What Worked in the Past, What Is Working Now, and What Comes Next

Airline revenue management has been running on automation for decades. The system does not ask permission before optimizing bid prices; pricing decisions execute algorithmically thousands of times a day across thousands of markets. The idea that AI represents some sudden new relationship between a commercial team and a machine is mostly just a framing problem. For airlines, the relationship with automation has existed for a long time; what’s changing is the level of sophistication and the breadth of what it touches. I think it is worth understanding this history before the conversation moves to what comes next.

RM practitioners have relied on automation for years because the scale of pricing decisions is far beyond what people could manage manually. What’s changing is how intelligent that automation can become. AI is the next frontier in that space. The implication is not that airlines have solved the problem and are therefore ahead of consumer goods and retail. It is that the airline industry has been living long enough with questions of how much to trust the system, when to intervene, and how to measure whether an intervention was right to have formed some considered views.

The real issue concerns the nature of the relationship between the person and the system. It’s not a set-it-and-forget-it. You are steering the ship. You have to know what it’s doing, why it’s doing it, what it’s good at, and where its blind spots are. Guiding all of this are leaders who simply cannot accept an answer from an analyst of “I don’t know, the machine did it.” The pricing organization is responsible for the prices that customers see, and that responsibility does not transfer to the algorithm, regardless of how sophisticated the algorithm is. Understanding the system’s reasoning well enough to explain it—not at the formula level, but at the narrative level—is a prerequisite for the kind of accountability that boards and leadership teams are increasingly going to require.

Looking backward at what worked in the past, the honest answer for most pricing organizations is that the systems have been running on established science for long enough that the hard work of calibration has already happened for the core problems. The challenge now is that the core problems are no longer the only ones. AI opens up a much wider surface area of commercial decisions, including a long tail of medium-sized problems that never justified building a dedicated solution before, and reaching those requires a different kind of infrastructure than the one already in place.

What is working now is a model that may best be described as steering: the system handles the volume, and the person handles the judgment that the system cannot exercise on its own—the competitive move that does not fit any historical pattern, the market where the data science is technically right but contextually wrong, the situation where something has changed that has not yet shown up in the training data. Sometimes you just know something changed in the marketplace and you need to try something. The system tends to be highly factual; the person brings the situational awareness that facts alone cannot capture. Neither is sufficient without the other.

The piece that is hardest to get right, and most relevant to an audience of CCOs and RM Executives who are navigating this for the first time, is measurement. The temptation is to measure AI adoption by volume, such as how many recommendations were generated and how often users engaged with the tool. However, those metrics do not tell you whether the resulting decisions were good ones. AI isn’t creating a measurement problem; it’s exposing one we’ve always had. The pricing function has historically measured outcomes such as revenue, load factors, and yields. What it has not measured is the quality of the individual decisions that produced those outcomes, and without that infrastructure, there is no honest basis for evaluating whether AI is making things better or just more automated.

The fix is not complicated to describe. State a hypothesis before taking a pricing action—here is what I expect to happen, here is how I will measure it, here is what would cause me to change course. Record decisions at the moment they are made rather than reconstructing them after the fact. Track outcomes over whatever horizon is meaningful for your category and your customer base. Over time, a system that can reflect your own decision history back at you in similar situations becomes something different from a recommendation engine: it becomes a calibration tool, one that makes the judgment embedded in institutional knowledge visible and testable rather than tacit and personal.

What comes next, for any pricing organization serious about this, is the compounding of that calibration over time. The organizations I expect to be ahead in three years are the ones that started small, validated carefully, built trust incrementally, and built the measurement infrastructure alongside the capability rather than after the fact. Look back in six months and you’ll be surprised at how far you got. The pace of change in the underlying technology has made it tempting to go all in quickly, but the organizations I have watched navigate this well did not do that. They steered. They corrected. They compounded. And that, more than any particular capability they adopted, is what put them where they are.