How We Built an AI Agent That Calls Hotels Like a Human

If you’ve ever checked into a hotel late at night only to be told your room was given away, or that your company’s card wasn’t accepted, you’re not alone. These “last-mile” issues have persisted for travelers and finance teams.

At Navan, we saw an opportunity to help solve that problem using AI — not by building another chatbot, but by creating a voice agent to call hotels and talk to real humans at the front desk. This agent confirms payment details ahead of time, notifies hotels of late arrivals, and handles unpredictable workflows that pop up once a call is in motion. It’s designed to speak naturally, handle messy real-world scenarios, and get smarter with every interaction. It hasn’t covered every edge case yet — some hotels still prefer manual follow-ups — but it already handles the vast majority of real-world scenarios.

Here’s how we built it.

Solving Two Painful Problems for Travel and Finance Teams

This agent was designed to do two things well:

1. Confirm hotel payment in advance.

Travelers often arrive without knowing the hotel doesn’t have the virtual card or payment details on file. This can lead to personal card use, out-of-policy spend, and headaches for finance teams trying to reconcile it all. But our agent proactively calls the hotel before check-in to confirm that it has the correct payment method and that it’s ready to be used. That single interaction helps prevent manual reimbursements, misaligned records, and expensive follow-ups.

2. Help ensure late-night check-ins don’t result in lost reservations.

If a traveler’s flight is delayed or they arrive after hours, the hotel might cancel the booking or give away the room. The agent detects delays in the itinerary and calls ahead to let the front desk know the traveler is still coming. A simple confirmation adds a note to the reservation to help prevent this cause of friction at check-in.

It sounds straightforward, but building a reliable, voice-based AI to solve these problems required some serious iteration.

Why Voice, Not Chat

Hotels don’t run on APIs. Each front desk has its own systems, policies, and processes, so we were unable to rely on structured integrations. We had to build something that could talk like a person and adapt to different hotels and situations.

So we built a voice agent.

What happened to our first version? It sounded robotic. Calls would get dropped. Some hotels would hang up. Others didn’t trust it.

We went back to the drawing board and started experimenting with personality-driven prompts. We asked the agent to try on different roles:

• A helpful executive assistant
• A support agent
• A professional with a sense of urgency
• Someone new to their job — a little nervous, polite, and deferential

A slightly nervous voice worked best. When we told the AI to act like a support agent just starting — the kind who says, “Uh, I just wanted to check if…” — the results were striking.

The hesitation made it feel more human. Front desk agents sometimes asked, “Are you a robot?” -— but they kept engaging with our agent. That was the signal we were after. The goal wasn’t to fool anyone, but to make our agent believable enough to have a productive conversation.

Training the Agent to Navigate Hotel Complexity

No two hotels are alike. Some want a confirmation number. Others ask for a faxed form (yes, still). Some will say, “Sure, but email me.” Others want the card physically present.

Because of this, we couldn’t write deterministic flows. We needed to design an agent that could learn and improve with every call. As the agent makes thousands of calls each week, we evaluate a sample of calls with an LLM-as-a-judge, adjusting tone and responses to improve the calls for continuous improvement.

We’re also starting to teach the agent new skills, like how to fill out payment forms or send emails to front desks. That’s coming next. Powered by our work with AWS on multimodal AI, this next iteration will be designed to be able to use a computer — just like a person would.

How the Agent Understands and Responds

To interact naturally with hotel staff, the voice agent uses a combination of:

• AI speech model engineered to understand and navigate complex phone IVR systems and derive context from the speaker to drive more natural conversation.
• Agent engineering that dynamically adapts based on the task type (e.g., CCA verification vs. late arrival).
• Context engineering (e.g., guest name, hotel name, arrival time, virtual card details) that enables personalized, task-specific conversations.

This structure allows the agent to stay focused, maintain accuracy, and complete tasks and avoid deviating from or overstepping what it’s supposed to say.

We also integrate with programmable telephony APIs to handle the outbound call process, including:

• Call scheduling during hotel business hours
• Call routing logic based on time zone and local practices
• Real-time speech-to-text transcription and sentiment tracking

The agent is trained with custom speech parameters to generate a human-like tone — adjusting for warmth, pacing, and uncertainty when needed.

Multi-Agent Flow Design

While the agent makes the call, it’s part of a broader multi-agent architecture:

• A monitoring agent watches the traveler’s itinerary to determine if a late arrival notification is needed.
• A payments agent confirms if a virtual card is associated and whether Navan is responsible for delivering the credit card authorization (CCA).
• A call orchestration agent packages this information into an instruction set and coordinates the outbound call.

This design allows the system to scale and adapt to future use cases, like amenity pre-requests or loyalty number confirmations. We also log call metadata (duration, hotel response type, outcome) to inform future iterations and system tuning. Every call helps the system improve, because:

• We use human-in-the-loop feedback loops to review failed or ambiguous interactions.
• Patterns from call transcripts are clustered and classified, helping improve agent behavior.

Why Scale Matters

Perfecting this process isn’t just about helping a single traveler avoid a frustrating check-in. It’s about scaling reliability across thousands of trips.

Before we built this agent, travelers and finance teams dealt with missing CCAs. When something went wrong, it resulted in last-minute work for everyone: the traveler, the travel manager, and the finance team.Solving it consistently would have required a team of human agents working around the clock.

Instead, we built an agent that does it automatically.

Doing that and delivering consistent, human-quality interactions at scale isn’t just an engineering achievement — it’s also a win for business.

Built for the Real World, Not the Lab

This creation isn’t AI for AI’s sake. It’s a real product solving real problems in the real world.

To be clear, it hasn’t replaced every edge case yet. Some hotels still want manual follow-ups, for example. But even in those cases, our internal teams now know exactly what to do, because the agent logs the interaction and flags when human help is needed.

And that’s just where we are today.

We’re building toward an AI system that goes beyond automation to achieve true collaboration — a system that helps people do their jobs better, faster, and with fewer headaches.

To get there, we’re training it to manage more workflows for travelers and finance teams — from form-handling to multi-lingual support. And as Navan continues to expand globally, this AI will be ready to scale with us — one call at a time.