You Never Needed a Fad to Invest in Your People
Companies wait for management fads before helping their own people. Then blame AI when the real problem is neglect.
The One-Person AI Company Is a Greed Fantasy
11 min read
A Swarm Does Not Give You a Second Mind
15.09.2026, By Stephan Schwab
An AI agent swarm can multiply one person's output. It cannot supply a second independent mind, different experience, the incentive to expose a bad assumption, or a human willing to say no and carry responsibility. Every agent inherits its operator's framing, context, permissions, and blind spots. LLMs give skilled people extraordinary leverage, but ten fluent variations of the same mistake are still one mistake.
The dashboard looks magnificent.
One agent plans. Another researches. A third writes. A fourth reviews. A fifth watches the others and produces a neat summary for the human sovereign at the top. The boxes have different names, different icons, and perhaps even different synthetic personalities. From a distance it resembles an organization.
It is not an organization.
It is one person’s intent, replicated through a collection of probabilistic systems and ordinary software. The swarm may explore more options than that person could explore alone. It may work in parallel. It may catch inconsistencies. It may even produce a useful internal argument by assigning different roles to different agents.
But all of it still begins inside one frame. One person chose the objective, selected the available evidence, granted the permissions, defined success, and decided which disagreement counted. The apparent committee has one owner, one source of authority, and usually one set of blind spots.
That is leverage. Sometimes astonishing leverage.
It is not omniscience.
LLMs Do Not Want Anything
An LLM can pursue a bounded goal. It cannot decide what deserves to matter and bear the human consequences of that choice.
The word agent smuggles a great deal of humanity into a software component.
An agent appears to have a goal. It can choose among tools, perform several steps, inspect a result, try again, and ask for help when the surrounding software tells it to. This is useful operational agency: enough delegated freedom to complete a bounded task without a human approving every keystroke.
Human agency is something else.
A human can question the goal itself. A developer can refuse to build a surveillance feature. A lawyer can say that a technically defensible tactic is still wrong. A financial analyst can notice that the requested forecast exists mainly to flatter an executive’s preferred decision. A support worker can recognize that the policy is hurting a customer in a way the policy never anticipated.
Those acts are not failures to follow instructions. They are judgment.
The LLM has no career to risk, no professional reputation to defend, no customer whose anger it must face tomorrow, and no private conviction that the requested outcome should not exist. It does not want the company to succeed. It does not want the company to fail. It produces the next action inside the structure humans supplied.
Giving it more tools does not change that distinction. Giving twenty instances different job titles does not change it either.
One Operator, Many Mirrors
A swarm can diversify answers while preserving the same underlying ignorance.
The fashionable answer to model fallibility is to add more models.
Let one agent propose. Let another criticize. Let a third compare. Let a fourth vote. The pattern can improve results. Independent passes catch mistakes, adversarial prompts expose weak reasoning, and specialized context can make each agent better at its assigned task.
The mistake is treating this technique as the equivalent of independent human judgment.
The agents may share the same training biases. They may receive the same incomplete company documents. They may all lack the tacit knowledge held by the dismissed operations specialist. They may all accept the same misleading success metric because the operator wrote it into every task. They may confidently debate three solutions to the wrong problem.
Even the critic is performing a role defined by the person who wants the work done.
A real colleague brings an inconvenient history the operator did not select. She remembers the customer who nearly left when the company tried this three years ago. He knows why the apparently redundant reconciliation step exists. She notices that the new policy will shift unpaid work onto a team that was never invited into the discussion. He has an incentive to challenge a decision because he will be on call when it fails.
That independence can be frustrating. It can slow the meeting. It can ruin a beautifully simple plan.
That is often its value.
The Part Greed Finds Irresistible
Efficiency asks what work no longer needs to be done. Greed asks how many people can be removed before anyone can prove they were necessary.
There is nothing immoral about using AI to do more with less effort. Tools have always removed drudgery, compressed tasks, and allowed smaller groups to compete with larger ones. A skilled person should use the leverage available.
Greed enters when the desired conclusion arrives before the evidence.
The operator does not ask which work the agents perform reliably, which decisions still need independent review, or which responsibilities become more important as output accelerates. The operator asks how quickly the payroll can disappear.
Every objection then becomes self-interested resistance. Developers are protecting jobs. Lawyers are blocking innovation. Editors are nostalgic. Analysts are defending their spreadsheets. Operations people are afraid of change. The agent never complains, so the agent is declared more aligned.
