The pilot never got replaced

At 35,000 feet, the pilots are still there.

Commercial aviation is among the most automated working environments most of us will ever trust with our lives. Autopilot can hold altitude, follow a route, manage speed and assist with landing. Yet on commercial aircraft built for two pilots, two qualified pilots remain required in the cockpit. Automation did not make expertise obsolete. Instead, it changed where that expertise matters.

The machine handles sustained execution, while the pilot holds the intent, understands the environment, monitors the system and intervenes when reality departs from the plan.

That is the pattern I believe is coming to knowledge work.

The current conversation about AI is still largely about making an individual faster or completely replacing them. Employees are given a copilot. Companies are sold access by the seat, charged by the token and encouraged to add an agent to a workflow. Without a doubt, these tools can produce real advantages, but they leave the actual structure of work almost untouched. The AI has capabilities, but no durable place in the organization since it has no defined responsibility, no clear human owner and no unit of output against which its cost can be judged.

Tenor is built on a different conviction. The same way companies expect an upside from paying salaries to employees, AI should enter the organization as accountable workers with defined responsibilities, sitting under the employees who direct them. Employees become managers of AI workers that can perform long-horizon tasks and never repeat the same mistake. Their existing manager remains their people manager. These are different jobs.

The people manager sets direction, allocates resources and remains accountable for the organization they lead, while the AI worker manager directs digital execution. They define the task, provide context and permissions, set the standard, review exceptions and remain accountable for the result.

One leads people, direction and organizational outcomes. The other directs work performed by machines. The first remains necessary. The second should become universal.

The chart was an invention

The modern organization can feel permanent because nearly every company inherited the same basic shape. What many of us forget is that the org chart was invented to solve a technological problem.

In the 1850s, the New York and Erie Railroad stretched for hundreds of miles and employed thousands of people. Informal methods could clearly no longer coordinate trains, stations, maintenance and telegraph messages at that scale. Daniel McCallum, the railroad's general superintendent, developed a system of delegated authority and reporting, then represented it in what is widely recognized as the first modern organizational chart. The chart was an operating technology for routing information and responsibility through a network too large for one person to see.

Frederick Winslow Taylor later applied the same logic to work inside the hierarchy. Scientific management divided labor into measurable tasks, separated planning from execution and searched for repeatable methods. McCallum gave the company reporting lines and Taylor gave it decomposed work. Managers now coordinated information and judgment while employees executed the tasks assigned to them.

Technology kept altering that settlement. The spreadsheet moved accounting away from maintaining rows of figures by hand and toward interpreting them, part of a broader shift from clerical production toward professional expertise.

But history offers no comforting law that automation always elevates the worker. Research by Daron Acemoglu and Pascual Restrepo found that industrial robot adoption reduced employment and wages in exposed American labor markets. Technology creates leverage, but institutions decide where that leverage goes. It can be used to remove the worker, to control the worker or to elevate the worker.

I believe AI gives us a chance to choose the third path deliberately.

Two views of what comes next

In their March 2026 essay, “From Hierarchy to Intelligence,” Jack Dorsey and Roelof Botha offer an interesting perspective on how AI could reshape the company itself.

Their argument is that hierarchy is an information-routing system built around the limited number of people any human can manage. More people require more layers, more layers slow information, and AI creates a way out. A company-wide world model can maintain context, coordinate work and distribute intelligence without requiring human managers to relay information up and down the chain.

At Block, they envision intelligence living in the system and people operating at the edge, close to customers, craft and reality. Individual contributors build capabilities and directly responsible individuals own outcomes. Player-coaches develop people while continuing to build. Much of middle management disappears because AI performs the coordination work that justified it.

There is a lot in this diagnosis that I share. Copilots alone do not change a company. Hierarchy slows information. People closest to the work should have more authority. AI should reshape the operating system of the firm, not decorate its existing software.

I differ on the route.

Organizations resist structural change with astonishing force. In a McKinsey survey of organizational redesign, fewer than 30 percent of respondents rated any phase of a redesign as very successful. Asking a company to adopt an unfamiliar technology while simultaneously dissolving the hierarchy through which it allocates authority compounds the risk. Even firms in decline often defend their structure longer than they defend their future.

Tenor's approach is to change the work without first demanding that the company redraw who reports to whom.

AI workers embed into the layers that already exist. A finance analyst can direct workers responsible for reconciliation, variance investigation and reporting while continuing to report to a finance manager. The analyst moves upward from repeated execution to direction, evaluation and judgment. The manager remains responsible for the person, the priorities and the performance of the function. The human hierarchy remains intact while productive capacity grows beneath every employee in it.

We agree on the destination: people should be closer to judgment, creativity, customers and consequential decisions, while machines handle more coordination and execution. But I do not believe companies must first abolish hierarchy to get there. Embedding AI into the organization can deliver the shift at a fraction of the switching cost.

What cannot be written down

This is not only an argument about adoption, but it also rests on a view of human judgment.

In 1960, J.C.R. Licklider described a future of human-computer symbiosis in which people would “set the goals, formulate the hypotheses, determine the criteria, and perform the evaluations,” while computers handled routinizable work.

His ambition was to redesign the division of cognitive labor, not just to build a faster typewriter.

The philosopher Michael Polanyi supplied the reason humans remain central: “We can know more than we can tell.” Expertise is not a collection of explicit rules. A skilled operator recognizes what matters before they can fully explain why.

Modern AI can search more material, track more variables, generate more options and execute more steps than a person could unaided, but more information does not decide what deserves attention. A model can produce an answer without bearing the consequences of choosing the wrong question.

Calling such a system an AI worker means it is given an organizational designation, not claiming that it is human. It gives a capable system bounded work, a named owner, an observable record and a place from which a person can direct and correct it.

