Make sure the right things are on your P&L
AI is becoming embedded across hospitality. With such rapid adoption, the conversation needs to move beyond what the technology can do and the more fundamental question of: can we understand and trust how it reached its conclusion? needs to be asked.
Explainability is sometimes presented as a characteristic of more sophisticated, next-generation AI. But I believe it should be much more basic than that. If technology is informing decisions that affect revenue, margin or operational performance, understanding how it arrived at an output should be a minimum requirement. Put more simply: anything that can’t show its working shouldn’t be near your P&L.
For hospitality operators, explainability doesn’t need to mean understanding the inner workings of an AI model. Instead site managers should be able to answer some straightforward questions:
- What triggered this?
- What are the actual numbers behind it?
- Where did the threshold come from? And
- Can I see the underlying transactions?
If they can’t independently explain the reasoning to their own team, then I would question whether the system is genuinely explainable.
Don’t ask AI to make the decision
This becomes particularly important as AI starts influencing more consequential decisions.
There are some areas where I don’t believe AI should ever make the final call. Anything affecting someone’s employment or reputation is an obvious example.
The same principle applies where there are safety, allergen, legal or contractual consequences.
That means providers need to be very deliberate about where AI sits within a process. At Lolly, our starting point is to give the AI one clearly defined job. We keep consequential logic deterministic, log inputs and outputs so results can be reconstructed, and put thresholds in the customer’s hands. If the AI component fails, the system should fall back to showing the underlying information rather than attempting to guess.
Finding the handful of things that matter – Lolly’s Profit Protection Platform
Our Profit Protection Platform provides a useful example. A hospitality business can generate thousands of transactions, making it difficult for a manager to identify the relatively small number that genuinely warrant attention.
The platform looks for patterns that can quietly erode margin: such as voids and refunds clustered around particular operators, discount and staff-pricing misuse; cash variance; and sites beginning to drift away from their established norms.
The objective isn’t to produce another lengthy report. It’s to tell a manager which handful of things among thousands deserve attention – ASAP.
AI doesn’t decide what constitutes the exception. The problem from AI reports is that the result is that the model can potentially introduce information that isn’t supported by the underlying data.
Our own detection model therefore runs as deterministic code. The same data produces the same result, which means it can be recalculated and evidenced. Only then does AI become involved. Its job is to translate an already-calculated exception into straightforward language that a busy manager can quickly understand.
AI stops at the full stop.
Everything afterwards is human, including: whether the issue warrants investigation and what action should be taken. The technology helps someone make a decision, but it shouldn’t make the decision for them.
Explainability is commercially important
People are more likely to act on findings they understand and to ignore scores and recommendations they don’t trust. Explainability can expose poor thresholds more quickly, because a manager who can see the reasoning behind a result can identify when something doesn’t reflect what should be the operational reality.
And making complex information understandable puts useful information into the hands of individual site managers, which is where the potential value begins to multiply across a hospitality estate.
Knowing exactly where AI sits
Our experience of achieving ISO/IEC 42001 certification has reinforced the importance of defining precisely where AI is – and isn’t – being used.
For every AI-enabled feature, we document its purpose, inputs and boundaries and establish clear ownership. Our working rule is simple: if we can’t describe the AI’s job in one plain sentence and explain what happens when it gets something wrong, then we don’t and won’t use it. Hospitality operators should increasingly expect the same clarity from their suppliers.
I believe that within three years we’ll see operators demanding evidence. AI questionnaires could become as commonplace.
- Which model are you using?
- Where is it hosted?
- Is our data used for training? Can you show us the audit trail?
- Can we switch the AI off?
Certification is also likely to move from being a differentiator to becoming a condition of tendering, particularly in areas such as education and healthcare. The most sophisticated hospitality AI will be technology that knows precisely what job it is there to do, and, equally importantly, where its job ends and human judgement begins.
