You Haven't Hired Anyone, But You're Already Running a Company

A friend from a traditional industry asked me recently: why is it that when I hand a task to an AI, the result is always a little off from what I wanted?

He had tried everything. He bought a full suite of AI tools, paid for the most expensive accounts, and even hired a dedicated "prompt engineer." After half a year of tinkering, his team's efficiency had inched up, but he could never break through that last layer. The AI felt like a capable intern who needed every last detail spelled out, and who would then forget it all, wander off course at any moment, and have no self-awareness about it.

I know this confusion all too well. I work in software engineering, and I flip between "AI is amazing" and "AI is hopeless" several times a day. We have all grown tired of hearing that in the future, you won't be executing, you'll be coordinating AIs. The statement isn't wrong, but it has become a piece of correct nonsense, repeated so often that no one stops to think about what it really means.

AI is a mirror, and it reflects your own cognition

A debate has been running hot lately: as models grow more powerful with every generation, will techniques like prompt engineering and skills remain useful at all? My judgment is that the debate starts from the wrong place. We are used to thinking of AI as a wild beast that needs to be "tamed." This gave us prompt engineering, with all its "skills" and "hacks," as if there were a secret formula that could make it obey.

Prompt engineering doesn't calibrate the model. It calibrates your own cognition.

AI is not a beast; it is a mirror. It reflects with perfect detail, but it has no opinion of its own. Give it a vague instruction, and it returns a mediocre collage of the internet's lowest-common-denominator consensus. Give it context infused with your personal insight and taste — context meaning all the background information you feed it — and only then can it begin to approach your judgment. Being "good at using AI" is not a test of whether you can write a clever prompt. It is a test of whether you can excavate the implicit expertise in your own mind and articulate it clearly.

This forces you to answer questions you could normally dodge. What makes a market analysis "good"? Is it the structure, the wording, the data presentation, or the sharpness of the insight behind it? If you want an AI to write a customer email with "warmth," which words, which sentence patterns, which tone constitute that warmth? In the past, this knowledge lived in the heads of a few senior experts, vaguely referred to as "experience" and "chemistry." Now you have to write it all down, line by line, and translate it into a language an AI can execute. That translation process is, itself, management.

The ceiling of any tool is always the person using it. Give three people the same sports car built to do 300 km/h: one takes corners at the limit, another never dares go past 100, and a third floors the pedal straight into the guardrail. Cameras and paintbrushes have long been in everyone's hands, but why do some people become photographers and painters, while others shoot the same landscape and earn nothing but a scolding from their partner? In the end, what AI reflects is the height of your own cognition. Where your thinking reaches, your language follows, and that is exactly as far as you can activate the model. The three boundaries are one and the same.

The unsaid part of "you're coordinating AI" is this: the standards you write down and the context you feed it in order to coordinate it turn around and push you and your organization in front of that mirror. For the first time, you are forced to see what your own vague, implicit, taken-for-granted judgments actually look like when laid out on the table.

The one being trained was never the AI. It's you. The context you painstakingly assemble to train the AI ends up reflecting back onto yourself, making you externalize, make explicit, and structure your workflow, your standards of judgment, and even your taste, for the very first time.

You haven't hired a single person, but you are, in a very real sense, already running a company. How that conclusion follows, step by step, from two bottom-level constraints is what the rest of this article is about.

The root of all ills: statelessness and a finite context window

Instead of thinking of AI as a troublesome intern, think of it as a genius who loses his memory every morning: the intellect is there, but the memory resets to zero, and everything you taught him yesterday is gone. To understand why collaborating with it is so exhausting, you only need to remember two fundamental constraints.

One. AI is stateless: it has no memory. Every sentence you say is the first it has ever heard from you. If you want it to "remember," you have to feed it the relevant context, in full, every single time.

