100 Emails at 3 AM: The Hidden Cost of Getting Omnichannel AI Wrong

At 3 AM, a hospitality client received around 100 unexpected email notifications.

An AI agent had been connected to the property's inbox. Instead of processing only current guest inquiries, it began replying to unread messages, including emails dating back to 2024.

The technology was working. The deployment wasn't.

And the distinction matters.

For a hotel, an incident like this isn't simply a technical inconvenience. It creates work for employees, introduces unnecessary guest communication, and puts confidence in automation at risk.

The hotel invested in AI to take work off its team's plate. Instead, employees had to deal with the consequences of an automated workflow behaving in a way nobody intended.

This is the hidden cost of getting omnichannel AI wrong: automation can create more operational work than it removes.

Everyone wants omnichannel AI. One digital employee answering calls, managing WhatsApp messages, handling email, and responding to guests across different platforms.

The commercial argument is compelling: more consistent service, broader coverage, and less repetitive work for hotel teams.

But there's a dangerous assumption behind many omnichannel strategies.

If an AI agent works well in web chat, surely connecting another communication channel should be straightforward.

In production, every new channel is not an enablement. It is a deployment.

And every deployment introduces different ways to fail.

Why More AI Channels Don't Automatically Mean More Operating Capacity

Imagine an independent hotel using AI to handle routine guest communication.

Website chat works well. Guests receive answers to common questions. The front desk gets fewer interruptions.

The operator decides to extend that success to email, telephone calls, and messaging platforms.

On paper, the business case looks obvious. More conversations handled by AI should mean less manual work.

But that calculation assumes every channel behaves like the one already working.

It doesn't.

Email carries historical threads, system notifications, and automated senders.

Phone calls demand immediate recognition and recovery from misunderstandings.

WhatsApp and OTA messaging environments impose platform-specific delivery rules.

Facebook introduces its own formatting and communication expectations.

The same conversational intelligence can perform well in one environment and produce costly mistakes in another.

For hotel operators, that means the decision isn't simply whether the AI supports a particular channel.

The decision is whether the provider can operate that channel reliably enough to reduce workload without introducing new risks.

We've seen the difference firsthand at Polydom.

Email: When Automation Creates Work Instead of Removing It

The Axxos Incident

In one deployment for Axxos, connecting the AI agent to email caused it to respond to unread messages across the mailbox.

Some messages were nearly two years old.

The result was around 100 notifications sent to the client overnight.

Nothing was fundamentally wrong with the model's ability to write an email.

The failure came from treating an email inbox as though it behaved like a live chat window.

In chat, an unread message is usually a recent interaction requiring attention.

In email, unread can mean almost anything.

It might be an old reservation inquiry, an OTA notification, a forwarded document, a booking update, or a message someone deliberately left untouched.

We encountered other email-specific failures, too:

• The agent requested a guest's email address even though the guest was already communicating by email.
• It attempted to reply to no-reply addresses.
• It responded to Booking.com and Expedia notifications.
• It treated system-generated messages as ordinary conversations.

Individually, these may sound like minor defects.

For an operator, however, every unnecessary reply creates another opportunity for confusion and another interaction someone may need to investigate.

If the purpose of AI is to reduce manual communication, an inbox full of inappropriate automated responses undermines the investment.

The Business Lesson

Email automation needs to know more than how to answer.

It needs to know when not to answer.

That requires controls for identifying senders, recognizing message threads, filtering notifications, and determining which messages fall within the scope of automation.

In another client case, stabilizing the email workflow required three iterations and an architectural change: separating AI processing from manual review before allowing autonomous handling in production.

That additional implementation work was necessary to make the channel dependable.

For hotels, the lesson is practical.

Don't evaluate email AI by the quality of a sample response.

Ask what happens when the system encounters 500 unread emails, an automated OTA notification, or a thread that hasn't been active for months.

An AI assistant that answers everything isn't necessarily helping.

Voice: A Three-Minute Call That Should Never Have Happened

Telephone automation presents a different commercial risk.

When a guest sends an email, a mistake may be discovered later.

When a guest calls, the experience unfolds in real time.

Hesitation, repetition, silence, and recognition failures are immediately noticeable.

A guest trying to book a room doesn't care whether the problem came from speech recognition, telephony routing, or the language model.

They experience one thing: the hotel's phone service isn't helping.

When Una Talked to an Answering Machine

In one Axxos router case, an agent spent three minutes talking to an answering machine.

From a technical perspective, it was processing audio.

From the hotel's perspective, the system was spending time on a conversation with no guest.

The problem was distinguishing a human caller from voicemail, an automated announcement, or an IVR system.

