AI 原生公司如何重建管理职能

There is a crazy rumour going around that AI will make managers unnecessary. That AI will allow you to operate a company with teams that stay small on purpose, where agents do real work alongside people, and one person delivers what used to take a department.

In late 2024, Gartner predicted that by the end of 2026 one in five organisations would use AI to flatten their structure, cutting more than half of their middle management positions.

This is more than a rumour. AI-native companies like Block, Cognizant, Lovable, and Every have begun to put these ideas into practice. And they are seeing results. We should all be paying close attention to what they are doing.

It’s akin to the first factories after electrification. Most owners simply swapped the steam engine for one big electric motor and left the factory floor exactly as it was. Their productivity didn’t move for decades. The winners put a small motor in every machine and rebuilt the layout around the work. They didn’t just adopt a new technology and retrofit it to their existing workflows. They redesigned around it.

Look closely at these companies and it becomes easier to see why they are getting results from AI, and why others are shelving their AI initiatives entirely. They clearly don’t have it all figured out just yet. Nobody finds the clean version on the first try. But there are some interesting patterns emerging.

The best descriptions of where this is all headed come from the tech companies at the frontier. Jack Dorsey, CEO of Block, and Roelof Botha wrote a widely shared article, From Hierarchy to Intelligence, about how they’re rebuilding Block around AI. Then, Elena Verna, a well known growth operator working at Lovable, wrote her field notes from a year inside a truly AI-native company. And finally, Dan Shipper, CEO of Every, wrote “After Automation” about how his company works now that it has automated everything it can.

To make this transformation, you first have to change your beliefs about how a company is supposed to be managed: the belief that a company is its org chart, that managers keep the people beneath them aligned, and that meetings are mainly for passing information around.

“The problems that are most resistant to solutions are the system problems. In a system problem, if you’re part of the system, you’re part of the problem.
Your biggest blind spot is yourself.”

— Dave Gray , author of “Liminal Thinking”

This is my attempt at translating what these new beliefs mean for everyone else. The hard part of reading this won’t be the new ideas presented. It will be to let go of the obvious ones you assume are settled.

Here is what this article will cover:

1. What’s changing for the individual

2. What’s starting to break as the system transforms

3. Where the bottleneck in the system is shifting

4. What new roles are emerging as agents join the team

5. What management infrastructure is missing to make it all work

Let’s start by looking at how this shift is impacting employees. There are seven shifts going on at the individual level. The seventh is the most important.

1. The lines between job titles are fading: At Every, people in operations and customer service do work that used to need an engineer, and all engineers talk directly to customers. At Lovable, senior people who used to have dozens of direct reports are happily building things themselves. Traditional functional job titles hold less meaning. The impact of this has been measured too. In a randomised trial with 791 professionals at P&G, people without AI stayed in their lane: R&D professionals proposed technical solutions, commercial professionals proposed commercial ones. Given AI, both groups produced balanced solutions whichever side they came from, and the quality of work improved.

2. Everyone can ask the company what it knows. Jack Dorsey called it a World Model. Lovable calls it the Company Brain. Both are centralised systems that ingest every useful data point from the company’s various disparate systems, every Slack message, document, meeting transcript, and KPI to create a continuously updated digital twin of the organisation. It’s highly relevant context available to everyone, all the time, so everyone operates from the same basic information.

3. Agents do real work and everyone is a manager of agents. At Every, when you tag a colleague in Slack it’s 50/50 whether you’re talking to a person. At Lovable, every agent has a named human responsible for what it produces; a knowledgeable “agent parent”. Directing, reviewing and correcting an agent is a normal part of everyday work.

4. People take on more ambitious work than they could have tried before. The ceiling on what one person can attempt has gone up, so the work that was too expensive to start becomes possible.

5. With fewer permission gates, everyone becomes a founder. When an experiment costs a few tokens and a day, nobody needs a committee to approve it. The controls haven't gone. They’ve just moved. As Alex Lieberman describes it, this means a non-technical employee can take an idea to production with the governance built in. This used to be called intrapreneurship, and was reserved for a designated group. Not anymore. A tool someone builds for themselves can become a team’s agent, then a shared pattern, then an internal product.

6. It feels good. In Lenny Rachitsky and Noam Segal’s survey of tech workers this summer, 82% said AI makes them measurably more productive and 49% are said they feel amplified (able to do more, and better). The P&G trial measured the emotional side and found that people who worked with AI came away with more excitement, energy and enthusiasm, and less anxiety, frustration and distress, with the biggest lift of all for teams using AI. Working alone with AI matched or beat the emotional boost people get from having a human teammate.

