Recommendation Engines Need an Operating Design

A recommendation engine can rank options. It cannot, by itself, make a promise deliverable. In hospitality and service businesses, relevance depends on more than an algorithm. It depends on accurate data, current availability, suitable content, commercial rules, and teams who can fulfil what is recommended.

Recommendation is an operating capability

The recommendation should reflect the customer’s moment, respect their preferences, draw from trustworthy information, and connect to an action the organization can honour. If any of those pieces fail, a technically impressive recommendation can create disappointment rather than value. Before scaling a use case, ask a few practical questions. Is the input trustworthy? Is the recommendation understandable? Can the customer act on it now? Can the operation deliver it? What will we learn when it is ignored? These questions keep recommendation work grounded in the experience it is meant to improve. The point is not to add more suggestions. It is to make the next choice easier, more relevant, and more credible. The strongest recommendation programs connect customer understanding with service design and operational readiness. That is where a feature becomes a capability.

Recommendation systems are socio-technical systems

NIST’s AI Risk Management Framework describes AI systems as socio-technical: their outcomes are shaped not only by models, but also by data, organizational behaviour, human decisions, and context of use. Its core functions—govern, map, measure, and manage—treat risk management as a continuous operating practice rather than a one-time technical review. (NIST AI RMF 1.0, 2023) That framing is especially useful for recommendations. A team should evaluate the complete decision system:

• Govern: Who owns the use case and the acceptable trade-offs?
• Map: Which customer moment, data, inventory, and operational dependencies shape it?
• Measure: How will relevance, fulfilment, fairness, and unwanted outcomes be assessed?
• Manage: How will teams correct weak recommendations and retire unsuitable uses? Click-through rate is not sufficient. A recommendation can attract attention while creating operational failure, customer regret, or an unsettling sense of surveillance. Pair engagement with fulfilment, acceptance, correction, opt-out, and post-action satisfaction.

Sources

• NIST, “AI Risk Management Framework Aims to Improve Trustworthiness of Artificial Intelligence,” 2023
• NIST AI RMF Core