AI is changing the organisation around the job

I read Elena Verna’s field notes from working at Lovable expecting another account of an AI-native company doing familiar work faster. That is not really what interested me.

The striking part is how much of the organisation around the work appears to have changed. Lovable has very few titles. Senior people are happy working as individual contributors. Meetings are treated with suspicion. Roles overlap. Teams reorganise constantly. People try to automate parts of their own jobs. AI agents have human “parents” responsible for keeping their output useful. When something is being judged, one of the questions people apparently ask is simply: is it lovable?

Verna is careful not to present any of this as a universal playbook. She has one data point, and she explicitly says she is not always sure which parts are AI-native, which parts are Swedish and which parts are simply Lovable being Lovable. That caveat matters. I do not think one unusually fast-growing software company tells us how everybody will work.

But it does surface a more interesting question than whether AI can write an email or summarise a meeting.

How much of the modern organisation was designed around the old cost of moving information, coordinating people and getting work done?

If those costs change, the job changes. But perhaps the structure around the job changes too.

Titles are partly an information architecture

The no-titles point sounds superficial until Verna explains what disappeared with them.

At many companies, titles do much more than describe responsibility. They determine who gets invited to a meeting, which Slack or Teams channel somebody joins, what information they receive, which decisions they are allowed to make and whose opinion carries weight. Promotion changes the label, but it can also change the information system surrounding the person.

That makes a hierarchy partly an information architecture. Someone knows something. It travels up. Somebody more senior gets involved. A decision travels back down. Certain people are copied because their position says they should be. Management becomes one of the mechanisms through which the organisation moves context from one place to another.

Lovable appears to weaken some of that machinery. Verna describes public Slack channels, people opting themselves into and out of conversations, and information that does not move exclusively through management chains.

AI potentially makes that easier because the cost of recovering context is also changing. If people can search a shared knowledge base, interrogate company information, reconstruct why a decision was made and use agents to surface expertise, then access to context does not have to depend quite as heavily on being in the right meeting or knowing the right person.

That does not make hierarchy obsolete. Legal accountability, financial authority, regulated decisions, sensitive information and genuine differences in expertise do not disappear because retrieval gets better. It does suggest that some hierarchy may have been doing two different jobs at once: allocating authority and routing information.

AI can interfere with the second without necessarily changing the first.

Management starts to look different when it is not the router

This connects directly with something I wrote in AI is making your job description less useful. I argued there that AI can reduce dependence on a manager for information while increasing the value of a manager who provides context and direction.

Verna describes almost exactly that shift from inside Lovable. The management tasks she says seem less important are distributing information, coordinating work, approving decisions and assigning ownership. The things that remain are harder to automate: creating context, raising the quality bar, developing people and making sure autonomous people are moving in roughly the same direction.

The old manager could become valuable partly because they sat at an information junction. They knew what leadership wanted, what another team was doing, which project had priority, who needed to approve something and where a decision had stalled. If more of that becomes visible and retrievable, management has to justify itself somewhere else.

The stronger version of the role becomes less I know what is happening, so come through me and more I can help you understand what matters, make a better decision and become more capable of making the next one yourself.

That is not less management. It may be better management. It is also harder, because passing information down a chain is relatively observable. Creating enough context for autonomous people to make good decisions without turning autonomy into drift is a much less tidy job.

Seniority may stop meaning distance from the work

The section I kept coming back to was Verna’s description of very senior people returning to, or remaining in, individual contributor roles.

There has always been something odd about the conventional knowledge-work career ladder. You become good at the work, so eventually you are rewarded by doing less of it. Seniority becomes people management, then management of managers, then greater distance from the material that originally made you valuable.

AI makes another route easier to imagine.

One capable person can now research, analyse, prototype, write, build, test and iterate across a much wider surface area than before. That does not mean one person suddenly becomes excellent at every discipline. It does mean the leverage attached to an experienced individual contributor can rise dramatically.

I have been seeing a related signal in editorial work. The editorial jobs I keep finding are moving upstream: closer to evidence, narrative, systems, executive thinking and organisational knowledge rather than simply producing assets.

The Lovable version is broader. Seniority does not have to be expressed through the number of people beneath you. It can be expressed through the quality of the problems you can own, the amount of ambiguity you can absorb and the leverage you can create around your judgement.

