How to spot fake AI in hospitality software

Artificial intelligence is in the marketing copy of almost every hospitality software vendor. AI-powered insights. Machine learning optimisation. Intelligent automation. The terminology proliferates faster than the underlying capability it describes. Understanding what AI forecasting in hospitality can and cannot do gives useful context for evaluating these claims.

Some of it is real. A significant proportion is not. Knowing the difference is worth the effort before a purchasing decision.

What real AI in hospitality software looks like

Genuine machine learning capability in a hospitality context means one of a small number of things.

Pattern recognition in historical data. A model that has been trained on a sufficient volume of clean historical data can identify patterns that would be difficult to spot manually: correlations between weather, day of week, time of year, and specific demand signals. This is useful when the data is clean and the patterns are stable.

OCR with learned improvement. Optical character recognition — the technology that reads invoices and extracts line items — improves with training data. A system that has processed many thousands of invoices from specific suppliers produces better extraction results than one that has not. This is genuine machine learning in a practical application.

Anomaly detection. A system that can flag when a figure is anomalous — this week's spend in a category is 40% higher than the twelve-week average — is applying a form of pattern analysis. Whether this requires machine learning or is a simpler statistical calculation depends on the implementation.

These applications are real. They produce genuine operational value. They are also not magic: they depend on data quality, they have failure modes, and they are best understood as tools that assist human decision-making rather than replace it.

What fake AI looks like

The most common form of fake AI in hospitality software is manual labour presented as automation.

The clearest example is invoice processing. A platform that claims AI-powered invoice processing but takes twenty-four hours or more to process an invoice is not using AI. Genuine OCR processes an invoice in seconds. A twenty-four-hour turnaround means the invoice is being processed by a person, typically at an offshore processing centre, who reads and enters the data manually.

This model is cheaper to run than genuine OCR development. It produces results that look similar on the surface. The operator sees invoice data appearing in the system and assumes automation is responsible. In practice, a person in a processing centre is doing the entry, with no operational context and no accountability for accuracy.

The tell is the turnaround time. Under a minute is automation. Hours or days is people.

Other signals of overclaimed AI

Vague descriptions of what the AI does. "Our AI provides intelligent insights" is a marketing sentence. It describes nothing specific. Genuine AI capability can be described specifically: "Our model identifies historical demand patterns at the category and day-of-week level, using a rolling twenty-four-month training window." Vague language usually indicates there is nothing specific to describe.

AI features that do not affect the outputs you can test. If the platform claims AI-powered demand forecasting, you can ask to see the forecasts for the last twelve months against actual outcomes. If the vendor cannot provide this comparison, or is reluctant to, the forecasting capability is not producing outputs that can withstand scrutiny.

AI described as a replacement for operator judgement. Genuine AI tools in hospitality are most useful as inputs to human decision-making, not replacements for it. A platform that suggests operators can rely on AI recommendations without review is either overclaiming the capability or building a system that enables operators to be less engaged with their own data — neither of which is beneficial. Responsible AI in hospitality is built around surfacing patterns for the operator to investigate, not automating the judgement itself.

What to ask before buying

When evaluating a platform that claims AI capability, the questions that expose genuine versus claimed capability are direct.

For invoice processing: how long does it take from upload to extracted data? If the answer is more than a minute, ask how the processing is done. If the answer is unclear, assume manual.

For forecasting: can you show me forecast accuracy data from existing customers, comparing predictions to actuals across a twelve-month period?

For pattern detection: what specifically triggers an alert, and what is the basis for that threshold?

Genuine capability produces specific, verifiable answers. Claimed capability produces marketing language. A broader evaluation checklist for hospitality software that looks good in a demo but fails in practice covers what to test before you commit.

The operator control question

The best AI implementations in hospitality share a characteristic: they put the operator in control of the final decision, not the algorithm.

An invoice processed by OCR, reviewed by the operator before posting, is more reliable than one posted automatically without review — because the operator catches the extraction errors and applies the operational context the algorithm cannot have. The AI does the heavy lifting. The operator does the judgement.

A forecasting model that shows its work — here is the prediction, here is the historical pattern it is based on, here is how confident the model is — is more useful than one that produces a number without explanation. The operator can engage with the prediction, adjust it for context the model does not have, and use it as input to their decision.

Hops uses genuine OCR for invoice processing: under a minute from upload to extracted data, with operator review before any posting. We are not interested in claiming AI that is not there. What we have, we describe specifically. What we do not have, we do not claim.