Using Guest Messages to Improve Reviews
Guest messages show problems before they appear in public reviews. Learn how guest feedback AI for hospitality groups them by topic and turns them into fixes.
AI demand forecasting means using a model to predict how much of something will be needed, such as units sold, patients arriving or electricity used, so you can plan ahead. A practical start is small: pick one thing to forecast, gather its history, beat a simple baseline, show a likely range instead of a single number, and keep a person in charge of the decision.
This article is for planners, operations leads and finance teams who want to try forecasting with AI without a large programme. It covers what to forecast first, the data you need, how to know the forecast is any good, and what tends to go wrong.
What demand forecasting is, in plain terms
A forecast is a statement about the future built from the past plus what you know is coming. Retailers forecast sales per store. Hospitals forecast admissions. Utilities forecast electricity load. Logistics teams forecast parcel volumes.
The AI part is that a model learns patterns from history, such as weekly cycles, seasons, holidays and the effect of weather. It can use many signals at once, which is hard to do in a spreadsheet. But the model is only one part. Data, checks and the decision process around it matter as much.
Start with a baseline, not a model
Before you build anything clever, write down what you do today. Many teams already forecast with a simple rule, such as "same as last week" or "average of the last four weeks." These are called baselines.
Baselines matter for two reasons. First, they tell you the bar the AI must clear. Second, they are surprisingly strong. A seasonal baseline is hard to beat on stable products, and a model that cannot beat it is not worth deploying.
| Approach | How it works | Good for | Weak at |
|---|---|---|---|
| Last period | Repeat last week or last year | Stable, steady demand | Trends, events, weather |
| Moving average | Average of recent periods | Smoothing noise | Sudden changes |
| Statistical models | Fit trend and seasonality | Clean, regular series | Many outside factors |
| Machine learning | Learn from history and extra signals | Demand driven by weather, promotions, events | Little history, messy data |
Use the simplest method that meets the need. Move up only when you can show the gain on your own numbers.
Pick a first forecast worth making
Choose one target that passes these checks:
- A decision depends on it. Staffing, stock, procurement or capacity changes based on the number.
- There is history. At least a couple of seasonal cycles, such as two years of daily data for something with a yearly pattern.
- The data is recorded the same way each time. A change in how you count breaks the pattern.
- Errors are tolerable at first. Start where a miss means extra stock or a manager's phone call, not a safety incident.
- You can measure the result. You need a before number to compare against.
Weekly branch sales, daily call volumes and monthly material orders are all good candidates. A single new product with no history is not.
What data you need
The core is a clean history of the thing you are forecasting, with a date on every record. Then add signals that plausibly drive it.
- Calendar: weekdays, public holidays, school terms, Songkran and other local events.
- Weather: temperature and rainfall, especially for energy, tourism, retail and agriculture.
- Business drivers: promotions, announced tariff or fee changes, new store openings, marketing campaigns.
- Supply side: stock-outs. If a product ran out, low sales reflect supply, not demand, and the model will learn the wrong lesson.
Spend time on missing values, duplicates and changed definitions. Most forecasting projects spend more effort here than on modelling. If your data lives in many systems, read building a data foundation before you build AI first.
See it working: the power grid demo
Reading about forecasting is one thing, and seeing it is another. Our AI power demand forecasting demo is a working demo with sample data, not a real client result. It shows the pattern that applies well beyond electricity:
- Live load and weather feed the model.
- A 48-hour forecast is shown with a likely range, not just one line.
- Accuracy is checked against simple guesses, which is the baseline idea above.
- Questions can be asked in plain language.
- A person decides, and every answer is logged.
The same layout suits a retailer planning stock or a hospital planning staffing. Only the data and the decision change.
Show a range, not a single number
A single number gives a false sense of precision. Demand is uncertain, and good forecasts say how uncertain.
A range such as "between 80 and 120 units, most likely around 100" lets a planner decide how much safety stock or spare capacity to hold. For example, a retailer may accept more risk on a slow-moving item and almost none on a staple. The range makes that trade-off visible. It also helps you judge the model later: if actual results fall outside the range far more often than stated, the model is overconfident.
Measure whether the forecast is good
Set up testing before you trust any model.
- Hold out recent data. Train on the older history and test on a recent period the model has not seen. Never test on data used for training.
- Compare against the baseline. Report how much better, or worse, the model is than the simple rule.
- Check errors by segment. A model can be accurate overall and wrong for your busiest store or your peak week.
- Track bias. Consistent over-forecasting or under-forecasting is a different problem from random error, and it is often easy to fix.
- Monitor after launch. Forecast accuracy falls when behaviour changes. Review it on a schedule.
Choose one or two measures that planners understand, such as average percentage error, and use them throughout. A measure that only the data team understands will not build trust.
Connecting the output to your systems
Extraction only pays off when the result lands where work happens. Decide early where each field goes: an accounting system for invoices, a customer record for ID data, a case system for permits.
Send structured data through an interface, not through a spreadsheet that someone re-keys. Keep a link from every record back to the source image, so an auditor or a colleague can see what the system read. Store the confidence score and the reviewer's name alongside each corrected value. These details make audits easier and give you the data to improve the system.
What can go wrong
- Leakage. The model sees information it would not have at forecast time, so it looks brilliant in testing and fails live. Always ask, "Would we know this on the day?"
- One-off events. A flood, a closure or a viral post distorts history. Mark these periods so the model does not treat them as normal.
- Structural change. A new competitor, a new product mix or a new counting method makes old history less relevant.
- Over-trust. Planners stop questioning the number. Keep a way to override it and record why.
- Sparse data. Items with few sales are noisy. Forecast at a higher level, such as category or region, and then allocate.
When not to use AI forecasting
If demand is flat and steady, a moving average may do the job and be easier to explain. If you have only a few months of history, wait or forecast at a coarser level. If nobody will act on the number, the project has no point. And where a wrong forecast could harm people, such as in healthcare or grid operations, the forecast should inform a trained person and never act alone. See how this plays out in logistics and supply chain and our wider data and analytics practice.
How to start
- Name one decision and one forecast that supports it.
- Collect the history and write down your current method as the baseline.
- Build the simplest model that beats the baseline on held-out data.
- Present the forecast with a range and a short explanation.
- Run it beside your current process for a few cycles before switching.
If you want help choosing the first forecast or checking your data, our consultants can walk through it with you. Talk to our team.
Frequently asked questions
What is guest feedback AI for hospitality?
It is software that reads guest messages and reviews in several languages, tags them by topic and tone, and shows trends per property. It helps managers see repeated problems. People still decide and carry out the fixes.
Why analyze messages instead of only reviews?
Messages arrive before or during the stay, when a problem can still be fixed for that guest. They are also more specific than star ratings, which makes them easier to act on.
Is it accurate for Thai and Chinese messages?
It can be useful, but accuracy varies by language, slang and politeness style. Have staff label a sample of your own messages and compare their labels with the tool before you rely on it.
Do we need guest consent to analyze messages?
Messages contain personal data, so the PDPA applies. Tell guests in your privacy notice how messages are used, minimize personal details in analysis, and confirm your lawful basis with counsel.
Ready to put this to work?
Turn this playbook into a working system
From a first agent to a national platform, the same team plans it, builds it and runs it with you.
