What If AI Is Making Companies More Efficient and Less Intelligent?

We have become very good at measuring what AI can save: hours, headcount, cost, time. We are getting better at measuring what it can produce too.

There is nothing inherently wrong with that. If a company can achieve the same result with fewer resources, it would be strange not to explore the possibility.

But I have been wondering whether we are measuring the easiest part of the equation.

A mathematician friend of mine, someone who understands AI from behind the scenes rather than from the headlines, recently sent me a paper called “The AI Layoff Trap”, by Brett Hemenway Falk and Gerry Tsoukalas.

Their argument is economic and, at its core, relatively simple.

When a company replaces workers with AI, it captures the full cost saving. But displaced workers are also consumers. If their income is not replaced quickly enough, some of their spending disappears with it. The company that automated bears only part of that loss in demand; the rest is spread across the market.

The authors describe what follows as an “automation arms race.” Each company has a rational reason to continue automating even when, collectively, they could end up worse off.

That was the part that stayed with me.

Not the prediction about layoffs, but the paradox behind it.

Nobody needs to be stupid for the system to behave stupidly.

The companies in their model can see what is happening. They can understand the consequences. And still, individually, they have an incentive to continue. Knowing where the road leads doesn’t necessarily give any one of them a reason to take another road.

I started wondering how often something similar happens inside companies.

After all, organizations are full of decisions that make perfect sense from where someone sits. Finance can build a compelling case for reducing costs. Technology can make another for automating a process, and HR may have the numbers to support operating with fewer people. The productivity figures may support all three.

Nobody has to be wrong.

But organizations are not collections of independent business cases.

And this is where, for me, the AI discussion becomes much more interesting than the debate about how many jobs it will eliminate.

Suppose AI can now perform 60% of someone’s work.

What exactly does that 60% represent?

Most likely, the things we can observe and measure: reports produced, emails written, documents reviewed, transactions processed, code generated, presentations prepared.

But I am not sure a person’s contribution to an organization has ever been simply the sum of those tasks.

Anyone who has spent enough time inside a company has probably encountered the person who looks at a perfectly reasonable number on a dashboard and says, something is wrong here.

Sometimes they cannot immediately explain why. They remember something from two years ago. They know that a particular customer behaves differently. They recognize an exception that isn’t documented anywhere. They know that two systems define the same thing differently. Or they know whom to call because they have seen something similar before.

None of this looks particularly impressive on a job description, and it is remarkably difficult to put into a productivity calculation.

Yet organizations depend on it every day.

There is knowledge in databases, reports, policies, process maps and systems. There is also knowledge sitting between people, buried in old decisions, accumulated through mistakes, embedded in relationships and carried around in people’s heads.

Much of it becomes visible only when something doesn’t go according to plan.

That is what concerns me about some of the current conversation around AI and headcount. Reducing headcount is not inherently wrong. Sometimes it is exactly what a company should do.

What worries me is the assumption that because AI can reproduce a percentage of someone’s observable tasks, we have therefore measured the same percentage of their contribution.

We haven’t. And the gap between those two numbers is not a rounding error. It is precisely the part that doesn’t show up on the dashboard until the day the organization needs it: the exception no one documented, the mistake no one wants to repeat, the context everyone assumed someone else would remember.

Maybe what remains has little value. Maybe it can be transferred to somebody else, captured properly or replaced by AI as well. If so, automate it.

But it would be useful to know that before doing so.

A company could reduce a team from ten people to three, maintain exactly the same output and quite reasonably declare the transformation successful.

It might be.

Then something unusual happens.

A market changes, a supplier fails or a customer behaves unexpectedly. Two sources of information contradict each other. A model produces a plausible answer that happens to be wrong. Suddenly, maintaining yesterday’s output is not the problem anymore.

The organization has to understand something it hasn’t seen before.

And I suspect this is where some of our current measures of AI success may eventually prove insufficient.

AI is getting very good at doing things organizations can describe, document and feed into systems. Companies will inevitably use that capability to remove work that no longer requires a human being.

But before asking How many people do we need now?, there may be another question worth asking:

What did those people know that the organization still needs to know?

I don’t think most companies can answer that question particularly well.

And if they can’t, we may become very good at automating what people did while having very little idea what disappeared with what they knew.

The numbers could still look excellent: lower costs, higher productivity, faster execution, more output per employee.

But those numbers won’t necessarily tell us whether the company still has the context to understand an exception, the memory to avoid repeating an old mistake, or the judgment to recognize when a perfectly plausible answer is wrong.

Perhaps that is the distinction we should pay more attention to as AI changes the way companies work.

Not simply whether an organization can do more with less.

But whether, after doing so, it still understands what it is doing.

And that leaves me with a question I haven’t quite answered yet:

What if the companies that ultimately win the AI race aren’t the ones that need fewer people, but the ones that become more intelligent with the people and machines they have?