The organizational dead end
2026-11-12
There's a misconception around what it looks like to have an AI-native tech company, and how those organizations will come into being.
People seem to really believe it'll go something like this:
- Start with a large or mid sized company as it exists today
- AI is introduced - first ICs are replaced one by one, then line managers
- Eventually you end up with perhaps the few key decision makers and no-one else
This is completely and utterly wrong and ignores the reality of how the organization came to be in the first place and whether this ablative process can actually produce a better organization given the people in the company right now.
I'll describe what happens in practice, why it's a dead end, and how these orgs get replaced.
Hitting the brick wall
Most large software-producing organizations have worked the same way for a very long time - on yearly, quarterly, and finally biweekly cycles, various levels of decision makers plan how to execute requirements after a sometimes long and drawn out process.
Product managers, sales, executives, directors will opine about strategy, and that'll filter down to the lowest levels.
Despite the comedic level of inefficiency of such an organization, it usually still is true that the bottleneck of execution is still the engineer. It's much easier to state an ambitious idea than it is to execute it. The jobs of people, product, and project managers up and down the org is to break those strategies down into executable chunks which engineers can execute.
The stated or purported reason hierarchical organizations exist is to break up a domain into chunks that an individual can hold in their heads - the average human's capacity for this has been static over time. The only way to have a team effectively wear more hats is to select humans with greater capacities to hold ideas in their heads, but even then greater focus is likely to produce better results.
The stated goals are usually to execute faster, increase revenue, please customers.
Wait, the stated reason, the stated goals? Well I'm a big fan of the purpose of a system is what it does - if big companies execute slower and do not actually even try to select for polymaths, and do not even try to have cross-functional Renaissance men who can make fast, correct decisions, then that's not their purpose.
This has been laid bare as we hit the brick wall of productivity exposed by AI.
Now, execution is not the bottleneck. Just like any program you profile the execution of a project and find most of the cores laying idle - they can write programs faster than the machinery of communication and coordination can come up with them.
Now, a project spanning a dozen teams and a hundred engineers spends almost all of its time waiting on decisions from stakeholders. AI has 100xed productivity in producing software without 100xing decision making speed.
This is an obvious case of Amdahl's law - previously, execution and planning took up an equal amount of time. Because AI won't, on its own, reorganize your company and make planning faster, the most you can speed anything up is 2x.
I see this on a daily basis - teams and contributors who adopt AI the fastest and are the most productive trivially exhaust their queues of work and spend their time blocked waiting for code reviews or for other teams to okay their ideas.
This is not the end state of a large organization adopting AI - staying pancaked against the brick wall is not a permanent state of affairs.
I don't know what the real purpose of the system is, but if it doesn't get things done and still doesn't get things done, then getting things done is not its purpose. The principal-agent problem has many failure modes.
Getting around the wall
For a short time it will be possible for blocked engineers to focus on valuable but neglected work like testing, performance, tech debt, and so on. But there's a point at which that's no longer useful and we'll be forced to examine what's going wrong.
There has been a lot of buzz around this article on accelerating AI research. The key difference between how old world managers and execs view AI and the reality is that they believe AI is like having a team of hundreds at your disposal. Instead, it's more like being in a time warp.
Their beliefs are very far away from reality and this lays bare the cost that they don't understand: a team of 100 people wouldn't get anything done due to communication overheads and the inevitable politicking that comes with any group of more than 1 person.
This is Brooks' Law:
Adding manpower to a late software project makes it later.
That's a simplification, but it does illustrate the idea. Collaboration is a COST that you pay, not a benefit. The fewer people that need to be involved in anything, the better.
If decision making for a group of a dozen can be collapsed into one or two people who both make decisions and execute, then we'll see real order of magnitude speedups in development.
But can large, inefficient, slow organizations become small high-perfoming organizations? I don't think so, for one simple reason: management is the problem and management will never decide to eliminate itself.
Compare three managers:
- One manager is a generationally talented product manager but is non technical
- One is a wizard at running teams of software engineers
- One manager is a polymath - he is technical, product-minded, and capable of both running a team and doing some work but has less depth of experience than either of the first two managers.
The last manager will be able to produce software orders of magnitude (10x-100x) faster than the first two put together. The elimination of the collaborative or communication overhead will allow us to bypass Amdahl's law.
But this kind of manager is not selected for at any large company - you have many levels of managers who aren't fit for the new world at every level of the organization.
If we descend to the fifth circle of hell and consider the scourge of non technical non expert managers who do not have any expertise in anything, then such a person would not benefit from mass layoffs and running a team of agents. Having frontier AI is less like you having a team that has greater expertise than you do, and more like you being in a time warp, as I said. So having expertise in managing teams doesn't help.
If the current rulers are in charge of implementing the new system, it probably will not acknowledge that the value of coordinating teams of people and politically maneuvering between groups to get your particular proposal accepted, building consensus, and so on, has dropped not just to zero, but is significantly negative.
If you crash your car into a wall and it's totaled, there is no getting around it. You crashed and you need to get a new car.
Getting out of the car
It seems unlikely that these organizations will just fail. After all, incumbent, profitable entities like Big Tech companies are likely to adapt and remain profitable. But why is that, why wouldn't missing out on a 100x productivity boost mean that you get outcompeted?
The truth is that most of what goes on in large organizations that have achieved product market fit decades ago already doesn't generate any value, and startups having a 100x advantage is not enough to overcome market failures.
If Discord moves a text box around or Reddit shifts to a new UI framework, that probably generates no value. The battery bar on Android changing shape probably doesn't generate any value either. The design language of macOS changing likely destroyed value.
If you 100x the speed at which projects in a company move, you're changing the temperature of an object by increasing the speed of random motion - that doesn't result in macro-scale properties of the company like the position of the object changing. It's stationary and it's going to stay that way.
It might be unpleasant for most engineers to hear, but you know it to be true: your job was already pointless, generated no value, was akin to ditch digging or a welfare system.
Leaving behind cars and getting into the rocket
The only way for the existing owners of companies to build new companies that follow the new model is to invest in them. The old superstructures of project management and people management are entrenched and cannot be removed even as the difference in execution speed between new and old organizations becomes farcical.
Startups have always had this advantage - being smaller and moving faster is enough of an advantage on its own. Large companies cannot pivot or compete as well as you can.
Evolution doesn't happen through individual organisms changing into other organisms, it happens through competition.