Follow ZDNET: Add us as a preferred source on Google.

ZDNET’s key takeaways

• Agentic AI is often more about talk than production services.
• Smart professionals focus on use cases and supporting tech.
• They test processes, refine the approach, and seek new opportunities.

Conversations with digital and business leaders about agentic AI often revolve around a similar sentiment: we’ve explored agents, but there’s nothing in production yet.

But while everyone talks about AI experimentation, no business can afford to run endless pilots without creating business value. And with experts suggesting professionals who fail to exploit AI risk being left behind, there’s an imperative to deploy successful agents sooner rather than later.

Also: How to build better AI agents for your business – without creating trust issues

At online travel specialist Booking.com, Huy Dao, director of data and machine learning platform, is charged with delivering value from AI, including agentic services. He has produced results by taking a structured approach to service rollout, creating targeted solutions to the challenges customers face today and tomorrow.

Dao referred to this approach in a conversation with ZDNET as the “connected trip,” in which Booking.com attempts to ensure all elements of a customer’s trip, whether flights, hotels, or attractions, are considered as an integrated experience.

Also: Worried AI agents will replace you? 5 ways you can turn anxiety into action at work

Creating the connected trip means working across disparate information. The data stack Dao’s team has created has allowed Booking.com to develop new AI-enabled services, including the firm’s first agentic application, a partner-to-guest system that facilitates communication between customers and hotel partners.

Here’s what he has learned so far, with five key lessons for other professionals who want to turn agentic AI pilots into brilliant production services.

1. Identify a business challenge

Dao said the key to exploiting emerging technology is finding the right use. While some professionals remain unsure about the potential of AI, he said companies can use agentic technologies to overcome intractable challenges.

“In my opinion, AI is not like a flavor of the day, or even the year — it is the real thing,” he said. “I see that every day at work how AI can impact the way that we do things.”

Also: 5 ways to use AI when your budget is tight

At Booking.com, Dao and his team identified that timely responses to customer inquiries were a key challenge for hotel partners. They recognized that agentic technology could help hotels reply to questions faster and more accurately.

“Before we rolled out the agentic solutions, whenever a customer wanted to connect to the hotel partner — for example, if you wanted to check if the hotel had a pool, or if you wanted to arrive one or two hours later — you’d contact the partner and say, ‘Hey, can I have this information?'” he said.

“However, when the hotel staff replied, they’d often need to do more work to get the response right. Also, sometimes they were unavailable when the customer asked a question. So, it could take a few hours or more before the customer receives an answer.”

2. Build a data platform

Dao said the data stack his team has created allows Booking.com to accelerate the adoption of AI and machine-learning technologies for use cases, such as the one outlined above.

The Snowflake data platform forms part of an integrated stack that includes ThoughtSpot for analytics, Astronomer and Airflow for orchestration, Immuta for access control, Arize for machine-learning observability, and AWS for cloud computing. The data team also tests and uses AI models from major providers, such as OpenAI, Amazon Bedrock, and Google Gemini.

Also: Why enterprise AI agents could become the ultimate insider threat

Booking.com’s bespoke partner-to-guest communication system was developed internally in Python, and the data team used LangGraph, an open-source agentic framework, to help the agent reason about guest inquiries.

Dao said effective agentic systems aren’t just about backend systems. His team also thought carefully about the user interface.

“We want to integrate technologies or AI capabilities wherever it makes sense to our users,” he said.

“And in this use case, our partners already had a web-based portal to view their messages, so it was clear we should integrate the agent right there to help them.”

3. Test the use case carefully

With a business challenge identified and the technology platform perfected, Dao and his team focused on implementation, which occurred in two phases.

In the first phase, they developed a trusted assistant to help hotel partners deal with customer questions.

The result was an agentic technology known as Smart Messenger, which gathers partner, property, and reservation information to support hotel staff communicating with guests.

Also: 90% of AI projects fail – here are 3 ways to ensure yours doesn’t

In this initial phase of agentic service, Dao said the human is still very much in the loop.

“We want to make sure the partner is the one who has the final say on how they want to respond to customers,” he said.

“But we give them an assistant, so that instead of taking five minutes to respond, it might be just a one-second click if they are happy with what the agent provides as an answer.”

4. Delegate as confidence rises

Over time, Dao said confident hotel partners can start delegating more work to the agent — and this stage represents the second phase of the agentic implementation.

Here, Booking.com’s Auto-Reply tool allows hotel partners to define custom replies and create instant responses to guest questions, such as whether a hotel has on-site parking.

“This phase is where the agent says, ‘OK, if you trust me enough, I can act for you,'” said Dao.

“In this use case, the partner might be sleeping when the customer asks a question, because it’s late at night. However, the agent can respond on behalf of the partner — and that approach helps in a few ways.”

Also: 5 ways you can stop testing AI and start scaling it responsibly

Booking.com reported that early experiments yielded a 73% increase in partner satisfaction compared to previous messaging tools. Dao said the agent continuously learns from past interactions and user feedback, adapting its responses for accuracy and relevance.

“Now, with the agent, we measure the answer against everything we do; we experiment with it, and then we compare the improvement in satisfaction,” he said.

“Because the customer gets the answers they need, they don’t have to contact customer support, and that success also reduces support costs.”

5. Look for more opportunities

Dao said agentic exploitation must be tied to the individual use case. As his team refines the customer experience, they continue to hone the platform, creating a foundation to support other agentic explorations.

“We didn’t want to build the platform for the platform’s sake,” he said. “When we built the platform, we had the user in mind. We made sure that we picked the right agentic technology.”

Also: Is Google’s new $8 AI Plus plan worth it? How it compares to the $20 Pro subscription

Dao said his team has learned a lot from the agentic development process. He advised other professionals to take heed of these lessons.

“When you do your testing, you might think the agentic system is good,” he said. “But when you go into production, things like latency can become a problem that you need to deal with. Then, you must simplify your architecture and platform.”

Over the next 24 months, Dao expects further pioneering developments at Booking.com. “You should expect that, as a company, we will invest heavily in generative and agentic AI, not for the fun of it, but to increase the user experience,” he said.

“People are looking for a ChatGPT-like experience now, and we want to have a similar experience, or even better, when it comes to the travel experience on our sites.”