What responsible AI in hospitality actually looks like
Every hospitality software vendor now claims to use AI. Most of it is either window dressing or genuinely unhelpful. Here is what the useful version actually looks like.
HOPS Team
Product & Operations
There is a joke doing the rounds in hospitality technology circles at the moment. Every platform you speak to has AI. The invoicing tool has AI. The rota software has AI. The reservation system, the allergen tracker, the supplier portal. All of them: AI. Knowing how to spot fake AI in hospitality software is increasingly a necessary skill for operators evaluating these claims.
What none of them can tell you, if you push on it, is what the AI is actually doing and whether it has ever been tested by someone who works in a real venue.
That matters, because the hospitality industry is a terrible environment for AI that has been built by people who have not spent time in it. The stakes are real. An operator who follows a bad AI suggestion does not get to undo it when service is already running.
Why AI fails in hospitality contexts
Most AI features in software products are built by product teams optimising for something that can be demonstrated in a sales meeting. They look good. They produce output quickly. They are easy to show on a screen.
The problem is that the things which look impressive in a demo are often the exact things that fall apart in practice.
One example stands out. A platform rolled out a feature that suggested operators use their phone to film their stock room and let the AI count the inventory. It generated numbers. It looked like technology. The product team presumably thought it was solving a real pain point.
Anyone who has actually done a stock take knows why this is not useful.
Stock rooms in hospitality are poorly lit, stacked with unlabelled boxes, and organised around the logic of whoever last did a delivery rather than any consistent system. The same product might be stored in three different places because Monday's delivery went in one spot and Thursday's went in another. Boxes are often reused. Labels come off. The crisps your supplier switched to are in a box that still says the old brand.
A camera cannot read any of that. It produces a number that has no reliable relationship to what is actually on the shelves. That number then feeds into your GP calculation, your ordering, your variance analysis. Everything downstream of that count is wrong, and you do not know it until something does not add up weeks later.
This is not a technology failure. It is a failure of understanding. The feature was built without anyone asking whether it works in the real conditions operators face.
The pattern behind the problem
The stock room video example is one instance of a broader pattern: AI being applied to contexts where the human has crucial knowledge the system cannot access.
A forecast generated from sales history alone does not know that the pub three streets away has closed and your Thursday evening trade has quietly shifted. It does not know that you are running a limited menu because your head chef is away. It does not know that the private hire booked for Saturday is a works party that always runs two hours over and spends heavily on spirits.
An ordering suggestion generated from your average weekly usage does not know that your current storage is at capacity, that your key supplier has a lead time of four days this week, or that you are deliberately running down a product because it is coming off the menu next month.
In each case, the operator knows something the system cannot. When AI ignores that and produces confident-looking recommendations, the operator has two choices: override it every time (making the feature useless) or trust it and accept the consequences (making it dangerous).
What actually makes AI useful in hospitality
AI is genuinely useful in hospitality when it does one specific thing: surfaces patterns in your own data that would be invisible or very slow to find manually, and presents them as prompts to investigate rather than instructions to follow. The limits of AI forecasting in hospitality illustrate why that framing matters: the operator holds contextual knowledge the algorithm cannot access.
That framing matters. The shift from "here is what you should do" to "here is something worth looking at" is the difference between AI that respects the operator's judgement and AI that tries to replace it.
The contexts where this works are the ones where you have good data, a defined question, and a human who can interpret the output with their operational knowledge. Variance analysis is one of those contexts. Stock take reliability is another.
How Hops has approached this
When we started building AI-assisted features, the test we applied to every idea was simple: has someone who has actually worked in a venue confirmed that this is useful in real conditions?
The features that passed are narrow. They are specific. And they are built on the understanding that the operator knows their venue better than any algorithm.
Smart stock take guidance works like this. After a stock take, if you have meaningful variance, the system analyses your GP figures and variance patterns across your history and surfaces a structured prompt about where to investigate. Not a guess. Not a generic warning. A pattern it has identified in your own data.
If your wet stock variance is consistently higher on Saturdays than mid-week, the system surfaces that. It does not tell you why. That is your job. But it saves you the work of going back through months of stock take results to notice the pattern yourself.
