I spent a day this week at the third annual gathering the travel industry now holds to take stock of where its AI ambitions actually stand. It was a really, really good conference. The framing was honest, more honest than these events usually are. The host had stopped asking whether the industry would adopt AI and started asking how deeply it already had. The theme on the opening slide was a single phrase: operational AI. The implied question underneath it was sharper. Not are you doing AI, but how far has your execution traveled from your intention?

It is worth holding onto that question, because the rest of the day kept answering it, often without meaning to.

What the room talked about

The opening session organized the industry’s struggles into five tensions and put each one to a live audience vote. Pilots versus production. Speed versus trust. Restructuring versus readiness. Build versus buy. Control versus visibility. They are good tensions. They are real. The audience split close to the middle on most of them, which is the sign of a debate that is genuinely live rather than already settled.

But notice the shape of the list. Four of the five tensions are about how to build and deploy. Only the last one, control versus visibility, touches the commercial question directly, and even there the framing was defensive: how much of our first-party booking funnel do we surrender to AI intermediaries? It was a conversation about not losing ground, not about gaining it.

The executive panels that followed reinforced the pattern. The infrastructure stories were genuinely impressive. One major hotel group described spending eight years breaking a legacy monolith into a modern cloud architecture, taking its reservation system from roughly a hundred million transactions a day to well over a billion. A major airline described coming off the mainframe only last year and rebuilding around a single connected view of the traveler. These are real achievements and the right foundation to build on.

When the conversation turned to what that foundation was being used for, the examples clustered in three places. Cost takeout in the contact center and back office. Service automation at the front desk and in the call queue. And the consumer-facing dream: conversational search, the AI trip planner, the agent that helps a guest discover and eventually book. The marquee demos were all discovery and booking tools. The marquee savings were all operational.

What the room did not talk about

Pricing came up, but it came up the way the weather comes up. Acknowledged, not engaged. One technology partner on stage said plainly that dynamic pricing is where the industry is least mature, that there is only “a little dabbling” so far, and that it is “starting to evolve.” A taxonomy slide later in the day ranked operational AI across nine layers of depth, and revenue and commercial automation sat squarely in the middle of the stack, described in a single line as “dynamic pricing, human oversight exception-based.” It was named. It was ranked. Nobody on stage claimed to be operating there.

The most telling moment was an aside, not a headline. An executive described sitting with a hotel revenue manager and watching the manual reality of the job: the proliferation of spreadsheets, the processes held together by one person’s memory, the work that breaks when that person is out sick. The point being made was about workflow and change management. But sitting in the audience, I heard something else. I heard a description of the single most direct path to revenue that the company owns, being run on tooling that would embarrass most other functions, and being discussed as a change-management problem rather than the largest unclaimed opportunity in the building.

Why this lever can’t be the one left out

Here is the part that should round out the priority order. The investments the room is making are the right ones. Service automation and a stronger booking funnel matter, and they matter more, not less, as demand begins to move through agentic channels; a brand that fails to invest in capturing demand in its future form risks being absent from the very interfaces travelers will use to discover and book. None of that is in question. What is missing is a parallel investment in what happens once that demand is captured, because the economics of revenue optimization are unlike anything else on the list.

Consider the structure of the business. A hotel room or an airline seat is perishable inventory with a fixed, finite supply and a hard expiry. The marginal cost of selling the last unit is close to nothing, which means almost the entire value of a better pricing decision falls to the bottom line. A point of RevPAR, a point of yield, a point of RASM is not revenue that arrives with a cost of goods attached. It is close to pure margin. This is why the discipline exists in the first place and why it has historically been one of the highest-leverage functions in the company.

Consider the difference. Service automation lowers cost, and cost has a floor; you cannot save more than you spend. The booking funnel and conversational search are how a brand stays present in the channels where demand is increasingly going to form, which is essential work and only becomes more so as agentic AI reshapes discovery. But both of those efforts are about getting demand in the door. Revenue optimization is about what that demand is worth once it arrives: charging the right price, on the right unit, at the right moment, across thousands of markets a day. It is a different lever entirely, and for a brand with perishable inventory it is the one that converts the same captured demand into the most margin.

And revenue optimization is, by almost any measure, the textbook environment for the kind of AI the industry says it wants to deploy. The tasks are well defined. The objective function is explicit: maximize expected revenue against uncertain demand and perishable inventory. The data is high volume, structured, and arrives on a predictable cadence. The cost of a missed or delayed decision is quantifiable in dollars and attached to a specific unit on a specific date. If you set out to design a showcase environment for capable AI agents, you would end up describing a revenue management department. The fit is nearly perfect, and it is the one application the summit treated as a footnote.

Why the commercial judgement problem is a real opportunity

The neglect is not irrational, which is what makes it durable. But of the three forces holding it in place, one is structural in a way the others are not.

Most commercial operations cannot currently say whether the pricing decisions they made last week were good ones. They measure revenue against target and load factor against last year, but not decision quality at the level of the individual call. If you cannot measure whether today’s decisions are good, you have no way to prove that an AI made them better, and a function you cannot instrument is a function nobody is eager to hand to a machine. The honest first move toward revenue AI is not a model. It is the feedback loop that lets you grade the decisions in the first place.

The other two forces are real but secondary. Visibility: a conversational trip planner makes for a compelling keynote demo. A two-point improvement in yield does not as photograph well on a conference stage (but it plays really well in a boardroom). Difficulty: pricing and inventory management AI cannot be bought as a generic product and bolted on. It demands a system that respects both the genuine complexity of the domain and the way analysts actually work, built by people who understand the operational reality from the inside.

The takeaway worth carrying out of the room

The summit’s organizing question was how far execution has traveled from intention. On revenue, the honest answer is: not far, and not because anyone decided it shouldn’t. The intention is rarely even stated. The industry has quietly agreed that AI’s job is to capture demand more cheaply and serve it more smoothly, and has left the question of what to charge for that demand sitting where it has always sat, in spreadsheets and in capable revenue management systems, with analysts who deserve a judgment layer built on top of both.

That is the opportunity, and it sits alongside the others rather than in place of them. For a mid-to-large travel brand, the booking interface and the service layer are worth building, and worth building now. But the lever most likely to be forgotten is also among the largest: the function that turns the same captured demand into more margin, run on tooling worthy of how much it moves. The brands that get this right will keep investing in capturing tomorrow’s demand and add the discipline of pricing it well. They will not be buying a feature off a vendor shelf. They will be building a commercial operation that learns, and they will be doing it while the rest of the room is still applauding the trip planner.

A note on sourcing: this post reflects observations from the Skift Data & AI Summit and intentionally omits the names of companies and speakers. The themes, not the attributions, are the point.