The biggest shift for the individual is in the unbundling of their value. With AI agents doing more of the production work, they’re discovering what makes them uniquely valuable — their judgment, taste and intent.

To understand what 'human judgment' means, it helps to be precise about what kind of knowledge judgment actually is.

The cognitive scientist John Vervaeke splits knowledge four ways:

• propositional: knowing that

• procedural: knowing how

• perspectival: knowing what a situation looks like from where you're standing;

• participatory: knowing what it is to play your part in it.

AI absorbs the first two kinds of knowing. Real human judgment exists in the last two.

A diagram in Dan Shipper’s article does the best job of explaining how judgment becomes a unique advantage in the AI era.

Essentially, you have two converging realities. On the left side, LLMs learn from work that has already been done, which opens the floodgates for every individual regardless of title, experience or function to make an attempt at things. To make use of an increasingly intelligent corpus of propositional and procedural knowing.

And on the right, you have the humans who know the live situation. The people who can see it from where they stand, who know what matters right now, and who are part of the thing rather than just looking at it. That is judgment formed from perspectival and participatory knowing. And it is something AI can never do.

Jack Dorsey says something similar in his article, about how people sit at the edge of the company, where intuition, trust and “feeling the room” really matter. Elena Verna explains how people ask themselves: “but is this lovable?” before they commit to a path.

“The weird paradox is that the more AI removes the mechanics of work, the more the company seems to depend on human judgment. When information is everywhere, execution is cheap, and everyone has a crazy amount of leverage, knowing what to do, when to interfere, what good looks like, and when something is actually done becomes disproportionately important.”

— Elena Verna, Growth at Lovable

While everything described up until now sounds exciting, there are some less welcome by-products showing up as these companies transform. It's still early and correlation isn't causation, but the companies furthest ahead are already seeing where the costs accumulate. Here are a few:

• More gets made, but more of it looks the same. When everyone can produce a decent draft of anything using a tool that is designed to spit out an average of everything that has already been done, someone with real expertise still has to review it, improve it or throw it away. Otherwise we converge on sameness. And sameness is a commodity. Cheap execution makes bad judgment more expensive.

“Slop is visible sameness, repeated ad nauseam.”
- Dan Shipper, CEO of Every

• Work collides and stresses people out. Since starting something is cheap, people start all sorts of things. They don’t wait to be assigned things, and they don’t check with anyone first. So two people pick up the same problem from different ends, and neither finds out until the work is done. Collisions are a major source of uncertainty for people. And heightened uncertainty is one of the leading causes of burnout.

“There are definitely moments where ‘high agency’ is just a flattering way of describing duplicated work, unclear accountability, or someone stomping on somebody else’s toes. Sometimes the chaos produces a better outcome. Sometimes it’s just chaos.”
— Elena Verna, Growth at Lovable

• Managers stop hearing things. A TNG survey found that 60% of respondents ask AI, at least once a week, questions they would have previously asked their manager. When people take their questions to AI, managers lose one of the main ways they find out where things are unclear.

• Being an effective manager is getting harder. It was a hard job even before AI. In a 2025 survey of tech workers, only 26.6% rated their manager as highly effective, and 42.3% rated theirs as ineffective. AI makes it even harder. Managers have less to go on, because the questions that told them what was unclear now go to AI. They have more to hold together, because work moves faster and collides more often. And they're squeezed from above: in TNG's survey, line managers were the group most worried that leadership expects too much from AI, at 36% against 23% of senior executives. This matters because of what an effective manager is worth. People with great managers reported 63% more belonging and 31% less burnout than people with ineffective ones.

• People are moving faster and feeling worse. In the survey of tech workers, burnout rose from 44.7% to 55.7% in a year. People worry about being expected to do more for the same pay (51%), an unsustainable pace (46%) and the quality of their work going down (41%). The same fears show up in Sweden. In TNG’s survey, the top concerns were wrong data leading to wrong decisions (51%) and human judgment being pushed aside (50%).

There’s a paradox worth stopping for here. If judgment is where human value is moving, how can it be that it is the thing 50% of people surveyed are afraid of losing?

Part of the answer is that we're not all talking about the same thing. At Lovable, people automate themselves on purpose:

“’Automate your own job’ sounds terrifying if you think the next words are gonna be ‘...and then we no longer need you.’ But here it’s more like ‘...so you can go figure out what your job should become next.’”

