The AI ready company

For most of my career, building digital products meant organizing scarcity.

There were never enough designers. Never enough engineers. Never enough time. Never enough information in the right hands at the right moment.

So we designed organizations around this scarcity.

We created teams.

Then teams needed coordination.

Coordination needed managers.

Managers needed reporting.

Reporting needed meetings.

Meetings needed preparation.

Preparation needed documents.

Documents needed alignment.

And alignment needed another meeting.

I am exaggerating.

A little.

There were good reasons for all of this.

I spent years working with Agile methods and I still believe one of their great contributions was to shorten the distance between thinking, making and learning. Put different expertise around the same problem. Work in smaller increments. Show things early. Observe reality. Change direction before changing direction becomes terribly expensive.

That was already an attempt to reduce the friction of the organization.

But Agile was designed for humans.

And humans are wonderful, complicated, expensive creatures.

We forget things. We misunderstand each other. We have limited attention. We need sleep. We interpret the same sentence differently. We join projects halfway through. We leave with knowledge in our heads. We have egos, intuitions, politics, friendships, fears, ambitions, blind spots and, thankfully, occasionally very good ideas.

The organization became a technology for making all of that work together.

And then something changed.

Not everything.

But something fundamental.

Implementation started becoming abundant.

I noticed it first at a very small scale

I build products.

For years, that sentence implied a collection of people.

More recently, it has increasingly meant me, a computer and a strange collection of artificial collaborators that never sleep, type extraordinarily fast, sometimes reason brilliantly and occasionally do something so spectacularly stupid that I remember immediately why I am still there.

With Magimica, I wanted to see how far this could go.

Not a prototype.

A product.

Database. Backend. Frontend. Scheduling logic. AI assistance. Tests. Deployment. Production.

Five days.

This should feel absurd to anyone who remembers how software was produced not very long ago.

It feels absurd to me.

But the most interesting thing wasn’t the five days.

It was what became scarce instead.

Code wasn’t really the problem anymore.

Deciding what should exist was.

Understanding what mattered was.

Knowing when something was good enough was.

Seeing that an elegant technical solution was solving the wrong problem was.

Choosing when to trust probabilistic intelligence and when to force deterministic behavior was.

Knowing when to stop.

So I eventually wrote a sentence at the top of my portfolio:

Implementation is abundant. Coherent decisions are scarce.

I initially meant product development.

I am increasingly wondering whether it describes the future company.

Of course, we immediately tried to rebuild the company inside the machine

This part makes me smile.

Give humans powerful new technology and one of our first instincts is apparently to reproduce ourselves.

One AI agent wasn’t enough.

So we made teams of agents.

One writes.

Another reviews.

Another tests.

Another critiques the test.

An orchestrator tells them what to do.

A supervisor watches the orchestrator.

We add memory.

Roles.

Handoffs.

Escalation.

Evaluation.

Suddenly we have invented middle management at token speed.

I tried versions of this myself.

At one point, I introduced heavier QA agent orchestration into my own product workflow. The theory was beautiful. Specialized agents would inspect the work, challenge one another and improve the quality before anything reached me.

The result?

Roughly twice the time.

Roughly five times the tokens.

And no consistent improvement in the quality I actually cared about.

So I removed most of it.

That experiment left me with an uncomfortable question.

What if some agentic orchestration is simply bureaucracy with an API?

Not all of it.

Certainly not.

Specialization matters. Independent evaluation matters. Separation of concerns matters. Parallelization can be extraordinarily powerful. Some systems genuinely require multiple authorities, perspectives and controls.

But complexity needs to earn its existence.

That is true of organizations.

Why would it suddenly stop being true because the workers are artificial?

Meanwhile, large organizations have another problem

They know something important is happening.

Nobody wants to be the executive who wakes up one morning and discovers that a competitor reorganized around AI three years earlier.

So organizations experiment.

POCs appear everywhere.

Agents appear everywhere.

Copilots appear everywhere.

Innovation teams demonstrate increasingly impressive things.

And yet the distance between a beautiful demonstration and an organization that actually operates differently can remain enormous.

Because the technology isn’t entering empty space.

It enters existing authority.

Existing processes.

Existing budgets.

Existing systems.

Existing regulation.

Existing jobs.

Existing accountability.

Existing fear.

A POC can demonstrate that an AI is capable of doing something in fourteen minutes that previously required three days.

Wonderful.

Who is responsible when it does it wrong?

Ah.

Now we are talking about the organization.

Someone still has to be the fuse

I use the French word in my head: fusible.

A fuse.

The component that sits in the circuit knowing that when something goes sufficiently wrong, the current stops there.

