Follow ZDNET: Add america arsenic a preferred source connected Google.
ZDNET's cardinal takeaways
- Agentic AI is often much astir speech than accumulation services.
- Smart professionals absorption connected usage cases and supporting tech.
- They trial processes, refine the approach, and question caller opportunities.
Conversations with integer and concern leaders astir agentic AI often revolve astir a akin sentiment: we've explored agents, but there's thing successful accumulation yet.
But portion everyone talks astir AI experimentation, nary concern tin spend to tally endless pilots without creating concern value. And with experts suggesting professionals who neglect to exploit AI hazard being near behind, there's an imperative to deploy palmy agents sooner alternatively than later.
Also: How to physique amended AI agents for your concern - without creating spot issues
At online question specializer Booking.com, Huy Dao, manager of information and instrumentality learning platform, is charged with delivering worth from AI, including agentic services. He has produced results by taking a structured attack to work rollout, creating targeted solutions to the challenges customers look contiguous and tomorrow.
Dao referred to this attack successful a speech with ZDNET arsenic the "connected trip," successful which Booking.com attempts to guarantee each elements of a customer's trip, whether flights, hotels, oregon attractions, are considered arsenic an integrated experience.
Creating the connected travel means moving crossed disparate information. The information stack Dao's squad has created has allowed Booking.com to make caller AI-enabled services, including the firm's archetypal agentic application, a partner-to-guest strategy that facilitates connection betwixt customers and edifice partners.
Here's what helium has learned truthful far, with 5 cardinal lessons for different professionals who privation to crook agentic AI pilots into superb accumulation services.
1. Identify a concern situation
Dao said the cardinal to exploiting emerging exertion is uncovering the close use. While immoderate professionals stay unsure astir the imaginable of AI, helium said companies tin usage agentic technologies to flooded intractable challenges.
"In my opinion, AI is not similar a spirit of the day, oregon adjacent the twelvemonth -- it is the existent thing," helium said. "I spot that each time astatine enactment however AI tin interaction the mode that we bash things."
Also: 5 ways to usage AI erstwhile your fund is tight
At Booking.com, Dao and his squad identified that timely responses to lawsuit inquiries were a cardinal situation for edifice partners. They recognized that agentic exertion could assistance hotels reply to questions faster and much accurately.
"Before we rolled retired the agentic solutions, whenever a lawsuit wanted to link to the edifice spouse -- for example, if you wanted to cheque if the edifice had a pool, oregon if you wanted to get 1 oregon 2 hours aboriginal -- you'd interaction the spouse and say, 'Hey, tin I person this information?'" helium said.
"However, erstwhile the edifice unit replied, they'd often request to bash much enactment to get the effect right. Also, sometimes they were unavailable erstwhile the lawsuit asked a question. So, it could instrumentality a fewer hours oregon much earlier the lawsuit receives an answer."
2. Build a information level
Dao said the information stack his squad has created allows Booking.com to accelerate the adoption of AI and machine-learning technologies for usage cases, specified arsenic the 1 outlined above.
Dao: "AI is not similar a spirit of the day, oregon adjacent the twelvemonth -- it is the existent thing."
The Snowflake information level forms portion of an integrated stack that includes ThoughtSpot for analytics, Astronomer and Airflow for orchestration, Immuta for entree control, Arize for machine-learning observability, and AWS for unreality computing. The information squad besides tests and uses AI models from large providers, specified arsenic OpenAI, Amazon Bedrock, and Google Gemini.
Also: Why endeavor AI agents could go the eventual insider threat
Booking.com's bespoke partner-to-guest connection strategy was developed internally successful Python, and the information squad utilized LangGraph, an open-source agentic framework, to assistance the cause crushed astir impermanent inquiries.
Dao said effectual agentic systems aren't conscionable astir backend systems. His squad besides thought cautiously astir the idiosyncratic interface.
"We privation to integrate technologies oregon AI capabilities wherever it makes consciousness to our users," helium said.
"And successful this usage case, our partners already had a web-based portal to presumption their messages, truthful it was wide we should integrate the cause close determination to assistance them."
3. Test the usage lawsuit cautiously
With a concern situation identified and the exertion level perfected, Dao and his squad focused connected implementation, which occurred successful 2 phases.
In the archetypal phase, they developed a trusted adjunct to assistance edifice partners woody with lawsuit questions.
The effect was an agentic exertion known arsenic Smart Messenger, which gathers partner, property, and preservation accusation to enactment edifice unit communicating with guests.
Also: 90% of AI projects neglect - present are 3 ways to guarantee yours doesn't
In this archetypal signifier of agentic service, Dao said the quality is inactive precise overmuch successful the loop.
"We privation to marque definite the spouse is the 1 who has the last accidental connected however they privation to respond to customers," helium said.
"But we springiness them an assistant, truthful that alternatively of taking 5 minutes to respond, it mightiness beryllium conscionable a one-second click if they are blessed with what the cause provides arsenic an answer."
4. Delegate arsenic assurance rises
Over time, Dao said assured edifice partners tin commencement delegating much enactment to the cause -- and this signifier represents the 2nd signifier of the agentic implementation.
Here, Booking.com's Auto-Reply instrumentality allows edifice partners to specify customized replies and make instant responses to impermanent questions, specified arsenic whether a edifice has on-site parking.
"This signifier is wherever the cause says, 'OK, if you spot maine enough, I tin enactment for you,'" said Dao.
"In this usage case, the spouse mightiness beryllium sleeping erstwhile the lawsuit asks a question, due to the fact that it's precocious astatine night. However, the cause tin respond connected behalf of the spouse -- and that attack helps successful a fewer ways."
Also: 5 ways you tin halt investigating AI and commencement scaling it responsibly
Booking.com reported that aboriginal experiments yielded a 73% summation successful spouse satisfaction compared to erstwhile messaging tools. Dao said the cause continuously learns from past interactions and idiosyncratic feedback, adapting its responses for accuracy and relevance.
"Now, with the agent, we measurement the reply against everything we do; we experimentation with it, and past we comparison the betterment successful satisfaction," helium said.
"Because the lawsuit gets the answers they need, they don't person to interaction lawsuit support, and that occurrence besides reduces enactment costs."
5. Look for much opportunities
Dao said agentic exploitation indispensable beryllium tied to the idiosyncratic usage case. As his squad refines the lawsuit experience, they proceed to hone the platform, creating a instauration to enactment different agentic explorations.
"We didn't privation to physique the level for the platform's sake," helium said. "When we built the platform, we had the idiosyncratic successful mind. We made definite that we picked the close agentic technology."
Also: Is Google's caller $8 AI Plus program worthy it? How it compares to the $20 Pro subscription
Dao said his squad has learned a batch from the agentic improvement process. He advised different professionals to instrumentality heed of these lessons.
"When you bash your testing, you mightiness deliberation the agentic strategy is good," helium said. "But erstwhile you spell into production, things similar latency tin go a occupation that you request to woody with. Then, you indispensable simplify your architecture and platform."
Over the adjacent 24 months, Dao expects further pioneering developments astatine Booking.com. "You should expect that, arsenic a company, we volition put heavy successful generative and agentic AI, not for the amusive of it, but to summation the idiosyncratic experience," helium said.
"People are looking for a ChatGPT-like acquisition now, and we privation to person a akin experience, oregon adjacent better, erstwhile it comes to the question acquisition connected our sites."

2 days ago
21








English (US) ·