Where AI actually belongs in a restaurant — and where it doesn’t

Hospitality does not have an AI problem. It has a repetition problem, a systems-gap problem, and a marketing problem created by people who have never closed a venue.

Used carefully, language models are good at extracting, classifying, summarising and drafting. They are not good at running a Saturday night.

Where it belongs

• Reading a function enquiry and pulling date, pax, budget, and dietary notes
• Drafting a reply a manager can approve in ten seconds
• Summarising a week of numbers into a digest a group operator will actually read
• Classifying a guest email as booking, complaint, supplier, or general
• Turning SOPs into something staff can query without pinging the GM

Where it does not

• Autonomously changing payroll or rostered hours
• Sending guest messages outside venue-approved rules or making booking commitments without human approval
• Deciding who gets the best table
• Replacing a floor manager’s judgement on a walk-in
• Inventing a 300% productivity claim in a deck

The default is to draft guest messages for review. Routine messages can send automatically only under rules the venue has approved. Complaints, booking commitments, allergy advice and financial decisions always need human approval. The venue keeps control of its voice and decisions.

Automate the boring stuff. Keep hospitality human.

If a vendor cannot tell you what happens when the model is wrong, they are not ready to sit near your bookings, your labour, or your guests.