The value is in the framing: "this is a pattern worth looking at" rather than "your Saturday bar team are over-pouring." The first is genuinely useful. The second is a liability.
Stock take health checks address a different problem: not what the variance is, but whether the stock take figures are reliable enough to act on.
A stock take where every category comes in exactly on expected is statistically unusual in a real operation. Real venues have real variation. When every count lines up perfectly, that can mean the operation is exceptionally well-run -- or it can mean the count was not done properly: estimates were made, sections were skipped, someone completed it without actually walking the floor.
The health check looks for patterns that suggest the latter. It does not accuse anyone. It flags that the figures may warrant a recount before they feed into the GP calculation and become the basis for decisions.
This is where the integration layer matters. Because Hops connects to POS systems, accounting software, and supplier records, the AI features can draw on data from multiple sources. Not just the stock count, but the sales data from the same period, the delivery records, the invoice history. Pattern analysis on a single data source produces a partial picture. Pattern analysis on connected data is where you start to see things that would otherwise stay invisible.
“You can really tell HOPS has been built by operators. They understand our needs and provide a solution that is exactly what we want, instead of us having to adapt and change for a system.”
Dominique Fernandes
Head of Operations, Mildreds
The honest position
We are not trying to automate operational decisions. We do not think that is the right goal, and we are sceptical of vendors who suggest it is.
The best hospitality operators are not looking for a system that tells them what to do. They are looking for a system that helps them see what they would otherwise miss, quickly enough to act on it.
That is what responsible AI in hospitality looks like. It does not replace the knowledge that comes from running a venue, doing the deliveries, knowing your team, understanding your customers. It works alongside it. It handles the pattern recognition so that the operator can focus on the interpretation.
Anything that tries to do more than that, and especially anything that tries to do it without genuinely understanding the conditions of hospitality, is worth being cautious about.
If you want to see what this looks like in practice, Hops is built around that principle. Start with a stock take and see what comes back.
Frequently asked questions
What does responsible AI actually mean for hospitality operators?
Responsible AI in hospitality means using machine learning in narrow, well-defined contexts where the operator still holds the judgement. It surfaces patterns in your own data that would be slow to find manually and presents them as prompts to investigate, not instructions to follow. The distinction matters because the consequences of acting on a bad AI suggestion during service cannot always be undone.
Why do AI features in hospitality software often fail in practice?
Most AI features in hospitality software are built by product teams optimising for what can be demonstrated in a sales meeting, not by people who have spent time in a real venue. The stock room camera example illustrates the problem: it looks like technology, but anyone who has done a stock take in a real kitchen knows it does not work in the actual conditions. Features built without that operational understanding tend to produce results that look plausible but cannot be relied on.
Can AI help with hospitality stock takes?
AI can help with stock takes in a specific, limited way: by analysing the patterns in completed counts and flagging when the figures suggest the count may not have been done properly. A stock take where every category comes in exactly on expected is statistically unusual in a real operation, and the health check surfacing that as a prompt to recount is genuinely useful. What AI cannot do is replace the physical count itself.
What is the difference between AI that assists and AI that replaces operator judgement?
AI that assists presents a finding and leaves the interpretation to the operator: this is a pattern worth looking at in your wet stock variance on Saturdays. AI that replaces judgement tells the operator what to do without showing its reasoning or acknowledging the limits of what it knows. The first respects the operational context the operator holds. The second ignores it. Hops is built on the first model -- book a demo at hopshq.com.
How does connected data improve AI pattern detection in hospitality?
Pattern analysis on a single data source produces a partial picture. When a back-office platform connects POS sales data, stock counts, delivery records, and invoice history, the patterns it can identify are more meaningful. A variance that looks like a counting issue on its own may look like a delivery shortfall when the purchase records are in the same view. The integration layer is what makes pattern detection genuinely useful rather than superficially impressive.
Tags
Built for operators
See how operators are actually using Hops.
We could tell you what Hops does. Instead, read what the people running their businesses on it have to say.