Of course it is aligned. It has no interests of its own.
That is precisely why the comparison is dishonest. The human colleague who warns about a security boundary, an unsupported claim, a fragile supplier, or an angry customer is not producing friction because humans are obsolete. The colleague is exposing a cost the operator hoped not to pay.
Silence is cheaper than judgment right up to the moment judgment would have saved the company.
Output Is Not Capability
The ability to generate an artifact is not the ability to know whether the artifact should exist, is true, or will survive contact with reality.
LLMs are superb artifact machines.
They can produce a plausible contract, a functioning feature, a polished campaign, a market analysis, a policy, a research summary, a hiring plan, and a board presentation before lunch. The quality range is enormous, but the speed is real.
Knowledge work was never only the production of artifacts.
The hard part is knowing which contract term will become poisonous during a dispute. It is recognizing that the functioning feature encodes the wrong business rule. It is noticing that the market analysis quietly compares unlike categories. It is asking whether the campaign promise can be fulfilled by operations. It is understanding which sentence in the policy will make competent people route around it. It is deciding that the board should hear an unpleasant truth rather than a polished forecast.
Those decisions draw on domain knowledge, lived consequences, conflicting interests, ethics, taste, and the courage to refuse.
The one-person operator may possess some of that judgment. Perhaps a great deal of it. AI can extend the operator’s reach and free time for the decisions that matter most.
The delusion begins when success in one domain gets mistaken for universal competence. A founder who can judge product direction assumes the same instinct now covers security, employment law, accounting, user research, software architecture, and operations because an agent can produce fluent material in each field.
Fluency hides the boundary between what the operator knows and what the operator can no longer recognize as wrong.
The First Results Will Be Excellent
The fantasy survives because the early evidence is real.
The one-person AI company will not necessarily fail on day one. That would make the lesson easy.
It may thrive.
Backlogs disappear. Prototypes arrive overnight. Customer replies become instant. Reports that used to take a week appear in an hour. The operator feels vindicated, investors applaud the margins, and every former colleague starts to look like proof of historical waste.
This is where sober criticism often loses. It denies the visible productivity gain instead of explaining its limits. The gain is real. A capable person with modern models can perform work that recently required much more time and coordination.
But early throughput measures the easiest part to see: production.
It does not yet measure the decisions nobody challenged, the exceptions nobody recognized, the security model nobody understood, the relationships nobody maintained, the institutional memory nobody wrote down, or the risks that take months to mature.
Agentic coding becomes useful through boundaries, evidence, and judgment. The same is true outside software development. Remove the people who provide those boundaries and the agents do not become more autonomous. The remaining operator becomes more exposed.
The Awakening Has a Delay
Generated work can pass every visible check while the organization quietly loses its ability to notice the checks that are missing.
The bill usually arrives as an exception.
A customer has a case the workflow never modeled. A regulator interprets an obligation differently. A model update changes behavior. A supplier removes an API. An attacker finds the permission boundary the agents kept reproducing. A financial assumption survives through twelve generated reports because all twelve inherited the same source. A new employee asks why the system behaves this way and discovers that the answer exists only across hundreds of synthetic summaries.
Then the operator turns to the swarm for help.
The swarm reads the artifacts. It reconstructs likely intent. It proposes a migration, an explanation, a revised policy, and a reassuring list of next steps. The output is once again impressive.
What it cannot retrieve is the human context that was never captured. Why did the experienced developer object? Which customer promise overruled the clean design? Which exception was politically impossible to document? Which warning did the operator dismiss as resistance? Which number looked wrong to the analyst because she had spent ten years watching the business breathe?
That knowledge left with the people.
The operator did not merely reduce headcount. The operator destroyed the independent sensing system of the organization and replaced it with a highly efficient generator trained on whatever remained.
You Cannot Rehire Memory on Demand
Once the people are gone, recovery is not a matter of buying fresh tokens.
The standard fantasy includes an escape hatch: if the agents ever struggle, hire a specialist for a few hours.
Sometimes that works. Specialists can diagnose bounded problems and AI can help them understand unfamiliar material faster than before.
But a company is not a set of interchangeable answers waiting in a marketplace.