Organizations contain human fuses everywhere.

A designer approves.

An engineer merges.

A product owner prioritizes.

A manager signs.

A compliance officer clears.

An executive commits.

We often describe these people as decision makers, but decision making is only half of their function.

They are also locations of accountability.

AI makes this increasingly visible because execution can move much faster than accountability.

An agent may be able to analyze the problem, propose the solution, write the implementation, run the tests, inspect the result and recommend deployment.

But somewhere in the chain, today, a human being still says:

Yes.

Ship it.

Send it.

Approve it.

Move the money.

Change the policy.

Reject the candidate.

Treat the patient.

Take the risk.

And that yes is very different from generating another token.

It commits someone.

This may be one of the reasons AI adoption sometimes feels strangely slow from inside large organizations even while the underlying technology moves absurdly fast.

The bottleneck isn’t necessarily intelligence.

It isn’t necessarily implementation.

It may increasingly be responsibility.

And if that is true, simply inserting AI into every existing box of an organization chart misses the deeper opportunity.

Perhaps we need to redraw the boxes.

So I tried a thought experiment

A deliberately unpleasant one.

I asked myself:

What would a company look like if we designed it today around AI rather than adding AI to an organization designed yesterday?

And because thought experiments become more useful when pushed until they become slightly ridiculous, I removed things.

Departments?

Remove them.

Layers of management?

Remove them.

Permanent teams?

Why necessarily?

Status meetings?

Please.

Reporting chains?

Let the machine maintain state.

Information distribution?

The machine already knows who needs what.

Project coordination?

Machine.

Work decomposition?

Machine.

Progress monitoring?

Machine.

Cross functional synchronization?

Machine.

Performance synthesis?

Increasingly machine.

Keep removing.

Eventually I arrived at an organizational design that sounds so politically incorrect that I rather like its name.

The dictator CEO company

One CEO.

Not a management pyramid underneath.

An AI operating system.

And a fluid network of extremely capable human specialists and artificial agents executing around it.

The CEO defines intent.

The system translates intent into objectives, constraints, work, dependencies and decisions.

It distributes only the context each capability requires.

It observes execution.

It maintains organizational memory.

It evaluates progress.

It surfaces contradictions.

It escalates decisions.

It challenges assumptions.

It reports reality upward.

And specialists execute.

Some of them are employees.

Many might not be.

In earlier conversations about this model, I called them mercenaries.

The word is intentionally uncomfortable too.

Companies love saying they are families.

They aren’t.

Families generally don’t dismiss you because quarterly revenue missed expectations.

A company is a system of economic cooperation.

That doesn’t mean it needs to be cruel. Quite the opposite. It can respect people enormously while being honest about the nature of the relationship.

The mercenary in this model is not disposable cheap labor.

She is an elite independent specialist.

Highly competent.

Highly paid.

AI augmented.

Reputation driven.

Called because her judgment matters.

She doesn’t spend half her week navigating the organization around the work.

The organization comes to her with the problem, the exact context she requires and a clear definition of authority.

She does the work.

Then perhaps she leaves.

There is something simultaneously cold and strangely respectful about that model.

Less corporate theatre.

More explicit exchange.

But dictatorship has a famous engineering problem

The dictator becomes blind.

If every instruction flows downward and nobody can challenge the person at the top, the system eventually optimizes for telling the leader what the leader wants to hear.

Companies already do this without AI.

Artificial intelligence could make it catastrophically efficient.

So the interesting AI operating system cannot merely transmit authority downward.

It needs to protect truth upward.

A specialist must be able to challenge an instruction.

An agent must be able to detect a contradiction.

Evidence must be able to travel against the hierarchy.

The CEO can retain decision authority while losing the privilege of informational comfort.

This distinction matters.

Authority can be centralized while truth remains adversarial.

In fact, AI might make this easier.

The person challenging a strategic assumption doesn’t necessarily need access to the entire strategy.

The system can expose exactly the context required to test the assumption without exposing everything else.

Opacity can protect strategy.

Compartmentalization can protect information.

And carefully designed challenge channels can still allow inconvenient truths to reach the top.

The chain of command survives.

The game of telephone doesn’t have to.

This is where the thought experiment becomes less funny

Because if this architecture actually worked, what exactly would management be for?

That question deserves more respect than the usual prediction that “AI will replace managers.”

Good managers do many things that software cannot simply be assumed to reproduce.

They interpret ambiguity.

They develop people.

They absorb conflict.

They notice what isn’t being said.

They build trust.

They create meaning.

They carry responsibility.

Sometimes they protect teams from the stupidity of the organization itself.

But management also exists because information, coordination and attention are expensive.

