How We Built a Bilingual Support Workflow for a Kissimmee Property Manager

<i>An anonymized case study: We built a real-time English-Spanish-Portuguese support workflow for a vacation-rental property manager in Kissimmee. The system handled 80% of common issues without human intervention, saved 12 hours per week, and kept a human in the loop for refunds.</i>

(Client details are anonymized and some specifics composited at the client’s request.)

I got a call from a vacation-rental property manager in Kissimmee. They managed about 60 units near the theme parks—condos, townhomes, a few standalone houses. Their guests were mostly families from the U.S., but also a growing number from Latin America and Brazil. The problem? Their support team was drowning. Every day, they fielded 80+ calls and messages in English, Spanish, and Portuguese. They had two bilingual agents, but they were completely overwhelmed. Guests complained about slow responses, and the owner was losing bookings because of bad reviews about communication.

The owner had tried a few things before calling me. They bought a cheap chatbot that only did English. They tried hiring more part-time bilingual staff, but turnover was brutal and training took weeks. They even considered a call center, but the cost was just too high for a mid-market operation. Nothing worked. They needed something that could handle multiple languages in real time, understand the nuances of guest complaints (like a broken AC or a noise complaint), and—most importantly—never mess up on refunds.

The Situation: What Was Breaking

The core issue was simple: too many languages, too few humans. The property manager had three support channels: phone calls (mostly English and Spanish), SMS (all three languages), and email (all three). Their team of five agents—two bilingual in Spanish, one in Portuguese, two English-only—was trying to cover all shifts. Because of time zones and call volumes, though, the Spanish and Portuguese queues often went unanswered for hours. Guests would leave voicemails in Spanish, get frustrated, and write bad reviews. The owner estimated they lost about $4,500 per month in cancellations and lost bookings due to poor communication.

They also had a specific pain point: guest complaints. A family from Brazil might call about a broken air conditioner, or a group from Argentina might complain about noise from a neighboring unit. These complaints needed to be handled with empathy and accuracy. If the AI got the tone wrong—sounding robotic or dismissive—it could make things worse. And refunds? That was a minefield. The owner wanted a human to approve any refund over $50, because mistakes could cost real money.

What They’d Tried Before

Before I came in, they’d attempted a few fixes. First, they bought a generic chatbot from a large vendor. It was fine for English, but it couldn’t handle Spanish or Portuguese at all. They tried to use Google Translate to bridge the gap, but the translations came out awkward, and guests could tell they were talking to a machine. They also tried hiring more bilingual staff, but the labor market in Kissimmee is tight, and they couldn’t find reliable people willing to work nights and weekends. They even looked at outsourcing to a call center in Central America, but the quotes were $8-10 per call, which would’ve eaten into their margins.

The owner was skeptical about AI. He’d heard horror stories about chatbots that made up information or sounded rude. He wanted something that’d feel like a real person, not a robot. That’s when I explained the approach: we wouldn’t just build a chatbot. We’d build a workflow that combined AI with human oversight, using the right tools for each language and context.

The Actual AI Work We Did

We started with an AI readiness assessment to understand their data, systems, and goals. Then we built a custom solution using a stack of tools that are proven for multilingual support. Here’s the technical breakdown without the jargon.

Speech-to-text and translation: For phone calls, we used Whisper (OpenAI’s speech recognition model) to transcribe English, Spanish, and Portuguese in real time. Whisper is remarkably accurate across these languages, even with background noise from a theme park or a crying child. We then fed the transcriptions into a large language model (LLM) fine-tuned for customer support. But we didn’t just use a generic LLM. We used Retrieval-Augmented Generation (RAG) with vector embeddings of the property manager’s knowledge base—things like pool hours, parking rules, and common maintenance issues. This way, the AI could answer questions accurately without making things up.

Routing and tone tuning: We built a routing system using n8n, an open-source workflow automation tool. When a call or message came in, the system detected the language and the sentiment (angry, neutral, happy). For routine requests—like asking for the Wi-Fi password or checking checkout time—the AI responded directly. For complaints, we tuned the tone to be empathetic and apologetic, but not overly deferential. For example, if a guest complained about a broken AC, the AI would say: “I’m sorry to hear that. I’ll alert the maintenance team right away. They should arrive within 2 hours. Is there anything else I can help with?” We tested dozens of variations to get the tone right.

Human-in-the-loop for refunds: This was the owner’s non-negotiable. Any request for a refund or compensation over $50 had to be approved by a human. We built a workflow in n8n that flagged those requests and sent them to a Slack channel where a human agent could review and approve or deny. The AI would draft a response based on the human’s decision, but the human could edit it before it went out. This kept the owner comfortable, and it also meant that complex situations—like a guest who wanted a full refund because of a noise complaint—got human judgment.

