The loudest claims that AI will eliminate nearly all human work tend to come from a narrow group: people who build AI models and write software, and the people who cover them for a living.
This should not be surprising. Every new technology is first understood by the people closest to it.
The problem is that many of the people making sweeping forecasts about the future of work have spent their careers doing one particular kind of work—the exact kind of work that AI got good at first.
I have led companies most of my adult life. I have not written production software in decades. So I read these forecasts the way I read any confident prediction from a specialist: I respect the expertise, but I question the extrapolation.
Here is the extrapolation in question:
Great programmers are rare. They have one of the most cognitively demanding jobs in the economy.
If machines can do what programmers do—even the successful ones in Silicon Valley—what chance does the accountant, recruiter, manager, or salesperson have?
It seems like a reasonable leap, but I think it is a category error.
It assumes every job can be measured on one axis: intelligence required. Programming is at the top, and everything else sits somewhere below it. If AI clears the bar for programming, it has cleared the bar for everything else too.
But that intelligence axis is far from the only axis that matters. The picture is more complicated.
Why Programming Was an Early Target
Programming was an early target for AI because its basic conditions are unusually favorable to machines.
First, much of the work is fully represented in text. The code, documentation, and surrounding discussion are available to inspect, and the artifact is relatively self-contained.
Second, programming has been documented at enormous scale. There are decades of public code, technical documentation, questions, answers, and examples. Few professions have created a comparable body of accessible training material.
Third, and most important, programming produces unusually fast feedback. Code compiles or it does not. The software works, or it breaks. The feedback is often immediate and relatively objective.
Machines thrive in that kind of environment.
Programmers will correctly point out that this is not all software engineering is. Real software work involves unclear requirements, tradeoffs, organizational politics, legacy systems, and deciding what should be built in the first place.
Exactly.
AI has been strongest in the parts of programming with clear inputs, clear outputs, and fast feedback. It has been less reliable at deciding what the company should build, why it should build it, and what tradeoffs it should make.
AI’s success with code reflects the nature of that feedback loop. It tells us much less about professions whose hardest decisions do not produce a quick, objective answer.
Consider the Accountant
Everyone assumes accounting is easy prey for AI, and in some areas it is. Arithmetic was automated a long time ago. Reconciliation, categorization, reporting, and routine analysis will keep becoming more automated.
But the valuable part of the accountant’s job is often not arithmetic. An accountant may need to decide whether an unusual contract should be recognized as revenue now or over several years, or whether a reserve is prudent or aggressive. Sometimes the job is telling a CFO that the number they want is not supportable.
In other words, that kind of work does not grade itself. An auditor might challenge the judgment years later, or it might become an issue during a restatement. In many cases, no one will ever know with certainty whether another judgment would have produced a better result.
There is no test suite for materiality. There is no clean answer key for professional judgment under uncertainty.
The CEO Is the Extreme Case
The same feedback problem becomes even more pronounced in the CEO role.
Every significant decision made by a CEO carries a real possibility of being wrong, because the CEO must act before the information is complete.
Anything with a clear answer should be decided below the CEO level.
The decisions that reach the CEO are there because capable people disagree. Enter this market or stay out? Hire this executive or keep looking? Invest in growth or protect cash? Sell the business or build for another five years?
I might know the outcome in three years, but I will never know the counterfactual. I cannot run the company under two strategies at once or hire both candidates to see which one would have been better. Every choice closes off the alternatives that might have proved it wrong.
The real world will grade the decision slowly, and you may never get a definitive answer.
Then there is the CEO’s responsibility to balance shareholders, customers, and employees. These groups do not always want the same thing. The CEO has to make a judgment and explain it well enough to earn the trust required for execution, even when there is not a “right answer.”
What This Means for CEOs
For CEOs, this is a more useful way to think about AI than simply asking which jobs it will eliminate.
AI will eliminate some jobs. Its broader effect, however, will be to redistribute the tasks inside them and change what people are responsible for.
Programming is again a useful example. AI can now automate a great deal of coding, yet the U.S. Bureau of Labor Statistics projects that employment of software developers to grow 10 percent from 2025 to 2035. The programmer’s role is likely to shift toward more of the judgment around what to build and whether it works, even as the tool handles more of the implementation.
The question that best indicates which jobs will be affected most is: How quickly and objectively can we tell whether the work was done well?
When the feedback loop is fast and clear, AI will improve rapidly and automate a growing share of the work. When feedback is slow or contested, AI can still contribute, but people will carry more responsibility for the judgment and the consequences.
The same principle applies to the CEO’s own job. AI can do a lot for the CEO. It can act as a thinking partner, clarifying decisions, identifying incorrect assumptions, surfacing second-order consequences, arguing the opposing case, comparing options against the company’s strategy and values, and sharpening the explanation to the people who must execute. It can also bring relevant context to the surface faster than a human team can assemble it and help a CEO see patterns or inconsistencies they might otherwise miss.
But AI is an input, not an owner. It cannot carry the accountability of choosing a strategy, hiring an executive, committing capital, or deciding which stakeholder interest should prevail when they conflict. The CEO must still make the call and earn the organization’s commitment to it.
Bottom line: We should take technologists’ warnings seriously while recognizing that their experience with code is not a complete theory of the future of work.
The practical question for every CEO is: Where does your company have fast, objective feedback loops, and where does it rely on human judgment without a scoreboard? That answer should inform where, when, and how you put AI to work.
If programming is unusually favorable terrain for AI, the CEO role may be one of its hardest tests. That is part of what makes ChatCEO interesting to me. We are building an AI that can help where company context matters and the right decision may never be fully knowable. Sound interesting? Sign up at chatceo.co.
