World models for hospitality

The venue is a partially observable system

Walk into any restaurant or hotel at 7pm on a Friday and ask the manager what is true right now. The answer lives in at least six places at once: the POS, the PMS, the reservation book, the scheduling app, a group chat, and the manager's head. Tables are turning, two servers are out, a vendor is late, a guest is waiting on a call-back, and the walk-in list is growing — none of it in one place, most of it not recorded anywhere.

That is what makes hospitality hard, and it is why software has failed to fix it so far. A chatbot answers questions. A scheduler schedules. A POS records transactions. Each tool automates one slice and leaves the manager to fuse everything else in their head.

We think the missing layer is a world model: a representation of the venue's real state that fuses every signal — calls, texts, orders, invoices, schedules, sensors — into one continuously updated picture. Not a database of records, but a model that estimates what is true, predicts what breaks next, and simulates the outcome of a decision before it is made.

What a world model changes

• A call about a broken fridge stops being a message and becomes an event: the model knows the fridge matters, which vendor handles it, who needs to approve the repair, and what happens to tonight's prep if it isn't fixed by 4pm.

• A no-show stops being a data point and becomes a prediction: the model learns which reservations actually show up, when the floor gets overstaffed, and where the next gap is coming from.

• Shift coverage stops being a panic at 6pm and becomes a forecast: the model sees the call-out risk before it happens and lines up the backup before service starts.

• Guest recovery stops being a script and becomes a resolution: the model remembers the guest, the incident, the apology that was promised, and follows through until the outcome is closed.

This is the difference between software that reacts and a manager that runs. The operators we work with don't need more dashboards — they need something that already knows what is going on.

The research agenda

Our first working paper, World Models for Hospitality (2026), frames the problem formally: a venue is a partially observable Markov decision process, and the operations layer is a belief-state estimator that fuses observations from the PMS, POS, scheduling, messaging, and sensors.

The paper proposes an architecture for belief-state estimation, dynamics learning, and action interfaces — plus a benchmark dataset and three experiments that run on today's data. The second paper, The Body Needs a Manager (2026), extends the argument to hospitality robotics: navigation and manipulation answer "can I do this?", but a venue needs a layer that answers "should this happen, for whom, at what time, and what does it change?".

Why this matters for operators today

You don't need to care about POMDPs to feel the difference. Every feature Concya ships — calls answered with full context, vendors dispatched with intent, coverage found before service, recovery followed through — is an exercise in world modeling. The model is why the phone agent knows the fridge matters. It is why the night manager inherits the morning's promises. It is why the house runs when the operator isn't looking.

Software that records is table stakes. Software that knows is the operating layer. That is the world model, and it is what we are building.