Judgement in the Age of AI

By: Jordan Locke

Most conversations about artificial intelligence in revenue management eventually circle the same question: what will be automated?

It is a reasonable place to start. For decades, progress in this field has meant faster calculations, larger datasets, and systems that shoulder work once done by hand. 

But that framing misses the more interesting shift. 

Most conversations about artificial intelligence in revenue management eventually circle the same question: what will be automated?

It is a reasonable place to start. For decades, progress in this field has meant faster calculations, larger datasets, and systems that shoulder work once done by hand. 

But that framing misses the more interesting shift. 

Their first week happened to coincide with Las Fiestas de la Calle San Sebastián in Old San Juan, four days when the city fills with music, color, and the particular optimism that convinces every owner they are sitting on a goldmine.

On Monday morning, I asked them to walk me through the coming weekend. Bookings were pacing ahead of last year. Much of the competitive set was already sold out on Friday and Saturday. Shoulder nights were tightening.

“We should raise rates,” they said

.

It was a sensible answer. Well supported. The kind you expect from someone who has learned the mechanics quickly.

After lunch, I asked a slightly different question. “Were we right to raise rates?”

This time, the answer shifted. “If compression doesn’t carry into Sunday,” they said, “we may want to drop rates … just enough to stay competitive.”

The data had not changed.

The market had not changed.

Only the interpretation had.

That unsettled me more than I expected. If interpretation could shift so easily, then what exactly was I managing?

In that moment, it became clear that the exchange was less about inputs and outputs and more about management.

It took a moment to name what felt different about the exchange. At first it looked like automation: a machine executing instructions, a system returning a deterministic result. There was no single correct answer being revealed, no calculation that neatly resolved the question.

What I had been doing, instead, was watching how what I said, and how I said it, became a recommendation. In other words, I was being a manager.

Revenue managers have always relied on tools to extend their reach. Calculators relieve us of arithmetic. Spreadsheets allow us to hold more complexity than working memory ever could. Rule engines enforce consistency at a scale no human team could maintain.

Those systems operate in a narrow and unforgiving logic. They do exactly what they are told. They do not reinterpret instructions. They do not improvise. They do not surprise.

But new tools like large language models, and the agents built on them, are built differently.

They operate in language: the same medium we use to reason, persuade, hedge, and explain. Rather than executing fixed procedures, they generate plausible continuations of text based on patterns distilled from enormous volumes of human expression.

The result is a tool that feels less like a calculator and more like a junior analyst: capable, fast, and occasionally persuasive in ways that exceed its actual understanding. As Ethan Mollick puts it, “AI is like a super-smart intern who is very confident and sometimes very wrong.”

A calculator will always return that two plus two equals four. A language model might not. Not because it has decided to lie, but because it is not, in any meaningful sense, deciding at all. As Stephen Wolfram has observed, “It’s not reasoning about truth. It’s predicting what text is likely to come next.” Or, after training on feedback, simply guessing what you want to hear.

Humans, of course, are not so different. We also make mistakes. We misread. We carry unexamined assumptions. We grow overconfident. We’re all kind of like AI, or at least like an intern, right?

The difference is not that people are flawless and machines are not. That distinction misses the point. What matters is that we have spent decades building professional cultures around noticing, interrogating, and correcting human intelligence and judgment.

We have not yet done the same for artificial ones.

Some domains lend themselves naturally to verification.

Arithmetic does. If two plus two equals four, it equals four every time. Chess, though vastly more complex, still resolves cleanly: the king is either in checkmate or it is not. Many engineering problems behave similarly. Code compiles or it throws an error.

Revenue management does not offer this kind of closure.

When a rate is adjusted for a holiday weekend, the outcome arrives slowly, and only once. Bookings accumulate. Cancellations appear. A final occupancy number settles into place.

What never materializes is the counterfactual. What could have happened.

We do not see, side by side, the weekend that would have occurred had the price been five dollars higher. Or ten dollars lower. Or … GASP! … left untouched.

Instead, we work with shadows of comparison: last year’s performance, a competitive set that is never truly comparable, a forecast built on assumptions that were already imperfect when they were made.

In this environment, there is rarely a single right answer. There are only better and worse judgments, visible in hindsight and debated in the present. As John deRoulet so kindly put it, “being a revenue manager is just second guessing yourself until you have an existential crisis.”

This ambiguity poses a challenge for artificial intelligence. It also poses the same challenge for humans.

A model can surface patterns, propose adjustments, and articulate tradeoffs. It can do so at speeds that would have been unimaginable only a few years ago. But it cannot tell us, with certainty, which recommendation is correct. Because in revenue management, correctness is not something the world reveals cleanly.

It is something we infer.

And inference, by definition, lives in the realm of judgment.

By the end of that first week, I had stopped thinking of the system as a piece of software.

The way it responded to my behavior began to feel familiar. When I gave it narrow, poorly framed questions, it returned narrow, poorly framed answers. When I supplied context (what we were trying to accomplish, where risk mattered more than upside, which properties could tolerate volatility and which could not) the quality of its recommendations improved.

When I corrected it, it explained why and adjusted. When I delegated loosely, its outputs drifted. None of this was especially mysterious. It was the same pattern I had seen with human analysts for years.

The difference was speed.

The system could generate in seconds what once took hours. Reports rendered. Numbers checked. Alternative approaches appeared without even being explicitly requested. What did not disappear was the need to decide what to trust.

In the Army, every mission is framed with two elements: a task and a purpose. The task defines what must be done and under what conditions. The purpose explains why it matters. That distinction exists because environments change faster than instructions can be updated. When circumstances shift, people who understand only the task wait. People who understand the purpose adapt.

That distinction mattered in environments where waiting could cost lives. It matters here for quieter reasons, but the principle is the same. Swapping the fog of war for the fog of market conditions, I found myself applying the same structure to artificial analysts. Not just what to evaluate but why we cared. Seen this way, the lesson turns out to be managerial.

Good managers produce better outcomes, regardless of whether they are managing people or machines. Bad habits like vague direction, blind delegation, and abdicated responsibility scale faster when AI is involved, because the system scales both insight and error.

The presence of AI does not absolve anyone of judgment.

It amplifies the consequences of how that judgment is exercised.

Complexity didn’t recede. It thickened.

When only a handful of options could be evaluated, the act of choosing among them felt manageable. When dozens of plausible paths can be simulated instantly, selection itself becomes the constraint.

Herbert Simon once observed that a wealth of information creates a poverty of attention. In revenue management, that poverty expresses itself as something more specific. When analysis becomes abundant, judgment becomes scarce.

Scarcity, in any system, is where value concentrates. The emerging advantage lies in recognizing which ideas deserve to survive contact with reality.

AI can propose, simulate, and suggest. What it cannot do is decide what matters.

That remains a human responsibility and it is becoming heavier, not lighter.

Economists studying the effects of artificial intelligence have begun to converge on a similar conclusion.

If there’s one thing I’ve come to believe while building Prometheus, it’s this: that reorganization feels less dramatic than headlines suggest.

AI has made me faster. It has increased my scope. It has not made decisions disappear. If anything, it has made them more visible. My role has gone from calculation to command.

When generating options becomes effortless, choosing among them becomes the real work. When analysis is cheap, discernment is expensive. This is the quiet truth underneath much of the current noise. Artificial intelligence does not eliminate judgment in revenue management; it concentrates it. The tools are converging while judgment is not.

I recently asked a colleague what he thought about working in a market where every revenue manager has access to the same technology.

He smiled.

“Good,” he said. “Now it’s about who knows what they’re doing.”