20% reduction this spring. Klarna has gone from 7,000 employees to under 3,000, with the CEO publicly targeting under 2,000 by 2030.

None of this is being framed as a downturn. It is being framed as an upgrade.

To make sense of what is being upgraded, it helps to remember what the manager job was originally for.

Peter Drucker, who invented modern management as a discipline, coined the term “knowledge worker” in 1959 and spent the next forty years arguing that knowledge work would replace industrial work as the defining activity of the developed economy. He was right about that. He was also right about something subtler: managing knowledge workers was fundamentally different from supervising factory workers, because knowledge workers, by definition, knew more about their job than their manager did.

That paradox forced a redefinition. If you cannot supervise the work, what is the manager for?

The honest answer, for half a century, was: routing. The manager existed because information could not flow efficiently across a large organisation by itself. Someone had to gather context from below, push priorities from above, align teams sideways, translate between specialist languages, and resolve the small daily conflicts that came from people not having each other’s full picture. The manager was the human protocol layer of the company.

That worked. It also created the modern white-collar career: the ladder where you started doing the work, then stopped doing the work, and then spent the rest of your career routing other people’s work.

That is the job that is now being rebuilt. For the first time, the routing layer can be replaced by software.

The manager job decomposes into roughly five things. Route information up the hierarchy. Route information down. Align teams sideways. Supervise the work being done. Develop the people doing it.

AI now does the first three better than any human, at any company, at any scale. This is no longer a frontier claim.

In finance, Hewlett Packard Enterprise built an internal agent called Alfred with Deloitte that has cut its financial reporting cycle by about 40%. More than half of surveyed CFOs (54%) say integrating AI agents will be a digital transformation priority this year. The 2026 finance department is described in the trade press as a “Human + Agent collaborative”: agents handle reconciliation, invoice processing, variance detection; humans interpret anomalies and partner with the business.

In HR, organisations using agents for screening, scheduling, onboarding workflows and routine employee questions are reporting 50% faster time-to-hire. ADP’s new Payroll Variance agent is saving HR operations teams roughly thirty minutes per payroll cycle just by answering plain-language questions about discrepancies.

In customer support, Klarna’s AI handles work equivalent to more than 850 human agents. Revenue per employee at Klarna has gone from about $340,000 in 2022 to roughly $1.2 million today.

These are not pilots. These are line items in the operating plan.

What AI does not do, and shows no sign of doing soon, is the last two of the five. It does not own outcomes. It does not develop people. Those are the two pieces that still require a human.

That leaves every existing management role with a stark question. If three-fifths of what you used to do is now being done by software, what is left of your job? And the answer that is emerging, across industry after industry, is that what is left is to own the outcome and to build the system that produces it.

Share Tom's newsletter

The CEO who said this earliest and most clearly was Tobi Lütke at Shopify. In April 2025 he sent a memo, later posted publicly, that became a quiet inflection point for how serious companies are now operating.

Two lines from the memo travelled. The first: “Reflexive AI usage is now a baseline expectation at Shopify.” The second, harder one: before any manager could request additional headcount, they had to prove the work could not be done by AI.

The memo was not a productivity exhortation. It was a filter. It told everyone in the building that the floor of the job had moved. Within eight months, AI fluency in job postings had nearly doubled, from 5% to 9% of all listings. The number of workers in occupations requiring AI fluency had grown from one million to seven million in the same window.

Shopify did not invent this. It said it out loud first.

In 2024, Paul Graham wrote an essay about something he called “founder mode.” He had noticed that the best founders he knew systematically refused to do what management consultants told them: namely, to “hire good people and give them room to do their jobs.” Those founders ran skip-level meetings. They got involved in decisions five layers down. They broke the org chart whenever the org chart got in their way.

Graham’s framing was that founders were a special case, with a special licence to operate this way.

He was half right. The founders he was describing were special, but not because they had a special licence. They were special because they were already doing what the rest of the management profession will be forced to do in the next decade. They were treating the routing layer as overhead, and going around it.

The reason that workaround now generalises is that the routing layer no longer needs to be staffed at all. It can be built. And once a thing can be built rather than staffed, the question for every manager, at every level, in every industry, becomes: are you the builder, or are you the routing layer?

There is no third option. The old role of the manager-as-router is being eliminated, by AI, on the order of tens of thousands of positions per quarter. The new role of the manager-as-builder is being created, by AI, faster than the supply of people who can do it.

This is what I mean when I say the old founder mode is dead and the new founder mode is everyone. Founder mode was never about being a founder. It was about being the kind of person who builds the system rather than supervising it. That kind of person used to be a luxury. They are now the standard.

The new job is not theoretical. It can be described concretely, role by role.

The head of HR no longer routes CVs to hiring managers and chases status updates. She builds the recruitment system: which agent screens, what the screening criteria are, where the human steps in, what the interview workflow produces, what the onboarding sequence looks like, what numbers come out the other end. She owns time-to-hire, quality-of-hire, and retention. The system is the job.

The CFO does not “oversee” the month-end close. He builds it. He decides which reconciliation runs as an agent, which exception handling is escalated to which human, which review is automated, which is preserved for judgement. He owns the close calendar in days, the audit trail, and the variance reporting cycle. The system is the job.

The COO of a manufacturing plant does not “supervise” the floor. She builds the optimisation loop: which sensors feed which model, which dispatch decisions are automated, which require a human, what the throughput targets are and how they are revised when reality changes. She is fluent in her own factory the way a software engineer is fluent in her own codebase.

The head of sales no longer “manages the pipeline.” He builds the system that produces qualified pipeline: the targeting model, the outreach sequencing, the qualification logic, the handoff to humans, the closing playbook. He owns conversion, ACV and sales cycle.

