AI in Hospitality: From Chatbots to Connected Workflows

For the past few years, much of the travel industry’s conversation around artificial intelligence has focused on what AI can say: how well it can answer a question, recommend a destination, write an email, or build an itinerary.

That is beginning to change.

The more consequential question now is not whether AI can produce a convincing response. It is whether it can do something useful with the request that sits behind it.

Can it retrieve live information from the right system? Can it recognize what must happen next? Can it complete an approved action? And can it recognize when the situation requires a person?

This shift — from generating answers to managing workflows — is emerging across travel. A recent PhocusWire analysis examined how it is changing the role of travel advisors. The same transition is now becoming visible inside hotel and short-term rental operations.

Removing the invisible work

Travel advisors spend a surprising amount of time on work their clients never see.

David Shull, co-founder and CEO of Tern, estimates that administrative tasks can take up 80% to 90% of an advisor’s day. That includes researching suppliers, checking confirmations, updating itineraries, writing routine emails, and looking for inconsistencies across a trip.

Hotels have their own version of this invisible work.

It is the midnight Wi-Fi question, the request for an early arrival, the search for a reservation, the check-in instruction that must be resent, or the guest who has completed four of the five steps required to access a property.

Each request may look simple from the guest’s side. Behind it, however, someone may need to check the PMS, consult the property rules, confirm a payment, review an access system, update a reservation, or contact another member of the team.

Individually, these are small tasks. Collectively, they consume a significant part of the working day — and continue to arrive after the working day has ended.

This is where AI has an opportunity to create value. Not by replacing hospitality, but by reducing the repetitive coordination that makes hospitality harder to deliver.

From tasks to workflows

The first generation of widely adopted generative AI tools was built around individual tasks. A person asked for an email, a summary, a recommendation, or a piece of research, and the system produced it.

The next generation is being built around sequences of work.

In travel, this can mean reviewing an itinerary, identifying a missing hotel night, checking a confirmation, and alerting the advisor before the traveler leaves. In hospitality, it may mean receiving a guest request, finding the relevant reservation, checking the property’s rules and availability, taking an approved action, and confirming the outcome.

The difference is easy to overlook, but operationally it is significant.

A task has an output. A workflow has an outcome.

This distinction is central to the development of agentic AI. McKinsey’s research into agentic AI in travel argues that the technology’s potential lies in connecting reasoning with action across complex and fragmented travel processes.

Fragmentation remains one of the industry’s largest obstacles. Hotels typically operate across a collection of property management systems, inboxes, booking platforms, calendars, access tools, messaging channels, and internal task systems. As hotel leaders recently discussed with PhocusWire, AI ambitions often meet the structural reality of disconnected technology and inconsistent data.

An AI assistant that sits outside those systems may be able to explain what should happen. An operational AI needs to work inside them.

What a connected hospitality workflow looks like

Una by Polydom offers one example of how this model is beginning to take shape in hospitality.

Una operates across 11 connected systems and 12 guest communication channels, supporting 10 defined workflows. The relevant number is not simply how many functions the AI can perform. What matters is whether it can connect a guest conversation to the system in which the next action must happen.

Those connected environments currently include seven booking and property management platforms — Apaleo, Mews, Clock PMS, Guesty, Hospitable, Zeevou, and Profitroom — alongside Google Calendar, Cal.com, Google Places, and SuiteOp. Custom webhooks can extend workflows into operator-defined systems and forms.

Guest conversations may begin through phone calls, email, WhatsApp, a website widget, social channels, WATI, Telegram, or supported PMS and OTA inboxes.

From there, a workflow can follow a simple operating pattern:

Guest request → Context → Decision → Action → Closed loop

Una first understands the request. It then retrieves relevant information from the connected system, applies the operator’s rules, performs a permitted action or provides the correct next step, and confirms the outcome. If the case falls outside the defined workflow, it is escalated with the conversation context attached.

The workflows currently supported include:

after-hours booking inquiries, from offer search to booking link and lead capture;
gated self-check-in based on identity, agreement, deposit, payment, and guest-detail requirements;
early check-in and late checkout feasibility checks;
pre-arrival and in-stay concierge requests;
stay extensions in supported PMS environments;
self-service cancellations where a direct PMS action is available;
appointment and property-viewing bookings;
escalation to a person or live call transfer with the conversation context preserved;
operator-defined workflows using custom forms and webhooks.

Capabilities differ by integration. The point is not that every system supports every action. It is that the AI is structured around the workflow itself: what information is required, where that information lives, which conditions must be met, what action is permitted, and what should happen when the request cannot be completed automatically.

Proactive assistance is the next step

Most AI systems remain reactive. They wait for someone to ask the right question.

Workflow-based AI can begin preparing for the next likely step before that question is fully expressed.

An early check-in request is a useful example. A generic assistant can explain the property’s policy. A connected system can check whether early arrival may actually be feasible based on the current reservation and availability context.

In some environments, this context can be calculated when an inbound message arrives, before the guest explicitly asks about arriving early.

That does not mean AI should guess what the guest wants. It means the information required to respond intelligently can already be available when the need emerges.

Phocuswright has observed a similar change on the traveler side. AI is increasingly moving beyond inspiration and planning toward assistance during the trip itself. Its latest research describes real-time assistance as a growing part of AI’s value to travelers.

For hospitality companies, the corresponding challenge is making sure that real-time assistance is connected to real-time operational information.

Humans remain in the loop

The move toward action inevitably raises a more difficult question: how much should AI be allowed to do on its own?

The answer is unlikely to be the same for every workflow.

An AI system may safely provide parking instructions without approval. A reservation change may require specific rules. A payment dispute, service recovery decision, or access request demands a much higher level of control.

This is why hospitality AI cannot be built around a single choice between full automation and full human supervision.

The more practical model is graduated autonomy:

routine, low-risk actions can be completed automatically;
conditional actions proceed only when defined requirements have been met;
sensitive or ambiguous cases are transferred to a person;
the context collected by AI follows the request into that handoff.

Self-check-in illustrates the difference. Releasing a door code is not simply another answer. The system may first need to confirm the guest’s details, identity verification, signed agreement, authorized deposit, and completed payment.

If the requirements are satisfied, the workflow can continue. If they are not, the guest should receive the precise missing steps without the protected access information being retrieved.

The value of human-in-the-loop design is not that a person must approve everything. It is that people remain involved at the points where judgment, authority, or accountability are genuinely required.

As a recent PhocusWire analysis of hospitality’s agentic AI challenge noted, hospitality is full of edge cases. AI systems must know not only how to act, but when to ask for more information and when to step aside.

The next evolution

The hospitality industry does not need another layer of software that creates more messages for employees to process.

Its next opportunity lies in systems that can quietly remove work: checking context, coordinating routine steps, monitoring whether conditions have been met, updating the appropriate platform, and bringing a person into the process only when necessary.

This does not make human hospitality less important. It may make it more visible.

When repetitive coordination moves into the background, employees have more time for the moments that actually require hospitality: judgment, reassurance, creativity, recovery, and personal attention.

The winning model is therefore unlikely to be AI alone or people alone. It will be a combination of connected technology and human service — with each handling the part of the guest journey it is best equipped to manage.

The industry has spent several years asking whether AI can communicate like a person.

The more useful question now is whether it can help the work get done.