Autopilot is not a perfect analogy. A cockpit is bounded and certified, while knowledge work is not. One crew operates one aircraft, while an employee may direct many autonomous workers. Pilot staffing therefore proves nothing about the economics of software leverage. However, what transfers is the supervision model: automation executes within defined limits while the human retains intent, judgment and authority over exceptions.

The warning against making that human ceremonial is Air France Flight 447. When inconsistent airspeed readings caused the autopilot to disconnect, the crew failed to recover from the resulting stall. Skills must remain current, control transfer must be explicit and intervention must remain possible. Accountability without the power to correct is only blame assigned after the fact.

The exception corpus

Most companies remember outcomes and forget how they reached them. A person notices something unusual, applies judgment, fixes the problem and moves on. The result remains, but the exception and the reasoning are buried in a conversation, an inbox or the memory of the person who resolved it.

AI workers change this. Because work is assigned and executed within a system, each task can leave a structured trace: what the worker attempted, where the normal procedure failed, when the human intervened, what they decided and what followed. Routine work becomes observable and, more importantly, the exception becomes reusable.

High-reliability systems already understand the value of such a record. NASA's Aviation Safety Reporting System was conceived after TWA Flight 514 struck a mountain in 1974. Six weeks earlier, another crew had misunderstood the same approach at the same location and narrowly avoided the mountain, but the warning remained inside that airline. NASA began operating a confidential reporting system so that one crew's experience could become aviation's knowledge. By April 2026, it had received more than 2.3 million reports and issued more than 8,150 safety alerts.

The NIST AI Risk Management Framework applies the same principle to AI: deployed behavior should be monitored, incidents and errors documented, human override preserved and adjudicated feedback returned to the system.

I think of the resulting organizational memory as the exception corpus. It is not only a log of machine mistakes but an account of how an organization exercises judgment when its rules run out. Each resolved exception can refine an instruction, create an evaluation, inform another authorized AI worker or become a precedent for the next unfamiliar case.

This is how tacit knowledge begins to compound without pretending it can all be formalized. The employee remains the teacher. Reports emerge from the work itself: what was completed, what failed, how the person corrected it, what it cost and what the organization learned.

From consumption to leverage

Companies are already spending heavily on AI, but much of what they can see is consumption. Seats. Tokens. API calls. Model costs. Usage rates. These numbers tell a company what it bought, not what the purchase accomplished.

The distance between adoption and value is already visible. In McKinsey's 2025 survey, nearly nine in ten respondents said their organizations regularly used AI, yet only 39 percent reported any enterprise-level EBIT impact.

An accountable worker changes the unit of measurement. Each worker owns discrete responsibilities. Its actions are logged, its output can be reviewed and its cost can be connected to the work produced. A company can ask not only how many tokens it consumed, but what was completed, to what standard, under whose direction and at what cost.

AI workers should have performance indicators and operating budgets just as human roles have responsibilities, salaries and expectations. Tokens, compute and tool calls are inputs, not achievements. The FinOps Foundation's unit-economics framework makes the distinction explicit. Cost per token measures a resource, but cost per agent action or case resolved begins to connect that resource to an outcome. The cost must still be weighed against speed, quality and risk. The cheapest work is not necessarily the most valuable.

Once AI workers have objectives and budgets, another measure becomes visible: how effectively the employee above them manages that capacity.

I expect this ability to vary widely. Two employees given the same models and the same budget will not produce the same result. One may spend heavily on work that never matters. Another may identify the decisive problem, divide it intelligently, assign it to the right workers, recognize weak output and redirect the system before resources are wasted. The difference is judgment, prioritization, delegation and control, not access to AI.

Management research gives us reason to expect this spread. An NBER study of more than 32,000 American manufacturing plants found that management practices accounted for roughly one-fifth of the variation in productivity. Forty percent of the variation in management appeared between plants belonging to the same company. Shared ownership did not produce uniform management before AI. Shared access to AI will not produce it now.

AI research makes delegation itself part of the picture. In an experiment with 758 consultants, people using GPT-4 performed faster and better on tasks inside the model's capability frontier. On a task outside that frontier, AI users were 19 percentage points less likely to reach the correct answer. The lesson is not simply to use AI more but to know what to delegate, what to inspect and when human judgment must take over.

Over time, leaders will allocate AI capacity as they allocate capital and headcount today. Employees and teams that repeatedly turn bounded resources into valuable outcomes will earn responsibility for larger AI workforces and budgets. These decisions must account for complexity, risk and strategic importance, not collapse into a ranking by tokens consumed or tasks completed. But the distribution of managerial leverage will become difficult to ignore.

This changes competition inside the company. The person who outperforms you may not work longer or produce more by hand. They may just be better at directing AI workers toward the company's goals. Access to intelligence will become common, but judgment about where to apply it will not.

Put people above it

Workers are right to be uneasy. In a 2025 Pew survey, 52 percent of American workers said they were worried about how AI would be used in the workplace. Only 6 percent expected it to create more opportunities for them, while 32 percent expected fewer.

No slogan can resolve that fear. AI can be used to eliminate roles, and some companies will use it that way. Augmentation is not an automatic property of the technology. It is an institutional choice.

The choice I want us to make is to give employees authority over this new capacity. Let repetitive execution move downward to AI workers. Let judgment remain with the person who understands the domain, the customer, the history and the cost of being wrong. Let that person grow from completing every task to directing a workforce that completes them. Let their manager remain a manager of people, responsible for human performance, development and trust.

The real question is not whether AI will enter the org chart but where the human will stand when it does: beneath the system, outside it or above it.

I believe the human belongs above it.

At 35,000 feet, the pilots are still there. The machine carries more of the load. The human carries the meaning of the mission and the authority to change course.

Knowledge work is ready for the same promotion.