Customer support · Production

Booking.com builds a GenAI agent to assist accommodation partners with guest message responses

The problem

Partners at Booking.com manually replied to each guest inquiry, and even when response templates existed they still had to search for and select the right one. During busy periods this extra effort delayed replies, leaving travelers without reassurance and sometimes leading to cancellations and lost bookings.

Workflow diagram · grounded in source
1
Guest inquiry arrives
trigger
“Guests often reach out with questions about check-in times, parking, or special requests, and partners usually reply manually through Booking.com's messaging platform”
2
PII redaction and category check
validation
“It redacts any personally identifiable information (PII) from the incoming message and checks whether the topic belongs to a "do not answer" category”
3
LLM selects tools
ai_action
“If the question passes these checks, an LLM helps determine which tools to use next”
4
Concurrent tool execution
integration
“Once the relevant tools are identified, the agent runs them concurrently, gathers their outputs”
5
LLM generates final response
ai_action
“uses the LLM again to reason over the results and generate a final response”
6
Response suggested to partner
output
“automatically suggesting a relevant response to each guest inquiry. Depending on the message, it can surface an existing template or generate a tailored free-text answer”
7
Partner reviews suggestion
human_review
“By keeping humans in the loop, we can balance innovation with trust and reliability”
Reported outcome

The agent handles tens of thousands of partner-guest messages daily, and in live pilots the human-in-the-loop approach boosted user satisfaction by 70%, reduced follow-up messages, and sped up response times, with partners reporting less time spent on repetitive questions.

Reported metrics
User satisfaction70%
Daily partner-guest messages handled by agenttens of thousands of guest messages every day
Follow-up messagesreduced follow-up messages
Response timessped up response times
Show all 6 reported metrics
user satisfaction70%
daily partner-guest messages handled by agenttens of thousands of guest messages every day
follow-up messagesreduced follow-up messages
response timessped up response times
partner time on repetitive questionsspending less time on repetitive questions
reply time for supported topicssuggest or send a reply within minutes
Reported stack
LangGraphPythonFastAPIGPT-4 MiniMiniLMWeaviateKafkaGraphQLSuperAnnotateArizeE5-SmallMCP serverKubernetes
Source
https://booking.ai/building-a-genai-agent-for-partner-guest-messaging-f54afb72e6cf
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The agent handles tens of thousands of partner-guest messages daily, and in live pilots the human-in-the-loop approach boosted user satisfaction by 70%, reduced follow-up messages, and sped up response times, with par…

What tools did this team use?

LangGraph, Python, FastAPI, GPT-4 Mini, MiniLM, Weaviate, Kafka, GraphQL, SuperAnnotate, Arize.

What results were reported?

User satisfaction: 70%; Daily partner-guest messages handled by agent: tens of thousands of guest messages every day; Follow-up messages: reduced follow-up messages; Response times: sped up response times (source-reported, not independently verified).

How is this customer support AI workflow structured?

Guest inquiry arrives → PII redaction and category check → LLM selects tools → Concurrent tool execution → LLM generates final response → Response suggested to partner → Partner reviews suggestion.

Related customer support cases