Portuguese was harder than expected. Look, I’ll be honest: Portuguese was the hardest part. Whisper handled it well, but the LLM’s responses in Portuguese sometimes sounded stilted. We ended up fine-tuning a smaller model specifically for Brazilian Portuguese, using a dataset of past guest interactions (anonymized, of course). That improved the naturalness significantly, but it added a week to the timeline.

Where We Kept a Human in the Loop

Beyond refunds, we kept humans involved in two other areas. First, any call that the AI couldn’t handle after three tries was escalated to a human. Second, the AI never made promises about specific times or guarantees—only humans could commit to a 10 AM check-in or a specific maintenance window. This was a deliberate choice to avoid liability and keep the guest experience consistent.

The human agents also had a dashboard where they could see all AI-handled interactions. They could jump in at any time if they felt the AI was going off track. In practice, they rarely needed to, but it gave them peace of mind.

The Measured Results

After three months, the numbers were clear. The AI handled 80% of all incoming support requests without human intervention. The average response time dropped from 45 minutes to under 30 seconds. The property manager reported saving 12 hours per week of agent time—time that could be spent on more complex issues or proactive guest outreach. Guest satisfaction scores improved by 22% in English, 35% in Spanish, and 40% in Portuguese. The owner estimated that the reduction in cancellations and lost bookings was worth about $4,500 per month, which more than covered the cost of the solution.

One specific example: a family from Brazil checked in at 2 AM after a delayed flight. They couldn’t find the key lockbox. The AI handled the entire interaction in Portuguese, guided them to the right location, and even sent a follow-up message in the morning to make sure everything was okay. The family left a 5-star review mentioning the “excellent communication.”

“I was skeptical at first, but the AI sounds like a real person. My guests don’t even realize they’re talking to a machine until I tell them. And the refund safety net means I sleep better at night.” — The property manager

What We’d Do Differently

If I could do it over, I’d spend more time on the Portuguese training data. We rushed that part, and it took extra work to fix. I’d also set up a more robust testing environment before going live—we had a few early hiccups where the AI misunderstood a Brazilian Portuguese phrase for a complaint and responded in Spanish. That was awkward. We caught it quickly, but more pre-launch testing would’ve avoided it entirely.

Another lesson: the owner wanted to add a voice agent for phone calls, but we advised against it at first because the voice synthesis wasn’t good enough in Portuguese. Six months later, that technology has improved, and we’re now planning a voice agent implementation for them. Timing matters.

Honest Caveats

This solution isn’t for everyone. If your support volume is under 10 interactions per day, a simple chatbot might be enough. If your languages are more exotic (like Mandarin or Arabic), the accuracy of speech-to-text and LLMs varies. And if you can’t stomach any AI mistakes, you’ll need a full human team. But for a mid-market property manager in a multilingual market like Kissimmee, this approach works.

We also learned that the AI needs regular updates. Guest slang changes, new policies come up, and the knowledge base needs to be refreshed. We set up a monthly review cycle with the client to keep everything current.

Final Thoughts

If you’re a small or mid-market business in Central Florida dealing with multilingual support, you don’t have to hire a dozen bilingual agents. You can build a smart workflow that combines AI with human judgment. The key is to start with a clear assessment of your specific needs—languages, volumes, and pain points. That’s exactly what we do in our fractional AI officer engagements: we help you figure out what tools fit your situation and where to draw the human-in-the-loop line.

Want to see if this could work for your business? Contact us for a no-pressure conversation. We’ll talk about your languages, your volumes, and your budget. No buzzwords, just honest advice.

“I was skeptical at first, but the AI sounds like a real person. My guests don’t even realize they’re talking to a machine until I tell them. And the refund safety net means I sleep better at night.” — The property manager

Frequently asked questions

How much does a bilingual AI support workflow cost?

Costs vary based on call volume, languages, and complexity. For a mid-market property manager handling 80+ interactions per day across three languages, the total cost (including setup and monthly fees) was under $2,000 per month. That’s less than the cost of one full-time bilingual agent.

Can the AI handle angry guests or complaints?

Yes, we tune the tone to be empathetic and apologetic. The AI is trained to recognize sentiment and respond appropriately. For example, it will never argue with a guest or sound dismissive. However, it escalates to a human if the guest remains upset or requests a refund.

What if the AI makes a mistake in translation?

We use high-accuracy models like Whisper for transcription and fine-tuned LLMs for responses. Mistakes are rare, but we have a human-in-the-loop system that monitors interactions. The AI also logs all interactions for review.

Do we need to have a lot of technical knowledge to use this?

No. We handle the setup and maintenance. Your team gets a simple dashboard to monitor interactions and handle escalations. We also provide training for your staff.

Can this work for other languages besides Spanish and Portuguese?

Yes, but accuracy varies. Models for French, German, and Japanese are very good. For less common languages, we may need additional fine-tuning. We’ll assess your specific languages during the readiness assessment.

How long does it take to set up?

A typical setup takes 4-6 weeks. That includes an assessment, building the workflow, training the AI on your knowledge base, testing, and going live. Ongoing updates are monthly.