And those costs are changing.

So perhaps the interesting question isn’t:

Will AI replace managers?

It is:

Which reasons for management survive when coordination becomes computational?

The same question can be asked of teams.

Departments.

Processes.

Meetings.

Reporting.

Even employment.

Some will survive because humans need them.

Some because regulators require them.

Some because trust requires continuity.

Some because tacit knowledge is real.

Some because concentration of power is dangerous.

Some because humans actually enjoy building things with other humans.

And some, surely, will survive for twenty years simply because organizations are extremely good at preserving things whose original purpose everybody has forgotten.

The company may become smaller without becoming less capable

This is the possibility that interests me most.

For most of industrial history, increasing organizational capability generally meant accumulating people and capital.

More customers required more operations.

More products required more teams.

More teams required more managers.

More managers required more organizational machinery.

Software already weakened that relationship.

AI may weaken it dramatically further.

A tiny permanent organization could command enormous temporary capability.

A handful of accountable humans.

A computational coordination layer.

Thousands of agents.

Specialized services.

Independent experts appearing when needed.

Capital.

Infrastructure.

Distribution.

Then disappearing again.

The boundary of the company becomes blurry.

Maybe the future organization isn’t primarily a collection of employees.

Maybe it is a persistent intention surrounded by dynamically assembled capability.

I find that idea beautiful.

And slightly terrifying.

There is one thing I would not delegate

Accountability.

At least not yet.

My own way of building with AI has gradually converged around something almost embarrassingly simple:

Define what matters.

Design the system.

Verify the result.

Then let reality change the next decision.

AI can do enormous amounts between those points.

But I remain responsible for what ships.

That sentence matters to me.

Not because I believe humans are magically superior.

Often the AI is better than I am at a specific task.

Sometimes vastly better.

But responsibility cannot merely be another step in the orchestration graph.

Someone has to own the consequence.

Perhaps the AI ready company therefore doesn’t eliminate humans from the chain of command.

It makes the remaining humans more consequential.

Fewer people may control dramatically more execution.

Which means their judgment matters more.

Their ethics matter more.

Their ability to admit error matters more.

Their exposure to contradictory evidence matters more.

Their capacity to resist their own certainty matters much more.

The dictator CEO sounds efficient until the dictator is wrong.

Then the entire architecture becomes a machine for scaling one person’s mistake.

That may be the most important objection to my own thought experiment.

And therefore the most interesting design problem inside it.

So no, I am not proposing that every company fire its managers on Monday

Tuesday would be more agile.

I am asking something else.

We spent decades improving how humans coordinate work.

Then intelligence capable of participating in that coordination arrived.

We are currently placing it inside our existing organizations, often one assistant, one workflow, one POC and one agent at a time.

That is probably necessary.

It may also be temporary.

Eventually someone will stop asking:

Where can AI fit into this organization?

And ask:

What organization would we build if AI had always existed?

I don’t know what that organization looks like.

I don’t believe the dictator CEO is the answer.

It is a boundary condition.

Push hierarchy compression as far as imagination allows and see what breaks.

Accountability breaks.

Checks and balances break.

Human development may break.

Trust may break.

Culture may break.

Institutional memory may break.

Perhaps innovation itself breaks when too much intent originates from one mind.

Good.

Now we know what needs designing.

That’s why I find the extreme useful.

The interesting company probably exists somewhere before it.

Fewer layers.

More autonomous expertise.

Far more machine coordination.

Far less reporting as a human activity.

Information delivered according to need rather than copied to everyone.

Explicit authority.

Explicit escalation.

Persistent organizational memory.

Independent evaluation.

Protected channels for disagreement.

Humans concentrated where judgment and responsibility genuinely matter.

AI everywhere else it earns its place.

Not because AI can do everything.

Because we should finally be able to ask every piece of organizational complexity the same question I learned to ask of agentic systems:

What are you adding that justifies your existence?

Perhaps that is what being AI ready ultimately means.

Not having more agents.

Not having an AI strategy deck.

Not running more POCs.

Not putting a copilot inside every application.

Being AI ready may mean becoming willing to redesign the organization around a new economic fact:

Execution is becoming abundant.

And then confronting the much harder question:

What remains scarce?

Judgment.

Trust.

Responsibility.

Taste.

Courage.

Meaning.

The ability to decide what deserves to exist in the first place.

For now, at least.

And if those are the scarce resources of the next company, the most important transformation ahead may not be technological at all.

It may be organizational.

We built companies to coordinate human execution.

Now we may have to learn how to build them around human consequence.

Unless AI comes with a warranty like a Tesla does in case of copilot crash?