— Elena Verna, Growth at Lovable

What the people at Lovable are automating are the tasks that run on propositional and procedural knowing (i.e. the what and how tasks). What they aren’t voluntarily giving away is the judgment that decides whether the work was any good. Judging AI output mostly means spotting mistakes, questioning assumptions and rejecting sub-par quality.

Unfortunately, almost nobody is measured on that. KPIs are often about speed, volume, output. You don’t get rewarded for slowing down because something isn’t quite good enough. Add the pressure to do more in less time, and the rational move is to hand more of the judgment to the AI. Even when you shouldn’t.

The second reason is more dangerous. AI has separated our ability to produce something from our ability to judge it. A lot of what we now attempt sits outside our expertise. So we send the draft to someone whose judgment we actually trust on the matter, someone with real domain knowledge. The cost of judging lands on the person who didn’t make the attempt in the first place. As explained in the diagram above, the experts are now flooded with things they need to pass judgment on. So who could fault them for outsourcing some of that judgment to AI? If there is little incentive to catch mistakes, why would you risk burnout trying to spot them in someone else’s work?

“The failure case now of lazy work is not lack of output, it’s over output […] We are all tossing slop grenades at each other.”

—

Tobias Lütke, CEO of Shopify

Lastly, there may be something more sinister going on. Something I suspect many experts are beginning to think about after Meta began installing software on employees’ computers to capture mouse movements, clicks and keystrokes in order to train their AI models. It’s what economists from economists from MIT, Washington University and UCLA wrote about in a paper called Some Simple Economics of AGI. In it they describe “the codifier’s curse”, which explains that our current “human-in-the-loop” equilibrium is unstable. The experts carrying the judgment are, in the course of doing their job, converting their judgment into training data. They are codifying their own obsolescence.

“Deploying unverified systems becomes privately rational […] measured activity rises, but hidden debt accumulates in the gap between visible metrics and actual human intent.”

— Christian Catalini, Xiang Hui and Jane Wu from MIT, Washington University and UCLA

Now let’s step back. Every system has one step that sets the pace for everything else. Speed up the other steps and work just piles up in front of it. That's the bottleneck.

Previously, production was the bottleneck. Making things was the expensive part, so we coordinated in order to produce. But AI has flipped that.

The cost of production has fallen below the cost of coordinating.

Producing is now the easy and cheap part. Deciding what to produce, agreeing on it, and making sure multiple capable people aren’t all solving the same problem in different ways without even knowing it, is not. Coordination, which includes alignment, is the new bottleneck.

The reason coordination has become more difficult than before goes back to Dan Shipper’s digram, which I have repurposed to make my point.

The numbers used are illustrative. It’s the gap between them that isn’t.

So let’s say execution has gone up by a hundred times. Expert judgment has gone up too, because an expert can now pass judgment on more work than ever before. But human cognition is finite, so let’s say that’s gone up by ten times.

But coordination hasn’t moved at all. Who decided what, who owns what, what’s still open: that still runs at the speed of people talking to each other, the same as it did twenty years ago.

So both sides of the diagram converge on the one step that didn’t get faster. That’s where the queue forms.

It’s tempting to think the Company Brain or World Model alone solves the coordination problem. It solves the first step. It gives every person information.

But individual access to more information isn’t the same as shared sense-making or committed action. Knowing more does not automatically mean people understand the situation in the same way, agree on what matters most, or know what should happen next. That still occurs in conversations between people.

Elena Verna’s story of work colliding happened in a company where almost everything is visible to everyone. So what was missing wasn’t information. It was agreement.

In their article, Jack Dorsey and Roelof Botha put forward three new roles. None of them are managers.

1. The Directly Responsible Individual (DRI), who owns a single outcome for a set period and can pull in whoever they need.

2. The Player-Coach, who still does IC work and develops the people around them. They replace the traditional manager whose primary job was information routing.

3. The Individual Contributor (IC), a specialists and experts who, because of the world model, can make better decisions without being told what to do.

“There is no need for a permanent middle management layer. Everything else the old hierarchy did, the system coordinates, and everyone is empowered, with a role that’s much closer to the work and the customer.”

—Jack Dorsey and Roelof Botha

In an episode of the AI Daily Brief entitled “Why agents make every job a startup”, host Nathaniel Whittemore goes further. He lists nine new jobs AI-native companies are likely to need, grouped into three categories:

1. Technical roles: agent ops engineers keep the fleets of agents running. Context librarians curate what the agents know. Eval engineers build the quality checks their output has to pass.

2. Managerial roles: experiment portfolio managers decide which bets to fund, scale, merge or stop. Intrapreneur coaches help people with judgment and pace. Internal agent product managers look after the company's own agents as products.