That distinction doesn't exist in the same way in text chat.

When a Guest Couldn't Get Past a Number

In another phone case, a guest dictated a number.

The agent failed to recognize it correctly and entered a repetition loop.

The guest kept repeating the information without making progress.

The operational consequence goes beyond an inconvenient call.

A guest who becomes frustrated while trying to communicate a basic detail may abandon the interaction, request a human employee, or decide not to continue.

We cannot claim that a particular booking was lost in this incident. But the risk to conversion and guest confidence is clear.

The Business Lesson

A voice AI deployment needs more than natural-sounding speech.

It requires reliable handling of:

• Answering machines and automated menus.
• Spoken numbers and keypad input.
• Interruptions and overlapping speech.
• Repeated recognition failures.
• Silence and response timing.
• Escalation to a human when the conversation cannot progress.

These capabilities determine whether voice automation genuinely expands operating capacity or simply creates another source of guest frustration.

In a separate Polydom analysis of 15,910 incoming hotel calls across 15 AI voice agents, a change to IVR routing exposed callers who had previously dropped out before reaching the AI.

The finding illustrated how telephone infrastructure can conceal friction in the guest journey.

Read what 15,910 hotel calls taught us about AI voice agents.

For hotel operators, the important measure is not merely whether the AI answered the phone.

It's whether the caller could make meaningful progress toward resolving their request.

Messaging: When the AI Gives the Right Answer but the Guest Never Receives It

WhatsApp and OTA messaging may appear easier to automate than phone calls.

The guest writes. The agent replies.

But these platforms introduce restrictions that can prevent an otherwise appropriate response from being delivered.

The Niche Case

In one Niche deployment, messages associated with Airbnb encountered blocks when responses mentioned a website.

The AI could generate a sensible answer.

The channel wouldn't necessarily deliver it.

This is an important distinction for hospitality operators using third-party communication platforms.

Guest messaging isn't successful simply because the AI generated a reply.

The guest must actually receive a permitted, relevant message.

And platform rules aren't identical across WhatsApp, Airbnb, Booking.com, and other messaging environments.

The Business Lesson

A messaging deployment needs channel-specific controls for message content, platform restrictions, and delivery failures.

The system should be tested against real situations before it becomes responsible for guest communication.

For professional vacation rental managers, this is especially important because guest conversations are spread across different booking sources.

A response that works in one environment may require a different approach in another.

The AI may know what to say. The channel decides whether it can be said.

If your AI vendor demonstrates message generation but cannot explain how delivery failures are handled, you're seeing only part of the workflow.

Facebook: When the Answer Is Correct but the Experience Is Wrong

Not every deployment failure prevents a task from being completed.

Some simply make the hotel look less professional.

In a Tailored Stay deployment, an AI agent replied through Facebook using Markdown formatting.

The information was there.

But the presentation looked broken.

For a guest, this is a visible signal that the interaction hasn't been adapted to the channel.

It may not create the same operational consequences as an incorrect reservation, but it can still undermine confidence in the property's communication.

A hotel wouldn't normally send the same message, with identical formatting and tone, through email, SMS, and a public social media conversation.

AI shouldn't do that either.

The lesson is straightforward: consistent information doesn't require identical communication.

A reliable digital employee should adapt its responses to the environment without changing the underlying facts or the property's policies.

The Commercial Cost of Treating Channels as Features

The four incidents reveal a broader business problem.

Hotels often evaluate omnichannel AI through a feature checklist.

Does it support email?

Does it answer calls?

Can it connect to WhatsApp?

Can it work with social media?

These questions establish compatibility. They don't establish operational readiness.

A channel that is technically available but poorly configured can introduce several types of cost.

Additional employee work. Staff may need to investigate inappropriate messages, resolve exceptions, and repair failed interactions.

Guest friction. Repeated prompts, blocked replies, and confusing communication can make routine requests harder to complete.

Commercial exposure. When booking inquiries fail or conversations become frustrating, the property risks losing opportunities to convert existing demand.

Reputational risk. Guests judge the hotel through the communication they receive, regardless of which technology produced it.

Implementation overhead. Problems discovered after launch require investigation, workflow changes, and sometimes human review.

These consequences shouldn't be presented as quantified losses without supporting operational data.

But they demonstrate why the lowest-cost or fastest omnichannel rollout isn't necessarily the most economical choice.

The real value of hospitality AI is the work it reliably takes off your team's plate, not the number of channels displayed on a sales slide.

What We Believe at Polydom: Deploy Channel by Channel

These experiences have shaped our approach to deploying Una.

Una by Polydom is a voice-first AI Digital Employee designed to handle guest communication and supported operational workflows for hotels and professional vacation rental managers.

