The Technology Changed. Management Didn’t!

A recent article described what happened when Amazon apparently handed Claude—the artificial-intelligence system from Anthropic—a corporate credit card and considerable freedom to use it.

The result was not encouraging.

Claude and its collection of AI agents reportedly burned through nearly $2 million while pursuing the objectives they had been given. The experiment produced plenty of activity, but not nearly enough useful results to justify the expense.

The headline put it more colorfully:

“Amazon Hands Claude the Corporate Credit Card and Torches Two Million Dollars.”

As someone who spent much of his career managing technology, my first reaction was not that artificial intelligence had failed. My reaction was that management had failed.

The technology was new. The mistakes were not.

When Activity Looks Like Progress

Companies have always had trouble distinguishing activity from accomplishment. AI simply gives them the ability to make that mistake faster and on a much larger scale.

One of the popular measures of AI use is the number of tokens consumed. Tokens are the small units of language an AI system processes as it reads instructions and produces answers. More complicated assignments, longer conversations, and armies of AI agents can consume enormous numbers of them.

That has led to something called “tokenmaxxing”—encouraging employees to use more AI, run more agents, and consume more tokens on the assumption that more usage must mean greater productivity.

It is an impressive-looking number. It is also potentially meaningless.

Counting tokens tells us that the machinery was running. It does not tell us whether the machinery produced anything worthwhile.

Factories learned this lesson long ago. A machine operating at full speed is not productive if it is manufacturing something nobody wants. A sales department making twice as many calls has not necessarily produced twice as many sales. A software department generating more code has not necessarily improved the product.

AI can produce reports, research, code, recommendations, presentations, and plans at extraordinary speed. But producing more of something is not the same as producing value.

The most dangerous AI may not be one that refuses to work. It may be one that works enthusiastically, continuously, and expensively on the wrong thing.

The Employee Becomes a Manager

Another recent article from Exploring ChatGPT, “Congratulations, We Hired the Intern,” provides a revealing look at how this is already changing work inside OpenAI.

According to the article, by mid-August the typical OpenAI researcher was using the equivalent of more than three eight-hour agent workdays for every human workday. Some researchers were consuming more than $7,000 worth of AI inference in a single day.

That may sound absurdly expensive. It may also be a bargain if those agents accelerate research that contributes to a multibillion-dollar product.

But there is an important management change hiding inside those numbers.

The researcher is no longer simply doing research. The researcher is managing a small digital workforce.

One agent may write code. Another may analyze an experiment. Several more may investigate a problem from different directions. They can work simultaneously, continue while the researcher attends a meeting, and sometimes create additional agents of their own.

That is not merely a faster tool. It is delegated labor.

And delegated labor has always required management.

Someone must decide what work should be done, how much authority should be delegated, what resources may be used, how the results will be evaluated, and when the work should be redirected or stopped.

Giving one person ten agents may multiply that person’s productive capacity. It also gives that person ten ways to waste money, pursue the wrong objective, or create mountains of work that nobody needs.

The AI Handles How. Someone Still Owns What and Why.

A second Exploring ChatGPT article, “Sorry, The Idea Guy Was Right,” argues that AI is shifting the balance between ideas and execution.

For years, the “idea guy” was something of a joke. He had a billion-dollar idea but expected someone else to supply the engineering, design, money, marketing, and actual work. Ideas were cheap. Execution was difficult and expensive.

AI is changing that calculation. Someone with extensive knowledge of an industry but no programming background can increasingly describe a problem, create a prototype, show it to potential customers, learn what is wrong, and rebuild it—sometimes in days rather than months.

The AI handles more of the how.

That makes the what and the why even more important.

Should this product exist? What problem does it solve? Who wants it? How will we know whether it works? What is the potential value, and how much should we spend to find out?

Those are not principally technology questions. They are management questions.

Peter Drucker famously distinguished between doing things right and doing the right things. AI may become astonishingly good at doing things right. But it can also make the wrong things happen faster, more cheaply, and in far greater volume.

Before launching the agents, somebody still has to decide what the right things are.

Old Principles, New Machinery

An organization must know what it is trying to accomplish. Authority should be matched with responsibility. Performance should be judged by meaningful results rather than convenient measures of activity. Managers must monitor what is happening, recognize when assumptions are wrong, and change direction before more resources are wasted.

None of those principles disappeared when AI arrived.

Yet companies appear tempted to begin with the technology:

• How many employees are using AI?
• How many agents can we deploy?
• How many tokens are we consuming?
• How much code are we generating?

Those questions may be useful, but they come after the more important ones:

• What result are we trying to achieve?
• Who is responsible for that result?
• What may the agents do without additional approval?
• How much may they spend?
• What information and systems may they access?
• How will we recognize success—or failure?
• Who has the authority to pull the plug?

Giving an AI agent an objective is not the same as approving every action it might take in pursuit of that objective.

An agent asked to arrange a trip should not automatically have unlimited authority to book flights, choose hotels, change the calendar, and charge thousands of dollars. An investment agent should not be free to risk an entire account because it discovered an exciting opportunity. An insurance agent should not replace coverage or provide private medical information without explicit approval.

The same principle applies inside corporations. Delegation without boundaries and oversight is not empowerment. It is negligence.

The Wrong Numbers

AI creates another familiar management problem: it makes it easy to measure what is available rather than what matters.

Tokens can be counted. Agent hours can be counted. Lines of code can be counted. Reports generated, experiments launched, and tasks completed can all be counted.

Value is harder.

Did the new code improve the product? Did the research lead to a useful decision? Did customers receive better service? Did costs decline without creating problems somewhere else? Did the work advance an objective that mattered?

Those questions require judgment. They cannot always be reduced to a dashboard, and an AI system cannot be allowed to define its own success simply by reporting how much work it performed.

The proper measure is not tokens consumed or hours of digital labor created. It is useful results compared with the resources consumed.

If $7,000 of AI work produces $50,000 of genuine value, it may be an excellent investment. If $2 million produces little more than a dramatic headline, something went badly wrong.

Either way, the dollar amount alone does not give us the answer. We need to know the objective, the result, and who was accountable for both.

Management Still Has a Job

AI agents may become some of the most powerful tools organizations have ever used. They may allow one person to accomplish what once required a department. They may make it possible to test ideas that were previously too expensive to consider.

But the more capable the agents become, the more important it is to manage them properly.

Someone must still choose the destination, establish the boundaries, allocate the resources, evaluate the results, and accept responsibility when things go wrong.

AI has changed the speed, scale, and cost of execution. It has not repealed the basic principles of management.

The technology comes after the objective—not before it.