That is a more attractive career model for people who are excellent at the work and never particularly wanted to become full-time managers. It also creates a harder performance question. Headcount is a crude but visible marker of scope. If senior people remain ICs, organisations need better ways to recognise influence, problem complexity, judgement and system-level contribution without quietly rebuilding the same hierarchy under different labels.

Removing titles does not remove status. Status will find another surface.

Meetings are a coordination technology

I obviously liked the part about being suspicious of meetings.

Verna says she spends less than 20% of her time in them, but the interesting part is why. It is not simply that Lovable has discovered calendars are annoying. Meetings perform functions: they distribute information, expose decisions, create shared context, resolve ambiguity and force an issue to become somebody’s problem for half an hour.

If you remove the meeting without replacing those functions, people simply become confused asynchronously.

Lovable appears to have different infrastructure underneath the low-meeting culture: accessible information, shared knowledge, AI tools that help people recover context, high autonomy and a willingness to correct decisions quickly rather than trying to perfect them in advance.

That is a system, not a meeting hack.

This is why I am sceptical whenever “fewer meetings” gets presented as an isolated productivity idea. A company with poor documentation, weak decision records, inaccessible information and unclear ownership probably needs the meeting because the meeting is compensating for everything else.

The better question is what information, decision or alignment job the meeting is doing and whether there is now a cheaper way to do it. AI may remove some of the coordination cost. It does not remove the need for coordination.

Automating your own job only works if the job can move

The strangest Lovable behaviour is probably the one that sounds most threatening outside an AI-native company: people apparently arrive and start asking how much of their own work they can automate.

That reverses a lot of career conditioning. Job security traditionally came from accumulating responsibility, knowing the awkward process, owning the spreadsheet, being the person who could answer the repeated question and understanding the workflow nobody had quite documented.

The implicit bargain was simple: if the organisation depends on the thing I do, it depends on me.

Lovable seems to offer a different bargain. If you can automate the thing you currently do, that demonstrates value because there is another problem waiting on the other side.

That works partly because the company is growing rapidly and the work itself is expanding. “Automate your job and find the next useful thing” feels very different inside a growing company from “automate your job so we can remove the cost centre”. I would not blur those two situations.

But the career logic is interesting. The durable asset stops being ownership of a repeatable task. It becomes the ability to understand the task well enough to improve it, encode it, hand it off and move to a harder problem.

That is a very different definition of expertise.

The expert becomes a system owner

Lovable’s “agent parent” idea may be the most useful detail in the whole piece.

The human is not necessarily the person repeatedly answering the question anymore. They are responsible for the agent that answers it: teaching it, updating it, correcting it and making sure the output remains useful.

That is close to the argument I have been making about editorial judgement becoming infrastructure. An editor who has corrected the same structural problem across twenty drafts should not need to wait for draft twenty-one before the lesson exists again. The useful move is to preserve the judgement so the next piece of work can inherit some of it.

The same principle applies far beyond editorial work. The finance expert who repeatedly explains a policy can improve the system that answers the first-line question. The analyst who knows which metric is routinely misunderstood can encode the explanation and examples around it. The product specialist whose context is constantly requested can make more of that context retrievable.

This does not make the expert less important. It changes the expert’s relationship with repetition. They move from being the endpoint of every request to being responsible for the quality of the system that handles more of those requests.

That can create enormous leverage. It also creates a governance problem. If an agent distributes one person’s expertise across an organisation, then an error, stale assumption or bad rule can scale just as efficiently as a good one.

The expert has not disappeared from the loop. Their responsibility has arguably increased because their judgement now travels further.

Remove explicit structure and judgement has to carry more

Then there is the wonderfully vague quality test: is it lovable?

It sounds almost anti-operational. There is no clean acceptance criterion for “lovable”. No spreadsheet cell turns green when the product has crossed the threshold.

Which is exactly why I think it matters.

The more explicit structure Lovable removes, the more it appears to depend on people having a strong internal model of what good looks like. No detailed title tells you exactly where your responsibility ends. No exhaustive process tells you how every decision should be made. No manager is expected to approve every move. Teams can reorganise. Roles can overlap. People can build things outside their nominal area.

Something has to stop that becoming complete randomness.

Part of the answer is context. Part of it is culture. Part of it is fast feedback. And part of it is judgement.