The objective is to expand operating capacity while reducing repetitive work.

But connecting Una to an additional communication channel isn't proof that the channel is ready for autonomous handling.

We treat each rollout as a separate production environment.

That means calibrating the channel, understanding its failure modes, establishing the appropriate safeguards, and evaluating how it performs with real interactions.

For email, that may mean sender classification, historical-message boundaries, and human review.

For voice, it means testing recognition, interruption handling, automated-call detection, and recovery from conversational failures.

For messaging platforms, it means respecting channel rules and identifying delivery problems.

For social media, it means ensuring replies are appropriate for the interface.

The underlying intelligence can be shared.

The deployment cannot.

Why Human Review Sometimes Makes AI More Valuable

It may seem counterintuitive to introduce human review into a product designed to reduce manual work.

But controlled review during implementation can prevent a much larger burden later.

The email rollout that required three iterations is a good example.

Separating AI processing from manual approval made it possible to evaluate the behavior before giving the system broader responsibility.

Human oversight wasn't the commercial outcome.

It was a safeguard supporting the eventual goal: more reliable autonomous handling of suitable requests.

For operators, this is an important distinction.

The goal should not be maximum automation on launch day.

It should be the highest level of dependable automation the workflow can support.

How Hotel Operators Should Evaluate Omnichannel AI

Before introducing AI into another guest communication channel, ask the vendor to demonstrate how the system behaves when something goes wrong.

A strong evaluation should cover five practical questions.

1. Will This Channel Actually Reduce Staff Workload?

Identify the interactions employees currently handle manually.

Then determine which portions the AI can complete and which still require intervention.

A system that forwards every exception without context may create a new inbox rather than eliminate a repetitive task.

2. What Could Go Wrong With This Particular Channel?

For email, test historical messages and automated notifications.

For voice, test difficult numbers, voicemail, and recognition failures.

For messaging, test delivery restrictions.

For social media, review how messages actually appear to guests.

The demonstration should reflect the channel's real operating conditions.

3. What Happens When the AI Cannot Proceed?

The system needs an appropriate recovery path.

Depending on the situation, that might involve asking for clarification, stopping an unsupported action, or escalating to an employee.

Repeatedly generating another answer isn't a reliable recovery strategy.

4. How Will We Know Whether It's Working?

Measure more than response time and conversation volume.

Useful channel-specific indicators include inappropriate replies, delivery failures, recognition problems, repeat contacts, escalation rates, and successfully completed workflows.

These measures can reveal operational problems that a single overall automation rate conceals.

5. Can We Expand Automation Gradually?

Ask how the vendor handles staged deployment, configuration changes, monitoring, and approval requirements.

A phased rollout can help establish reliability before extending the system's authority.

This approach may be less dramatic than switching everything on at once.

But for a hotel responsible for real guests, real bookings, and real service expectations, predictability has commercial value.

What Should Success Look Like?

An effective omnichannel AI deployment shouldn't be judged primarily by how many communication platforms it supports.

It should be judged by what the operator gains from those channels.

For an independent hotel, success might mean more consistent after-hours communication without creating an additional administrative burden.

For a hotel group, it might mean applying appropriate communication procedures across properties while retaining visibility into exceptions.

For a professional STR manager, it might mean guest requests reaching the right operational workflow without repeated manual coordination.

Those outcomes require AI that does more than generate messages.

They require reliable communication, accurate context, supported actions, and sensible escalation.

And each channel needs to contribute to those outcomes on its own terms.

Omnichannel AI Is Real. But So Are the Consequences of Getting It Wrong.

The idea of one AI Digital Employee working across phone, WhatsApp, email, Facebook, and other guest channels is achievable.

But it isn't one deployment multiplied across several interfaces.

Email has its own physics.

Voice has its own physics.

Messaging platforms have their own rules.

Social channels have their own expectations.

Each environment introduces new ways for automation to create value — and new ways for it to fail.

The 100 emails at 3 AM weren't a reason to abandon AI.

They were a reminder that operational reliability doesn't arrive automatically when a new integration goes live.

For hospitality businesses, the goal isn't to be present on every channel at any cost.

It's to make guest communication more dependable and the hotel team more effective.

The real promise of omnichannel AI isn't more automated conversations. It's more operating capacity without more operational chaos.

See How Una Works in Real Hospitality Operations

Considering AI for phone, email, WhatsApp, or other guest communication channels?

Talk to our human sales team about your current workflows, existing systems, and how Una approaches deployment across different environments.

See where Una is already operating and explore how a controlled rollout could support your property's needs.

Book a quick call with our team to discuss your hospitality AI deployment.