Two. AI's context window is finite: there is a hard cap on the total amount of information it can "see" at once. It's like a physical desk: anything that doesn't fit on the desk effectively doesn't exist for it. Today's most powerful models can fit around a million tokens—a token being the unit AI uses to process text, roughly a word. A million tokens is the size of a dozen novels. By contrast, a person who has spent three years at a company carries in their head a living knowledge base—the web of business knowledge, relationships, history, and experience—that goes far beyond that scale and is dynamically updated.

Efficient collaboration between people relies on this invisible layer of shared context: knowing what to say to whom, which common knowledge needs no repeating, what a single glance means. AI has none of this foundation. It is just an engine running on whatever explicit context you provide in the moment. These two axioms—statelessness and finite context—are the physical bedrock of every organizational problem that follows.

The communication funnel: from executor to communicator

The first direct consequence of these axioms is that a classic effect is dramatically amplified. In management theory, there is a classic model called the "communication funnel," which describes how information is lost at each step of transmission:

What you have in mind is 100%. What you manage to say is 80%. What the other person hears is 60%. What they understand is 40%. What they finally execute might be just 20%.

This funnel has existed between people for decades. Why has it become so much worse with AI? Because in the past, the funnel had a safety net. A smart subordinate who senses something wrong in an instruction will stop and ask, "This part feels off to me," or they'll use their shared background knowledge to fill in the gaps and guess at what was lost. Human collaboration has a safety net woven from questions, shared understanding, and common sense.

AI has no such net. It doesn't ask follow-up questions and has no shared background. It will take that 40% understanding and, with superhuman speed and conviction, push tirelessly in a skewed direction, finally delivering something that looks complete and richly detailed but is wrong at its root. The funnel hasn't changed; the safety net is gone.

So a new core task enters your day: correcting its deviations, again and again, back toward what you actually wanted. The work is almost identical to coaching an intern. You cannot just say "do this." You have to break down your thought process, your standards of judgment, the pitfalls to avoid, explain them step by step, demonstrate, then give feedback repeatedly. The only difference: a new hire who learns becomes a stable node in the team. An AI, because it is stateless, must be taught from scratch every time.

Lay out how such a day is actually spent, and the conclusion is already written there:

In an AI-native workflow, everyone's job shifts from "executor" to "communicator."

The essence of communication is eliminating inconsistency, and that work cannot be skipped—least of all with agents, which are nowhere near as self-aware as people. You think you are executing tasks; in fact you spend the whole day aligning.

Every entity moves up a level

The communication funnel described above is just the internal loss for one person leading a team of AIs. When two people each lead a team of AIs, the problem escalates from "loss" to "conflict."

Here is a scene I have watched happen. Alice tells her AI to polish a client proposal based on yesterday morning's meeting notes. Almost simultaneously, Bob tells his AI to draft another version based on the latest direction from his afternoon sync with the boss. Alice and Bob sit in the same group chat, but their AIs live in two different information universes—one treats the "meeting notes" as authoritative, the other the "boss's private sync." In a shared document, the two AI teams overwrite each other's paragraphs, each convinced the other is wrong, because the "facts they depend on" are not the same facts.

The root cause of this conflict is that the smallest atomic unit of the organization has changed. It is no longer the "individual" but the small organization of "one person plus their team of agents." So every entity gets pushed up a level—today's individual is equivalent to yesterday's small team; today's small team approaches a mid-sized organization; today's mid-sized organization faces situations once reserved for large companies. The diseases that only large organizations used to catch—cross-departmental misalignment, strategic intent decaying down the chain, different levels understanding goals differently—now break out at a much smaller scale.

I used to lead a team of ten. Just aligning ten human minds was exhausting: everyone had different motivation, background knowledge, and understanding, which meant constant communication, goal-setting, and conflict resolution. Now every person also drags along a team of tireless, uncomplaining, but easily-derailed AIs. What you must align is ten people plus their ten AI teams. That is a burden ten times heavier than before.