This is a pattern I keep returning to in my own work: AI makes the mechanics cheaper and therefore exposes the decisions around the mechanics. When content production is expensive, simply producing a competent article has value. When competent production becomes cheap, the question becomes whether the article should exist, what evidence it needs, what point it should make and whether it is actually good.

Lovable seems to be experiencing a similar shift at organisational scale. The more execution becomes available, the more valuable it becomes to know what to execute.

There is a cost to making the organisation this fluid

None of this is an uncomplicated argument for flatter, faster, AI-native companies.

Verna is unusually clear about the downsides. The pace is exhausting. Roles blur. People duplicate work. Ownership becomes ambiguous. Teams reorganise so frequently that returning from a week’s holiday can feel like rejoining a different company.

That matters because organisational structure exists for reasons other than bureaucracy. Boundaries reduce cognitive load. Stable ownership prevents duplication. Titles can clarify accountability. Repeated processes protect quality. Meetings can surface disagreement. A manager can stop five autonomous people solving the same problem in five directions.

The interesting question is not whether those structures are bad. It is what happens when the cost-benefit calculation around them changes.

A fixed structure that protects a bank from uncontrolled financial decisions is not equivalent to a fixed structure that makes a marketing team wait three days for permission to change a landing page. The right amount of hierarchy, process and role clarity will vary enormously with the stakes, regulation, reversibility of decisions and cost of error.

Lovable can move quickly partly because it appears comfortable correcting quickly. That is an important part of the system.

Maybe this is not actually about AI

There is an even stronger counterargument.

A lot of these ideas predate generative AI. Flat organisations, empowered teams, autonomous product squads, senior IC career tracks, asynchronous work, public internal information and high-agency startup cultures are not new inventions.

So it would be easy to overfit the entire organisational model to AI.

I think the more defensible interpretation is that AI can amplify a particular kind of organisation because it changes the economics underneath it.

A highly autonomous employee is more useful when they can investigate and prototype without assembling a team first. An open information culture becomes more useful when people can interrogate the information rather than merely search folders. A senior IC becomes more leveraged when tools expand the surface area they can operate across. A low-meeting culture becomes easier when decisions and context are recoverable. An expert can serve more people when some of their repeated knowledge is carried by an agent.

AI did not invent autonomy. It may make autonomy cheaper to support, and it may make the cost of poor judgement inside an autonomous system much more visible.

You cannot copy the visible behaviour

This may be the most important lesson in the whole Lovable account.

“Remove titles”, “cut meetings”, “give everyone autonomy”, “let people build agents” and “encourage people to automate their own work” can all be turned into management slogans by Monday morning.

That would miss the point.

The behaviours appear to work because they depend on one another. Fewer meetings depend on accessible context. Greater autonomy depends on a clear direction and tolerance for correction. Blurry roles depend on people being willing to collaborate rather than protect territory. Constant reorganisation depends on a culture that does not interpret every structural change as a personal verdict. Agents depend on humans who maintain the knowledge inside them.

Take one visible practice out of that system and the result could easily be worse.

I have been thinking about this in smaller form with my own work. AI can connect the silos, but somebody still has to understand which person needs which piece of context. AI can generate more possibilities, but knowledge work can become overloaded by how much remains alive at once.

Capability does not remove the need for an operating model. It increases the number of things the operating model has to make sense of.

The structure may be becoming less visible, not less important

That is where I land after reading Verna’s piece.

I do not think AI-native companies become structureless. I think some of the structure moves: out of titles and into decision rights, out of meetings and into shared memory, out of managers as information routers and into managers as context setters, out of individual expertise trapped in repeated conversations and into maintained agents and systems, and out of detailed process and into stronger quality bars, judgement and taste.

That is not obviously easier. In some ways it demands more from people because fewer decisions are pre-resolved by the hierarchy around them.

But it also explains why judgement keeps appearing at the centre of so many AI-at-work conversations.

When the mechanics are expensive, organisations build around the mechanics. When the mechanics become cheaper, the value moves towards deciding what should happen, making the context available, preserving what the organisation has learned and knowing what good looks like.

That is much closer to what I mean by Editorial Intelligence than simply using AI to produce more content.

The interesting AI-native organisation may not be the one with the most agents. It may be the one that works out what humans need to become responsible for once the old coordination machinery is no longer carrying as much of the load.

AI does not remove organisational infrastructure.

It changes what counts as infrastructure.