How this travel company's AI rollout drove a 73% satisfaction boost: A 5-step playbook for your business

Too many AI explorations get stuck at the starting gate. Here's how to ensure your agents reach the finishing line.

Key facts from the article

• Agentic AI often remains in the experimentation phase without reaching production.
• Smart professionals focus on use cases and supporting technology.
• They test processes, refine the approach, and seek new opportunities.
• Booking.com partnered with vendor integrations to create a data platform including Snowflake, ThoughtSpot, Astronomer, Immuta, Arize, and AWS.
• The first agentic application was a partner-to-guest communication system called Smart Messenger, built in-house with Python and LangGraph.
• An early experiment yielded a 73% increase in partner satisfaction compared to previous tools.
• Rollout occurred in two phases: first as a trusted assistant with human-in-the-loop, then as Auto-Reply delegating to the agent for instant responses.

1. Identify a business challenge

Booking.com director of data and machine learning platform, Huy Dao, explains that the key to exploiting emerging technology is finding the right use case. He emphasizes that AI is not just a passing trend but a real tool that can transform operations. At Booking.com, the challenge was that hotel partners often took hours to respond to customer inquiries about simple things like whether a hotel has a pool. Delays hurt guest experience and created frustration on both sides.

Dao recognized that agentic AI could bridge the gap. By automating responses to routine questions, hotel staff could focus on more complex issues while guests got instant answers. The goal was to make the 'connected trip' seamless across flights, hotels, and attractions.

2. Build a data platform

Once the business challenge was clear, Dao's team built a robust data stack to support AI and machine learning. They integrated Snowflake as the central data platform, ThoughtSpot for analytics, Astronomer and Airflow for orchestration, Immuta for access control, Arize for ML observability, and AWS for cloud computing. The platform allows the use of multiple AI models such as OpenAI, Amazon Bedrock, and Google Gemini.

The partner-to-guest system was developed internally using Python and the LangGraph framework for reasoning about inquiries. User interface was also a priority. Since hotel partners already used a web portal for messages, the agent was integrated directly into that portal to minimize disruption. The platform was built with the user in mind, ensuring that technology serves the workflow rather than the other way around.

3. Test the use case carefully

Dao's implementation happened in two phases. In phase one, they launched Smart Messenger — a trusted assistant that gathers partner, property, and reservation information to help hotel staff communicate with guests. The human remained in the loop, reviewing and approving AI-suggested responses. This reduced response time from about five minutes to a single click, boosting efficiency without sacrificing control.

The careful testing included measuring accuracy, latency, and user satisfaction. Dao stresses that what works in a sandbox may not work in production due to real-world constraints like response time and system load. Iterative testing was essential to iron out these issues before moving to the next phase.

4. Delegate as confidence rises

Once hotel partners built trust in the agent, Booking.com rolled out phase two: Auto-Reply. This feature allows partners to define custom replies and let the AI respond automatically to common questions (e.g., 'Do you have on-site parking?'). The agent acts on behalf of the partner, even when the partner is asleep.

Early experiments showed a 73% boost in partner satisfaction compared to the previous messaging system. Additionally, support costs dropped because guests got answers instantly without needing to contact customer support. The agent continuously learns from interactions and user feedback, refining its responses over time.

5. Look for more opportunities

Dao's team is now building on the success of the partner-to-guest system. They plan to apply the same structured approach to other use cases across the company. By focusing on real user needs and a solid data platform, they can accelerate adoption of generative and agentic AI in other areas such as personalized travel recommendations, dynamic pricing, and proactive customer service.

The lessons learned include the importance of simplifying architecture to handle production latency, integrating AI where it makes sense for users, and always tying platform building to specific use cases rather than building technology for its own sake. As Dao says, the goal is not to build AI for fun but to increase user experience.

Looking ahead, Booking.com expects to invest heavily in generative and agentic AI over the next 24 months, aiming to create a travel experience that rivals or exceeds the simplicity of ChatGPT. The company's five-step playbook — identify a challenge, build a data platform, test carefully, delegate gradually, and seek new opportunities — offers a replicable model for any organization looking to turn AI pilots into real-world successes.

Source: ZDNET News