3. Coordination roles: Coordination architects design how the organisation stays legible. Information pipeline owners route the right signal to the right place. Orchestration leads broker the conversation when work overlaps.

“Don’t let parallel brilliance become org entropy”

—Nathaniel Whittemore

Those three coordination jobs, which include keeping people aligned, are real work. They are what an effective manager does, and always has been. And today they are a company's greatest point of leverage. Look back at the diagram I adapted from Dan Shipper to understand why.

Everything AI produces and everything the experts know has to pass through this one step of coordination. Done well, it turns a flood of output into actual differentiated value. Done badly, you get commoditised sameness, faster. Whether it’s done well comes down to people.

The judgment, taste and intent that make the difference only show up when people are engaged and committed to their work. In the survey of tech workers, the people who rate their manager as effective are the ones who report being engaged and committed. And the gap is enormous. Workers with extremely effective managers are 8.2 times as likely to be committed to their role as workers with ineffective ones: 61.6% against 7.5%.

So what a company most needs from its people depends on the person with the least support. An engineer has a whole toolchain, and a salesperson has a CRM. The person responsible for getting the best thinking out of a group currently has a calendar and their own memory.

What’s missing is infrastructure that makes managers more effective.

“The most powerful retention lever in tech is also the most neglected.”

—Lenny Rachitsky and Noam Segal

Jack Dorsey and others are right that companies move at the speed of information flow, and that hierarchy slows it down. And the World Model fixes that. It gives people what they need to know to make good decisions.

But managers did more than that. They were coordinating the whole dang thing.

Coordination has two parts: information is one and alignment is the other. But alignment isn’t something you retrieve. It gets made, between people, in conversation. That’s where a group works out what the information means, what matters, and what they’re going to do about it. It’s where good judgment on complex problems is exercised.

As agents handle more of the routine and the complicated, complex problems will increasingly be the domain of people.

If the quality of collective human judgment is what separates differentiated value from commoditised sameness, then the place where that judgment is formed deserves its own infrastructure.

Put the whole picture together, and a company needs three things to become truly AI-native.

Take any one away and you get some or all of the problems listed earlier.

• Information and judgment without alignment is the colliding work that Elena Verna wrote about.

• Information and alignment without judgment is what the sameness Dan Shipper describes, well coordinated.

• Judgment and alignment without information is where most companies were before AI: people who agree, but on incomplete, uneven and sometimes wrong information.

In the AI era, differentiated value sits in the middle, where all three meet. The World Model is being built by every AI-forward company. The people with good judgment are already there. The third circle is the one with nothing underneath it.

And it matters, because having judgment alone isn’t enough. A company only benefits from what its people know, notice and decide if it can turn that collective judgment into coordinated action.

This is why the conversations among experts have to produce more than they do today, and why what they produce has to become an enduring company asset. A reliable record that lets increasingly independent people (and agents) act in a coordinated way.

For that, the record has to be reliable. And it’s only reliable if the experts who were in the conversation keep authority over what they meant. AI can interpret what a group of people said. It can’t decide what they are standing behind. The people involved have to confirm their intent before anyone else can act on it.

This is what we’re building at Afoot. It’s the coordination infrastructure beneath the people with the best judgment.

Afoot listens to the conversation and works out what was decided, who committed to what, and what’s still open. Then the people who were in the room confirm it. And from there, a team of agents does the coordination work. Because the agents maintain the record, they know who is doing what, they route the right information to the right people, and they prevent work from colliding. Their starting point is always the reliable record of human intent.

With that in place, the new roles described earlier can work as intended. The DRI who owns an outcome doesn’t have to chase it, and the Player-Coach doesn’t have to carry everyone’s context.

What’s left for both roles is harder and worth more. As Houda Nait El Barj writes, the more of our day we spend working with AI, the more we need human connection, inspiration and direction.

“The managers who thrive will look less like operations executives and much more like spiritual leaders.”

— Houda Nait El Barj

I’d put it more plainly. The leader’s job becomes orientation and facilitation: helping a group see the same situation, decide what matters, and mean the same thing when they agree. That was always the part of management worth having. We just never gave anyone the time and support to do it effectively.

AI can make each of us more intelligent. But agreeing is still something we have to do with each other.

Afoot goes live soon and the waitlist is growing. What we’re looking for now is the right design partners: a few companies that feel the pull towards working this way and want the coordination infrastructure built around how they operate.

If that’s you, email me at [email protected], or join the waitlist to follow our journey.

Join the Waitlist

No posts