A hotel chatbot answers a conversation; an AI agent may also take authorized actions in hotel systems. The label is not enough to distinguish them. Ask the vendor to demonstrate a complete task, including the changed record, guest confirmation, audit trail, and recovery when the action fails.
Test the verb, not the product name
“Handles bookings” can mean sending a booking-engine link, collecting a lead, creating a provisional reservation, or confirming a reservation after payment. Those are different services. A hotel that needs fewer missed enquiries might be satisfied with a lead collector. A hotel expecting unattended reservations needs a much deeper demonstration.
Oracle's interface overview distinguishes exchanging hotel data from writing changes back. That distinction is useful during procurement regardless of your PMS brand. It is not evidence that any particular vendor has implemented every available action.
Four levels of capability
| Level | Guest-facing behavior | Evidence to inspect |
|---|---|---|
| Information | Explains breakfast hours or parking | Approved source and current property |
| Assistance | Drafts a reply or collects an enquiry | Staff review queue and captured details |
| Transaction | Creates an authorized reservation or request | Record identifier and resulting status |
| Managed completion | Tracks success and handles exceptions | Delivery evidence, alerts and recovery history |
Higher is not always better for every task. Autonomous refunds may be inappropriate even when autonomous directions are helpful. Decide the allowed level separately for booking, payments, complaints, and housekeeping.
A demonstration that exposes the difference
Use a synthetic enquiry for two adults and a child, with a date change halfway through. Ask for an available room category and a package with meals. After the assistant responds, inspect the stored dates, occupancy, selected plan, and amount. Then interrupt the payment step and resume later. The resumed conversation should not create a second reservation.
Next, remove availability or simulate an unavailable connection in the vendor's test environment. Ask the assistant the same question. It should state the uncertainty and offer a supported next step. A polished sentence claiming success while no record exists is a failed task, even if the conversation reads well.
Authority is part of intelligence
An operational agent needs clear limits: which property it serves, whose instructions it accepts, and which changes require approval. A guest asking for a discount is not permission to alter the hotel's rate policy. A staff member with housekeeping access should not gain refund authority simply by phrasing a request persuasively.
The NIST AI Risk Management Framework provides background on managing AI risks. For a hotel buying decision, translate risk into concrete questions about permissions, evidence, and responsibility.
Measure completion without rewarding silence
A low handoff rate is not automatically good. An assistant can appear efficient by refusing difficult requests or prematurely closing conversations. Count successful authorized outcomes and separately record unresolved enquiries, human transfers, factual errors, and repeat contact about the same issue.
For housekeeping, the completed outcome is not “a ticket was created.” For a booking, it is not “a link was sent.” Define the endpoint before judging performance. Our request-completion guide develops this distinction for guest service.
Where to start evaluating Hotelary
Review Hotelary's WhatsApp AI alongside its reservation system. Use the implementation guide for rollout context, then bring the demonstration cases above to a product session. Confirm each required action for your configuration instead of assuming the word “agent” guarantees it.
A strong assistant should make its work inspectable. The most revealing question is simple: after this reply, what changed, where can the operator see it, and who acts if the change did not happen?
Sources and further reading
Sources reviewed on September 14, 2026. Check current vendor terms and policies before implementation. Examples and checklists are editorial guidance unless explicitly identified as reported research.
- Oracle's interface overview — docs.oracle.com
- NIST AI Risk Management Framework — nist.gov
