“AI” now appears on nearly every hotel software page, which makes it hard to tell what’s real. If you strip away the marketing, AI in revenue management is a fairly simple chain of steps, and understanding them makes you a much better buyer. Data goes in, a model learns patterns and produces a forecast, an optimiser picks prices, and the system checks what actually happened and learns from it. None of that is magic, and none of it needs a degree to follow. This article walks through each step in plain language, with small examples, and finishes with what AI can’t do, so you can tell a useful product from a label stuck on an old one. If you know the steps, vague sales answers become easy to spot.
Step one: the data it learns from
Everything starts with data, mostly your own. Past reservations, including the date booked, the stay dates, the rate, the room type, the channel and whether it was cancelled. Current bookings for future dates. Competitor rates. Local events and holidays. The quality of this data matters more than the cleverness of the model. If rooms that were out of order show up as sold, or bookings were entered under the wrong date, the model learns false patterns and gives confident wrong answers. Most of this data lives in your property management system, which is why the connection between the PMS and the pricing tool is so important. A tool that only gets a copy every few hours is learning from yesterday’s picture. It’s the least glamorous part of AI and the one that decides most outcomes. Some tools also pull in extra signals, like search activity or market data, which can help, but only once the basics are clean.
Step two: turning data into a forecast
Next, the model looks for patterns. How do bookings build up before arrival on a typical Saturday? How does the pace change in summer? What happens around a local festival? Take a simple example of hotel occupancy forecasting: 40 rooms are booked 30 days before a given date, and historically a date like it adds another 25 rooms by arrival. The simplest forecast is 65. A machine learning model does something similar but blends many signals at once, like day of week, season, lead time, cancellations, and events, and learns how much weight each deserves from your own history. The output is a forecast of how many rooms you’ll sell on each future date at current prices. Notice that it’s a prediction, not a certainty, so good tools show how confident they are, and how that forecast did against reality last time. It’s a bit like a weather forecast. A 70 percent chance isn’t a promise, but over a hundred days it tells you something real.
Step three: choosing a price
Once there’s a forecast, the system asks how demand would change if the price moved. On a date that’s filling fast, a ten percent rise might barely change bookings, so the rate goes up. On a slow date, a small cut may bring in a few extra rooms, so the rate goes down a little. The optimiser looks for the price that’s likely to earn the most, inside your limits: floor, ceiling, relationships between room types and restrictions like minimum stays. Good revenue management software shows the reasoning in plain terms, such as “pickup ahead of last year” or “competitor sold out”, so you can judge whether it makes sense. If a tool can’t explain a price, you’re being asked to trust a black box, and with your income at stake, that’s a lot to ask on faith.
Step four: learning from what happened
After the date passes, the system compares its forecast with the real result. Did it predict 65 rooms and sell 70? Where was it off, and why? It uses those misses to adjust. That feedback loop is why a tool usually gets better over the first few months, and why a brand-new hotel with little history gets weaker forecasts at first. It’s also why patterns change over time. If guests start booking closer to arrival than they did two years ago, a model that doesn’t update will be too pessimistic early on. Ask any vendor how often their model retrains and whether you can see accuracy figures for your hotel. A good answer is specific. A vague one, full of “proprietary” and “advanced”, usually means the vendor doesn’t want to be measured, or hasn’t looked closely. That’s also why a tool needs time to settle in, and why you should judge it over a season rather than a week.
What AI can’t do
AI can’t know about the concert your cousin mentioned or the road closing next month. It can’t tell that a booking spike came from a system error. It can’t decide whether you want to be the cheap option or the premium one. And it can’t fix poor reviews or a tired lobby. Some products labelled AI are really fixed rules, like “add ten percent above 80 percent occupancy”. Rules aren’t bad, but you shouldn’t pay AI prices for them. Ask what the model learns from, how it’s measured, and what happens when it’s wrong. Pair it with a channel manager that pushes rates everywhere quickly, because a good price stuck in a dashboard earns nothing. The best setup is the system proposing and a person supervising, with the person’s local knowledge as the safety net.
Once you know the four steps, AI in revenue management looks a lot less mysterious: good data, a forecast, a price, and a feedback loop. Aiosell’s revenue management software uses AI for the daily forecasting and rate suggestions, working from the same data as its property management system and channel manager, while leaving the controls and the final say with you. When you talk to any vendor, ask them to walk you through these four steps for your hotel. The ones who can do it clearly are usually the ones worth trusting.