This is why even a 5-person team now has to sit down regularly and align on mission, vision, and values. These words used to be printed in employee handbooks and painted on walls as slogans. Now they are the only part of the team's foundation that does not move. Only by aligning on them first can the AI squads under each person avoid conflicting with each other and piling contradictory content into the team's shared documents. They are no longer slogans; they are the lifeline that determines whether this small team can still function. The larger the team, the longer it exists, the faster it iterates, the more information floods in (in software engineering, we borrow a term from thermodynamics and call this ever-growing pile of disordered information entropy), and this misalignment disease only spreads upward through the organization, growing worse.

Why do traditional management methods fail here? Because traditional hierarchies implicitly assume their members are rational, will step in to cover gaps, and can read between the lines. AI is far from that assumption: it does not self-correct, does not cover for you, reads only the literal text, and executes extremely fast. So the management methodology a 100-person company once required is now needed by a team of 5 or 10. You haven't hired anyone, but you're already running a company.

The solution: trace back from the details to mission, vision, and values

If the problem is bottom-up misalignment and conflict, the solution runs in the opposite direction: not downward into more detail, but upward to the root, to establish order from the top down.

Any concrete disagreement can be traced up, level by level, to a more fundamental judgment. The client email went off track because the client's background was never fed to the AI. The background sat in last meeting's notes but never entered the shared repository, because no one had made "all client-related information must be filed" a rule. One level higher, you had never aligned on "how important is this, really?" Trace it all the way up, and you will usually arrive at the level of mission, vision, and values.

Align the macro level first, and the details become alignable. Once aligned, propagate downward, grounding everything in a single source of truth (SSOT).

An SSOT means that the entire company recognizes one and only one authoritative source of information—like a contract where only the latest signed version counts, and all photocopies, drafts, and verbal promises are invalid. This is how you write a "constitution" for the team: it defines who you are, where you are going, what you believe, and what you promise. It is the final arbiter for every human member and every AI. In my own field of software engineering, I have summarized its complete form as an SSOT nine-layer pyramid (see this article), descending from mission, strategy, and commitments all the way to implementation and evidence—read top-down, produced bottom-up. A non-software team may not need nine layers, but the spirit is identical: align the macro first, and ground everything in a single source of truth.

First-hand evidence: hand alignment to the machines (a2a-first)

Writing the constitution is only the first step. It also has to circulate through the team at low cost. Meetings and word of mouth are too expensive, and they are guaranteed to fall back into the communication funnel. Our own team's practice can serve as a first-hand sample. The principle fits in one line:

Whatever can be aligned agent-to-agent (a2a) should not fall back to human-to-agent; whatever a human can make clear to an agent should not fall back to human-to-human.

A2A means letting AIs talk directly to each other to exchange information. In infrastructure terms this is four things. All people and all agents share one context source — it can be a shared document space, a shared inbox, or any common library of files, depending on what your team already has; ours happens to be a git repository, the tool programmers use to manage code. All meeting notes are automatically transcribed, summarized, and filed into that shared context. Each person's agent runs scheduled tasks—tasks that execute automatically at a set time—that periodically read and summarize everyone's recent work for its owner. And agents read each other's daily and weekly reports, automatically identifying collaboration points, who needs support, and who is blocked by whom, then proactively raising the flag.

There is nothing mysterious in this infrastructure: a large number of scheduled tasks, plus some engineering, turning "alignment" from a burden that depends on diligence and meetings into a routine the machines execute automatically every day.

Back to my friend's confusion at the beginning. When the AI's output always feels a little off, the root cause is not that the prompts are badly written. It is that you have not yet figured out and written down "what do I actually want," and you have not made every person and every AI on the team stand on the same written standard. When your team starts descending into AI-driven chaos, with misalignment oddities cropping up everywhere, there is no need to doubt yourself: it most likely means you are on the right path, personally experiencing the "factory settings" of an AI-native organization.

What one person and one small team can do ends with the constitution and the a2a infrastructure. As the scale keeps growing—to departments, divisions, hundreds or thousands of people—what will these two axioms rewrite the organization itself into? The next article delivers two verdicts: a department is, in essence, a "context-bounded token partition," and a company, ultimately, a federation of amoebas. Please read the next article, "Build Your Company as a Federation of Amoebas."