The dismissed developer knew which tests lied. The former account manager knew which customer accepted an awkward compromise and which one merely stopped arguing. The operations worker knew that a duplicated step was protecting against a supplier’s erratic data. The editor knew when the founder’s favorite phrase made the company sound untrustworthy. The finance person knew which optimistic assumption was a ritual and which one would trigger a cash crisis.
You can hire competence again. You cannot instantly repurchase the exact history, trust, and pattern recognition you chose to discard.
Worse, the returning specialist now faces a mountain of generated material. More code, more documents, more decisions, more automated workflows, and less shared ownership. The apparent productivity gain has become an excavation project.
We have dreamed of removing developers from software development for decades. The one-person AI company merely extends the dream to every occupation whose real work becomes visible only when it is absent.
The Winner Will Not Be the Loner
AI favors small groups with dense judgment, not isolated rulers surrounded by obedient simulations.
AI will change the size and shape of organizations. Some tasks will disappear. Some roles will combine. Small companies will challenge incumbents with resources that would have been absurdly insufficient a few years earlier. Pretending otherwise is nostalgia.
But the strongest organization is unlikely to be one human performing sovereignty over a synthetic court.
It will be a compact group of capable people who use agents aggressively and retain genuine differences in knowledge, incentives, experience, and judgment. Fewer handoffs. Less ceremony. More individual range. Clear responsibility. Enough independent minds to catch a shared mistake before the market does.
That is not a defense of bloated departments or jobs preserved as museum pieces. If a task no longer creates value, stop doing it. If an agent performs it better, use the agent. If a team can become smaller without losing necessary capability, let it become smaller.
Just do not confuse the removal of labor with the acquisition of wisdom.
The greedy operator sees the payroll savings and imagines the whole surplus flowing upward. The serious leader sees the leverage and asks a harder question: which human judgments become more valuable now that producing work is cheap?
That question leads to better companies. It also leads to less theatrical ones.
Human Agency Is Not Overhead
The purpose of AI leverage is to expand what responsible humans can do, not to create a company in which only one human is allowed to matter.
Human agency is inconvenient because other people can choose differently.
They can challenge the founder. They can protect a customer the metric forgot. They can refuse a shortcut. They can leave. They can demand a share of the value they helped create. From the narrow perspective of control, an obedient agent swarm looks superior.
From the perspective of building something durable, that obedience is the danger.
A company needs more than generated options and fast execution. It needs people who can decide what should happen, understand why, notice when the premise has changed, and accept responsibility for the consequences. LLMs can inform those decisions. Agents can carry out bounded parts of them. Neither makes responsibility disappear.
The one-person AI company is therefore not mainly a technology prediction. It is a moral and organizational wish: maximum capability without dependence on other humans, maximum output without shared power, maximum gain without the irritation of competing judgment.
It may look brilliant for a while.
The awakening begins when reality finally asks a question the sovereign never knew to ask, and every agent in the kingdom gives a beautiful answer to something else.
Talk It Through
Tell me what is happening. I listen, ask a few practical questions, and reflect back what I see: where the risk may sit, what may be blocking delivery, and what looks worth checking next. No pitch, no obligation. Confidential and direct.
Talk it through. Practical reflection, no pitch.
Start a Conversation
🎭 This Week's Episodes
Telenovelas show what we can't say in client meetings. The drama is heightened, but the patterns are real.
Three Ways to Understand a Country
Stefan drives Sophie to the Pacific for two days. The beach is big enough to make her feel appropriately small. Back at the finca the rain keeps th...
The Resignations
Nathan Cole resigns on a Monday morning with two sentences and no argument left in his body. By Tuesday his desk is empty and the company's institu...
Technical Consultancy
Embedded Delivery Partner
Embedded into your team as an active contributor, reducing delivery friction and helping important work move cleanly.
Technical Advisor
Peer-style technical assessments before far-reaching decisions; reduce architectural and product risk early.
Product & Delivery
Ship working software to real users earlier. Measure impact and adapt based on evidence rather than assumption.
Custom Software Development
High quality, maintainable software. Short-term augmentation that leaves lasting capability in your own team.
Recent Articles
- When Everything Is Finally Under Control (28.09.2026)
- AI Makes Technical Execution Cheap. Judgment Stays Scarce. (24.09.2026)
- The LLM Will Not Come Back Tomorrow (19.09.2026)
- The One-Person AI Company Is a Greed Fantasy (15.09.2026)
- Test Production Without Leaving a Mess (10.09.2026)