The CEO is in the engine room with everyone else. She uses the same tools as her engineers and operators, because she needs to feel the constraints and the leverage in her own hands. She is not on the press tour. She is closing tickets and looking at dashboards she built herself.

None of this is a startup story. The function-level rebuild is happening at HPE, ADP, Klarna, Shopify, Amazon and Meta, none of which are startups. It is happening at banks, insurers, hospitals and law firms, more quietly but on the same schedule.

Two arguments are commonly made against this view. Both are wrong, but they are wrong in interesting ways.

The first is that AI is just a productivity tool, like the spreadsheet or email, and managers will absorb it without changing the shape of the job. This is the position you hear most often in established companies, usually from people whose job depends on it being true. The empirical record is now strongly against it. The shape of the job has already changed: spans of control have nearly doubled in twelve years, middle-management roles have collapsed by 40% in three, and entire functions are being rebuilt around agents at named companies. This is not the spreadsheet. The spreadsheet did not eliminate accountants; AI is eliminating the routing function that the entire middle of the org chart was built around.

The second is the opposite extreme: that AI will simply replace the job entirely, and there is nothing for the manager to learn. This is the Klarna-coded view, and it has the advantage of being half-true in the short term. But it ignores the most interesting voice in the debate, which belongs to Patrick Collison at Stripe. Collison’s view is that AI is “the GLP-1 for efficiency”. It does not just help you cut. It lets you do dramatically more of what creates value. Stripe is hiring more in 2026 because of AI, not less. The question Collison asks is not “what can we cut?” but “what would we elect to do much more of, if we suddenly had the leverage?”

Both arguments fail for the same underlying reason. They both treat AI as something that happens to the manager. The reality is that AI is something the manager has to do. It is the new tool of the trade. If you are not the person building with it inside your domain, you are the person being built around.

This changes what good looks like in a candidate at every level. The signals are the same whether I am hiring an analyst or a CFO. The artifact gets bigger. The shape of the person stays the same.

What I look for:

• A candidate who has shipped something this quarter that exists outside their head and that someone is using. Software, a workflow, a process, a product. They can describe it, defend it, and tell me what is wrong with it.

• A candidate who uses AI tools (Claude Code, Codex, Cursor, ChatGPT, the underlying agent platforms, whatever) daily, by reflex. They can show me their actual usage. They have opinions on which tool is better for which job.

• A candidate who can describe a workflow they own as if it were a system: inputs, transforms, outputs, exceptions, escalation paths. They do not say “the team handles it.”

• A candidate who owns a number (a P&L line, a conversion rate, an SLA, a defect rate) and recites it from memory.

• A candidate who, when asked what is broken in the system they own, has the list ready.

• A candidate who has built or rebuilt something that an organisation three times their size still does manually.

What I now treat as a red flag:

• A candidate who cannot describe their job without the word “stakeholders.”

• A candidate who optimises for the meeting rather than the work it is meant to produce.

• A candidate who has more direct reports than shipped systems.

• A candidate who talks about “the team” doing things they could not do themselves.

• A candidate who treats AI tools as something the engineers use.

• A candidate who has a brilliant strategy and no Tuesday.

• A candidate whose deck is better than the work behind it.

None of this requires the person to be a programmer. It requires them to be a builder. The HR director who has built the screening agent passes. The MBA who has shipped a working compliance workflow with an agent passes. The forty-year-old “experienced operator” who has been a director of operations at four companies and cannot show me a single thing he has built does not.

This is uncomfortable. Particularly for senior people who built their careers in the previous game and who, in many cases, are exceptional at the previous game. I have had to look at my own bench and make hard calls. So has every honest CEO I know.

I want to be careful with this part, because it is the most likely to be misread.

The Draghi report, published in 2024, laid out the European productivity problem with a clarity nothing before it had managed. Thirty years ago, EU productivity was about 95% of the US level. Today it is between 75 and 80%. GDP per capita in Europe has fallen to 77% of the US. The gap grew by 12% in 2023 alone, of which 72% was attributable to productivity.

A year and a half after Draghi, only about 11% of his 383 recommendations have been fully implemented, and another 20% are partially in motion.

There are many causes for the productivity gap, and I am not claiming to have the full picture. But one piece of it is now structural and is about to become decisive.

Europe has a long-standing overweight in operator talent and an underweight in builder talent. We are very good at running systems we did not build. We are less good at building systems no one has run yet. Our most prestigious talent funnels (the global consultancies, the top business schools, the corporate ladders of legacy industry) are excellent at producing routers. They produce, with great consistency, the kind of person whose career was perfectly suited to the manager-as-protocol-layer job that is now being eliminated.

This was a manageable disadvantage when the routing job still existed. It is becoming a structural one as the routing job goes away. Every quarter that we keep hiring routers when we should be hiring builders, the gap with the US, and increasingly with China, widens. The bottleneck is not capital. Europe has capital. The bottleneck is the talent stack.

The fix is not to import a different culture wholesale. It is to reset the hiring funnel inside European companies, at every level, including the C-suite, so that the candidates with the most prestigious CVs are not given automatic priority over the candidates who can actually show you something they built.

This is not a plea for less rigour. It is a plea for different rigour. The rigour of building, not the rigour of supervising.

There is a simple test now. What did you build this year? Not what your team built. What you built. If the answer is nothing, the job is going away.

That is true for the founder. It is true for the individual contributor. It is true for the manager in the middle who used to be safe.

For the routers, this is the end of an era.

For the people who can build, the ceiling just disappeared.