Original Photo: iStock.com/ozgurdonmaz
There is a specific profile of operating environment where AI agents perform exceptionally well. The tasks are well-defined, the decision criteria are measurable, and the data inputs are high-volume, structured, and arrive at a predictable cadence. The cost of a missed or delayed decision is quantifiable—in dollars, on a ledger, and attached to a specific flight on a specific date.
Airline commercial teams match that profile on every dimension.
An RM analyst is, at the core, solving the same problem Peter Belobaba formalized in the 1980s: given uncertain demand, finite and perishable inventory, and a set of fare classes with known prices, what is the optimal set of controls to maximize expected revenue? The inputs are structured, the objective function is clear, the feedback is eventually measurable, and the volume of decisions (across thousands of markets daily, with booking curves that shift and competitive signals that arrive without warning) is far beyond what any team of human analysts can fully cover.
If you were designing an operating environment from scratch to showcase what AI agents are capable of, you might end up describing a revenue management department.
So why has adoption lagged the opportunity?
The answer is not cultural resistance, though that would be the easy explanation. Airline commercial teams are not afraid of technology. They have spent decades adopting complex forecasting systems, revenue management platforms, competitive intelligence tools, and distribution infrastructure of considerable complexity.
The real answer is more specific: nobody has built a system that respects both the complexity of the domain and the way human analysts actually do their work. These are two different problems, and solving only one of them produces a tool that either oversimplifies or overwhelms.
The Data Is There. Whether It Gets Used Is a Different Question.
The conversation about AI in airline commercial operations tends to get stuck early, on data. Do we have enough? Is it clean enough? Is it labeled correctly? These are reasonable questions, but they are often asked in a way that obscures a more important point.
Airline commercial teams are among the most data-rich operating environments in any industry. Every booking, every fare transaction, every schedule change, every competitive filing, and every market performance metric is captured, timestamped, and stored. The underlying Passenger Service System collects this continuously, whether or not anyone is doing anything with it.
What varies enormously from airline to airline is not whether the data exists, but whether it has been procured, structured, and made accessible in a way that enables analysis. Some carriers have mature, well-governed data infrastructure: PNR data, ticket data, booking data, and ancillary revenue data visualized in robust business intelligence dashboards that an analyst can query. Others are in a position where answering a question as basic as “what was our average load factor in January?” requires a cross-functional effort, a SQL query nobody has run before, and a spreadsheet that takes three days to build.
The data is sitting somewhere in both cases. The difference is whether it has been made useful.
This is where the AI conversation has to get honest. AI is not magic, and large language models are not magicians who conjure insight from nothing. The tools are only as good as two things in combination: the quality and structure of the inputs they receive, and the competence and domain expertise of the team building and maintaining the system.
A well-architected AI platform can do significant work to clean, validate, and normalize messy data inputs. But it cannot manufacture data that was never captured or produce reliable signals from a foundation that was never built. Either the airline brings clean, procured data to the platform, or the platform needs to do that work; the platform doing that work is a solvable problem, but it is engineering work that has to be scoped, resourced, and executed with airline-specific domain expertise, not a just generalized data engineer.
What Analysts Should Actually Be Doing
Here is what is worth protecting: the people who populate airline revenue management and pricing teams are, as a group, highly capable. Entry-level RM analysts are frequently quantitatively trained, analytically sharp, and genuinely motivated by the intellectual challenge of the work. Senior analysts and managers carry years of market intuition, competitive pattern recognition, and operational judgment that is genuinely difficult to replicate.
The tragedy of the current state is what those people spend their time on.
Pulling data and building spreadsheets of considerable complexity that only one person on the team fully understands and that break when that person is out sick; reformatting outputs from one tool to serve as inputs to another; clearing exception queues because the system has no way to distinguish signal from noise.
None of that is analysis, or interpretation, or the work those analysts were hired for and are equipped to do well.
The case for AI in airline commercial operations is not that it replaces analysts. It is that it gives analysts their actual job back—the work of applying judgment, thought, and domain expertise to decisions that genuinely require it, rather than spending the majority of their hours on the infrastructure of getting to the question.
That is a straightforward value proposition. The reason it has taken this long to deliver on it is because building it correctly requires understanding the domain at the level of someone who has worked on the inside—not just theoretical revenue management, but the operational reality of how data flows, how decisions get made, and where the current tools break down.
That is the problem telos was built to solve.
